Hmm, 24%-42% performance increase (14/16TFlops vs 11.3TFlops) for 70-90% price increase... And prices were already inflated by crypto that is now collapsing. Not sure who is the target market for this tech honestly. Still only 11GB RAM, even if 50% faster, making it a nonsense purchase for Deep Learning enthusiasts (state-of-art models are already larger).
Unless somebody invented RTX-based coin of course, then this is the minimal price...
>Not sure who is the target market for this tech honestly.
These are gaming GPUs being announced at a games conference to an audience of gamers and games journalists. The focus of Huang's talk was how accelerated ray tracing will improve the graphical fidelity of games.
GPU compute is a spin-off from gaming. Gamers remain the primary market. They generate the volume that allows Nvidia to sell a supercomputer on a card for under $1000. If you want a professional product, Nvidia are more than happy to sell you a Tesla or a Quadro at a professional price point.
The point was different - it is way too expensive for regular gamers, 2080Ti likely won't be able to do 4k@60Hz like 1080Ti couldn't either and 10 games featuring RTX in the nearly future is not a sufficient draw, especially when some of the effects now look subjectively worse with RTX ON (see those shadows from multiple lights above dancing people from Jenson's demo). So the question remains - who is the real target audience that will actually buy those cards at these prices? Did NVidia turn into Apple and made RTX its own iPhone X?
Waiting this generation out until RTX is more wide-spread/tested and going for the next 7nm generation with hopefully AMD having a proper GPU to compete seems like a better strategy for most gamers out there.
The 80 series has always been a low-volume, high-margin halo product within Nvidia's gaming range. It's dirt cheap and high-volume compared to Quadro, but top-of-the-range for gaming. Cryptomania has revealed to Nvidia that they probably priced the 1080 too low at launch - many gamers were in fact willing to pay substantially inflated prices for The Best GPU.
If the mass market decides to wait for the RTX 1060 or 1050, that's fine with Nvidia, as they face no real competition from AMD at the moment. It's very much in Nvidia's interests to make the most of their market dominance.
The 70 series is traditionally pretty popular for gamers though. At $600, though, I mean... that's at the point where just the graphics card is more than one of the 4K consoles Microsoft or Sony is putting out. Obviously a PC game is going to look nicer, but someone has to start thinking about that comparison.
"someone has to START thinking about that comparison."
Sorry, but this comment is really kind of hilarious.
The "PC vs. Console" debate is something that almost predates the Internet and it has generated countless forum wars...
A high range PC has always been the more powerful and expensive gaming machine, since basically the first 3DFX cards in the late 90s, some people are OK with that, other prefer consoles as a perfectly acceptable alternative.
That's not really what I meant. Obviously some people have drawn their lines in the sand and will never consider switching. I don't think that's everyone. I play games on both console and PC, as I imagine do many others. If the price is too unreasonable, or the PC version doesn't work right without a bunch of configuration, or whatever, I can't be bothered with it and will just go to the console version of something.
> I mean... that's at the point where just the graphics card is more than one of the 4K consoles Microsoft or Sony is putting out.
The console might output 4K but that doesn't mean the GPU inside can handle higher settings than a 1060. The $600 GPU is irrelevant to that comparison.
> The console might output 4K but that doesn't mean the GPU inside can handle higher settings than a 1060
Which is kinda irrelevant too, since console games are highly optimized for exactly 1 graphic card & rest of the setup. You get all the details hardware is capable of smoothly, nothing more or less. No fiddling with tons of settings that most have no clue about.
I don't get this often used argument - my games look better on PC than on consoles. Yes some UI gimmicks look better, textures have higher res, but after 30 minutes of playing it really doesn't matter at all, quality of gameplay and smoothness of experience is the king. Of course if one pours 5x the amount of money into PC that would be spent on console, there needs to be some inner justification. But it still doesn't make sense to me.
This is a view of person who plays only on PC, never had any console.
If you set it to medium you should get a smooth experience. You don't have to do anything I'd call "fiddling", and being optimized for a specific GPU is overrated (and not even true with the current consoles). Especially when you have a much better CPU.
> Of course if one pours 5x the amount of money into PC that would be spent on console, there needs to be some inner justification.
You can get a prettier and smoother experience if you do that and don't put the settings on super-ultra.
But also, if you're already going to have a computer, popping in a better-than-console GPU is cheaper than a console.
To the extent that's true it kind of works against your argument, doesn't it? I doubt that PC sales look better in India because Indians all have top-of-the-line Alienware rigs.
The top of the line rigs are what you need to play new titles on Ultra settings, sold at $50-60+
With hardware comparable in pricing to what you'd find in a console (or using something that doesn't make much sense to mine with, like a GTX 780Ti) you can easily play a 3-5 year-old game at 1/2 to 1/4 of it's original price, which might even be reduced further by -50% to -90% during a Steam sale.
But it does open up the system's library of exclusive titles, which makes it seem compelling to someone considering a video card purchase who already has an older one that does OK with games.
I think the cryptomania revealed that people (young people, gamers, who got into cryptocurrencies) could earn a few bucks or so back with their investment. Some of whom used mom 'n pop's electricity grid for that purpose. If they had to pay that back, it was likely 0% interest.
I mean you're not countering his point in any way. He didn't say nobody would buy it, but it's a simple fact that most people can't afford or justify a thousand dollar GPU.
The vast majority of people buying Nvidia GPUs in the 10xx generation were going for 1060s and 1070s.
Yet, going by a scan of people on the train last time I caught it, heaps of people seem to find money for iPhone Xs at almost the same price point.
If you're sufficiently dedicated, even with limited funds going for every second or third iteration of halo products can be a great strategy. That way when you get it you'll have the absolutely best there is and a couple of years down the track it'll still be great but maybe it won't quite do Ultra anymore on the latest titles.
The 1080TI transformed my experience of my X34 Predator (it even paid for itself through timely sale of mining profits) which does 3440x1440 @95hz. I certainly wouldn't mind the new card but I'll wait for the next one after that minimum.
Don't most people get expensive phones because of subsidies from carriers? Or at least, they pay monthly installments for these devices, through data plans (basically).
Do people really take out loans to get super expensive video cards?
People seem to be missing something in this particular point of the conversation. It's not a function of absolute price. It's a function of price-per-performance. Sure, a lot of the crowd here can afford the 1080 or the Titan, but the bang-for-buck favors the lower-end cards.
'most people can't afford or justify'. Come one. People buy cars and other stuff. Someone working fulltime and buying a 1k cheaper car can already afford and justify a 1k graphics card.
We are already talking about a small percentage of people who wanna buy a graphics card.
From those people, it is easily justifyable to spend less on a car, a holiday or rent and instead having a nicer gaming rig. If you spend a lot of time playing games why not?
Modern technology is way cheaper than the previous/old status symbols.
I'm thinking about buying a car and one simple calculation is still what else i can do with that money.
And yes in munich, where i live right now, there are enough people with a car who could use public transport and don't us it.
The target group of a 1k graphics card is not someone who can barely afford the car he/she needs every day and would not be able to earn anything if the car breaks down...
If you played seriously then you knew enough to turn the settings lower, not higher.
The last thing you want is having your field of view obscured by colorful explosions, bloom, debris, etc. when your opponent has a crystal clear vision on you.
But that is a completely orthogonal point to the question as to whether doom runs well at 4k with all features on. Because I am asking that question does not mean I would play death match that way. But I might indeed go through the single player campaign that way.
Indeed I didn't play Doom at 4k at all, because as I said, it felt like garbage at 4k, no matter what settings, on 1070
Doesn't matter if they aren't "kids" anymore or not, there's a reason AMD focused hard on the $200-300 range for graphics cards - because that's where most buyers budgets are. There are people who spend more, but even many enthusiasts are shopping in the $400-500 segment for a card to support high-refresh-rate gaming or higher resolutions like 1440P, the number of people who blow $800+ on a GPU like a 1080 Ti are few and far between in comparison.
Certainly not, but the infuriating part is historically the performance of those cards trickles down to the lower tiers at more reasonable prices as the generations go by. The GTX 1070 beat the GTX 980 Ti at an MSRP $200 less just one generation later, meanwhile at least from pure TFLOPS numbers the RTX 2080 is less powerful than the GTX 1080 Ti while costing around the same.
One would be forgiven for expecting roughly GTX 1080 Ti performance in the RTX 2070 at around $449-499 USD.
This all sounds like normal r&d and market forces. New stuff is low volume and premium prices. Once it becomes more common and more production lines are switched, the prices fall and the features get included into other models. This applies to virtually every product.
Or did you want to highlight something else I missed?
The new products are launching at the same price point similarly specced parts from the previous generation have been selling at. "Low volume" doesn't really play when you're talking silicon manufacturing, when you spend millions of dollars to make a mask you want to get ROI on it quickly - the GTX 1070 sold at over $200 less than the GTX 980 Ti, for example, at launch.
When you sell a product that actually has less compute performance (the RTX 2080) at the same price point as the higher-end part from the last generation (GTX 1080 Ti) something has gone horribly wrong. A lot of this is likely due to the Tensor/RTX units on the new GPU's taking up die space and there hasn't been an appropriate process shrink to make up for it, but it's all the more reason these are REALLY unappealing options for anyone outside the top-end enthusiast segment (the GTX 1070 is the most popular enthusiast card this generation, because even enthusiasts have budgets - which is usually the $400-500 range for GPUs).
tl;dr; Prices here make no sense, cards with similar or less performance selling at the same price point you could get from the previous gen - just with raytracing support added on top (so it won't net you WORSE performance with this new functionality utilized). I don't know who Nvidia thinks is going to buy these from a gaming perspective.
The hype around "ti" is unreal. Now they're changing it to "RTX" and "ti". /shock /awe /s
To me, it's pretty clear NVIDIA is cashing out.
Have you not noticed the market slaps the word gaming on commodity hardware along with a bunch of christmas lights and people happily pay a premium for it.
Gamers aren't the brightest bunch and a $1000 is the right price point when people are gladly dropping that on a mobile phone now.
Sure compared to a few years ago I'd agree with you, this market? this hype? No.
Nvidia has a history of just sitting out their performance lead, see Geforce 8800 vs Geforce 9800.
Even in the initially released marketing material, by Nvidia, the 8800 GTX had the obviously way better raw specs than the 9800 GTX. Took them a couple of days until they changed the material to compare on performance % in different games.
But the 9800 GTX was actually a slower card than the one year older 8800 GTX due to lower memory bandwidth and capacity. As such it was competing against one generation older mid-range cards like the 8800 GTS 512.
NVDA also has a significantly higher market share than AMD does right now, that doesn't change that $200-300 is still the most common price point for consumer GPU purchases.
Current Steam user survey results (now that the over-counting issue has been fixed) shows the GTX 1060 as the single most popular GPU installed by active Steam users with a 12.5% market share, the GTX 1050 Ti and 1050 take second and third place with ~9.5% and ~6% respectively that means about 30% of Steam users have a current-gen GPU in the $200-300 price range.
So yes, volume != profit, but the consumer obviously trends towards more reasonably priced cards. Cards in the Titan price-point that NVidia is trying to sell the RTX 2080 Ti at are so uncommon that they get lumped into the 'other' category of the Steam survey - and since I highly doubt magic like doing integer operations in tandem with FP32 operations is going to bring that much of a performance improvement to the majority of gaming workloads in tandem with the really weak raw numbers of the just-announced cards (fewer TFLOPS on the 2080 than the 1080 Ti selling in the same price bracket) it's obvious Nvidia is really taking the piss with their pricing. You're paying more for raytracing, that's it - and while it's certainly a cool feature I don't really see gamers caring that much until it becomes usable at that $200-300 price point.
Thanks. Not sure where they got it from. The Anandtech article that is linked to there does contain some TFLOPS numbers but I think they came up with those numbers somehow based on the CUDA core count so could well not be accurate.
I guess it's just simply twice the number of Cuda cores times the operating frequency and so it's accurate as such but lots more goes into gaming performance of a GPU.
They're really not. Compared to many other popular hobbies gaming's yearly cost is really low. Things like audio, photography, cars, warhammer, travelling, winter sports each will have yearly costs that make gaming seem cheap as hell.
Football (soccer) is cheap, so is basketball, traveling can be done on a budget (backpacking, hitchhiking). Board games or card games (not the collectible or trade-able cash-grabbing variety) are also cheap.
There's many expensive hobbies, but also a ton of cheap ones.
I agree with you, aliasing is obviously less noticeable at higher resolutions, simply because the pixels are smaller (or the pixel density is higher, whichever way you want to see it).
Which ones though? Tbh I do have a 4k gsync so that does help when fps goes below 60. I find anything from 50-60fps to be smooth and below 50 it starts to get too choppy. It also helps with the input lag to run gsync @ 58fps. The most current game I run ultra is far cry 5, it's a pleasure @ 4k.
Ultra is ridiculous and unnecessary. I play 4K60 on an RX 480 with "smart" settings — max textures/meshes, min shader effects, no ambient occlusion, no dynamic reflections, etc.
The only point i'd like to make is that the only reason that you "can't do" 4k@60 is because devs decided to tune their effects scaling in such a way that this is the case.
This doesn't affect the argument that you're making. I just think it's actually incredibly absurd to complain as though it's the hardware's fault for not being able to keep up with 4k@60, when it's the devs who you should be looking at when you're disappointed with the performance on a given piece of hardware.
Oh yes it’s the developer’s fault for not “tuning” something the right way. Sure.
You can “tune” something all you want, you’re always going to have a lower quality representation if you want better performance. The hardware should give the developers the possibility of getting better quality graphics at higher resolutions. We can play the original doom at 4 and probably even 8K without much problems. But that doesn’t mean it’s because they “tuned” it better, it’s because hardware has gotten better and hardware will always be the limiting factor for games.
I think the point is that with PCs they make less effort to eek performance out of hardware. When you've got a console, you know exactly what you'll be optimising for and work really hard to get the most out of it. With a PC release I think Devs tend to make far less effort and simply up the requirements
They are likely incentivized to jack up the price on personal purchases so the big manufacturers can have more overhead integrating it with their consoles or pre-built gaming PCs.
They demonstrated real time super resolution on top of hybrid rendering. The meaning of 4k@60Hz has changed. They can render that just fine -- it's just a question of how many of the pixels are imagined by the model and how good the model is.
With my 1080 (not 1080Ti) I play most of the games (just couple of exceptions, really) with 60 FPS on 4k monitor. And look at 2080 with the same price tag as 1080 - gamers will wipe it out from stock in seconds.
Talking about two different things here... You mentioned that card is capable of outputting 8k HDR @ 60Hz, i.e. your Windows desktop can happily have 7680x4320@60. I mentioned that running games at 4k@60Hz or 3k@144Hz smoothly might not be possible for some demanding games and that many gamers expected that from the new generation.
You can slow any hardware with sufficiently inefficient program (or sufficiently detailed scene, if we're talking about GPUs). You can easily make any videocard work at 0.001 FPS if your scene is heavy enough. So it only depends on game developers, it's unfair to blame Nvidia for that. GPU progress is astonishing, at least when compared with CPU progress.
Ehm, if your server runs a reasonable modern x264, you should get significantly lower bandwidth at "transparent" quality compared to what the GPU's hardware encoders is even capable of reaching. The reason there being that the hardware encoder can't use some features x264 implements sufficiently fast to make them worthwhile to use at that sort of time investment.
Please don't try to measure lossy-anything-calculation by speed alone, always make sure that the required quality can even be reached, and even if, that it still exhibits the performance benefits, even after you tune these features up far enough.
It's not about blame, it's being realistic that 4K means 4x the pixels and 8K means 16x the pixels, and while these cards represent a lot of progress they're nowhere near that level.
Seems like gaming has long since hit an "eternal September" as producers are spewing out ever more impressive tech demos.
I have long since taken to killing every graphical effect game devs can think of throwing into a game (if they supplied a toggle for it) simply so i can tell where the enemy is without wearing sunglasses indoors in a dark room.
It's far from a nonsense purchase for Deep Learning professionals. At https://lambdalabs.com nearly every one of our customers is calling us up interested in the 2080Ti for Deep Learning. The reality is that very few people training neural networks need more than 11GB.
"Reality", in our case, is based on conversations with a few dozen customers who use our products as well as the actively publishing researchers we know personally. The majority of sampled customers are training convnets of some fashion, the most common being Yolo/SSD and ResNet.
Most people I know are just using Google Cloud. Directly integrated with Tensorflow, and way more scalable.
I can run 10 GPUs on my model training runs and finish in an hour now when they used to take at least 2 or 3 days, and it took absolutely no work on my end. It’s been absolutely wonderful for my productivity. The price doesn’t matter either because compared to how much the people are being paid at these sorts of companies, it really doesn’t matter. The boost in efficiency is so much more important.
GCP is pretty tightly integrated/optimizes for Tensorflow. That’s why scaling GPU amounts wasn’t a hassle for example since I was already using the TFEstimator framework.
I’m pretty sure for the next Tensorflow 3.0 update, they’re rethinking towards a Pytorch style with more dynamic computational graphs.
Do they know if tensor cores in the these processors are just good for interference or are they similar to the more pricey models (floating point precision)?
Then most likely they would not be successful in their businesses; training older models or newer ones with a batch of 1 is not a recipe for success these days when you more like need 100s of GPUs running slight modifications to models in parallel to find the one that works.
Most people I know who publish in NIPS/ICLR train their models using between 1-4 GPUs under their desk or 8 GPUs in a server. I would argue that these folks make competitive models and "are successful in their business" with only 1-8 GPUs. While having access to hundreds of GPUs helps you explore model space more quickly, it's not a hard pre-requisite to success in Deep Learning.
You know, there is more to deep learning research than meta-learning/architecture exploration. Sure you can explore the hyper-parameter space faster with 500 GPUs and get yet again a 0.05% better test score on ImageNet (or more I don't actually know), but there are other ways to do something meaningful in DL without using such compute power.
That's a fair point and I agree. It's just sometimes difficult to beat automated exploration; as a standard company you probably don't have access to top-end researchers/practitioners, just average ones, and those might get a significant boost by trading smartness for brute force and run many models in parallel in evolutionary fashion.
When you e.g. look at how Google's internal system constructs loss functions and how many combinations do they try, one has to have an unexpected idea to beat their results, and that idea can be usually quickly incorporated into their platform, raising the bar for individual researchers. At Facebook they basically press a few buttons and select a few checkboxes, then wait until best model is selected, leading to frustration among researchers.
It's just an indicator that extensive improvement is possible. But as with adding more and more shader cores you can get your FPS gainz to the point of Ray-tracing appears around the corner and now you need to add different kind of cores.
Same process for research. You suppose to find some insights on how to do one thing or another, find the direction of search, eventually there would be hardware to fully explore that direction. Then you move on to a different direction. Rinse-repeat
We used to call this model fishing and regard models that came from places that did it with fear and suspicion (due to sudden failure and poor performance in production)
What has changed that people think this is a wise approach?
They don't suddenly fail any more, and performance in production is fine... This is a rather empirical science right now.
If forced to speculate, I'd say far larger datasets are part of the answer. If you can afford to hold back 30%+ of your data for (multi-tranche) verification the difference between testing and production becomes a problem for the philosophers.
Hmm this isn't my experience. My experience is that highly complex models tend to reflect accurately the ground truth in the current dataset, but not the domain theory that generated it, so when things change (like, prevailing weather conditions, political alignments, interest rates, data usage) which move the distribution (but not the domain theory) they do fail. The question is : are you trading poorer compression for more fidelity or are you learning the system that generates the data?
Those are the "Founder's Edition" prices for the first cards from Nvidia, so it's not quite this bad either.
MSRP on cards from OEMs should start at $499 for the 2070 and $699 for the 2080 ($100 cheaper than Nvidia's).
Personally I still think it's insane. I have a GTX 970 from October 2014 and that was $369.
EDIT: Even my $369 was a bit of a premium (OC edition and shortly after launch, I don't remember but I'm guessing I bought whatever was in stock). Wikipedia claims the GTX 970 launched in September 2014 at $329.
Assuming the $329 MSRP is comparable to the $499 announcement, the __70 line is up 52% in two generations. I'm sure it's better hardware, but that's a big pile of money.
And if my 970 experience holds through today, the OEM cards that are actually available are going to be pricier than the MSRP, but maybe still lower than the Founder's Edition.
For those confused, these are about what the 10 series Founder's Editions were priced at originally. When the 10 series was released, the 9XX series was around the same price as the 10 series is now compared to 20. The price isn't going up by 70-90%. It's that the previous series (10) is going down.
Does that mean the 10 series isn't worth it? By all means, no. The 9 series was still an excellent series when 10 was released, and people did pick the up.
Ehhhh, not really, but the release seems a bit different this time around. The 1080 Founders Edition and 1080 Ti FE cards both debuted at $700 USD. (The 1080 Ti was released nearly a year later.) This puts it closer to the now released 2080 FE, which is $800 USD.
The 2080 Ti FE, however, is in a league closer to the Titan X/Xp, which were at $1,200. Also they're releasing the highest-end Ti edition at the same time as the ordinary version, which is a first for Nvidia, I think? (The Titan Xp was also launched after, not concurrently, with the 10xx series...) I think the concurrent launch of the 2080 Ti with the ordinary variant means they're positioning it more like an alternative to the Xp, while the non-Ti variants are closer to the ordinary gaming cards you'd normally get. In other words, for people willing to blow their price/perf budget a bit more.
For DL workloads the 1080 Ti is very cost effective (vs the Xp), so it remains to be seen which variant will have the better bang/buck ratio for those uses. I suspect the fact these include Tensor Cores at their given price point will probably be a major selling point regardless of the exact model choice, especially among hobbyist DL users. The RTX line will almost certainly be better in terms of price/perf for local DL work, no matter the bloated margins vs older series.
They may also be keeping prices a bit inflated, besides margins, so they can keep selling older stock and pushing it out. The 9xx series, as you said, continued to sell for a while after 10xx was released. I expect the same will be true this time around, too, especially with prices being so high.
Tensor cores (e.g. SIMD matrix FMAs) are extremely useful for training. There is nothing shaky about it.
I do not get why you're being so bizarrely negative about this card. Yes, there are an enormous number of applications, for both training and inference, where an 11GB ridiculously powerful card (both in OPS and in memory speed) can be enormously useful.
Yes, I agree, tensor cores are awesome, if your framework of choice doesn't have some rough edges you inevitably hit when you try to do advanced models on e.g. V100, but which work just fine on TPU. I think the presented Turing card is a masterpiece, just given I was hitting 11GB limit with some semantic segmentation and multi-object detection a year ago, I am obviously disappointed that this wasn't increased, and I am forced to buy RTX 5000 instead ($2300) or used K80/P40. Also, for a gaming card outside a few RTX games I doubt it will give adequate value to gamers that expected 144Hz or faster VR and similar goodies. For raytracing and as a fusion of RT and DL it's truly redefining computer graphics.
Tensors cores were built for training. For inference they added the int8 instructions (dp4a) which have lower precision. The Turing also has int4, and for inference this card blows the Volta out of the water since the v100 was only ~TOPS for int8. The Turing is 250 TOPS for int8 and 500 TOPS for int4.
The price of 1080's where $600-$650 for the founders edition cards when they first came out. I know cause I preordered at those prices. I think the price of the 2080 FE cards are fair given that they are not only faster at raster graphics, but they also have the Raytracing/AI capabilities.
For video professionals doing ray tracing this upgrade would likely be a must-have. This assumes the improvement (i.e. 6x or anything in that ballpark) is real and their software supports it... if so, hell, nVidia could double or triple the price of the card and it would still make sense for them. nVidia is just cashing in on a lack of any real competition at the high-end currently... hopefully we'll see AMD get back in the game at some point re: GPUs.
Not true. There is (or at least until now, has never been) a reason for video or 3D professionals in the "Arts" to use anything but GTX cards. 3D CAD and AI professionals on the other hand will profit from the Quadro cards.
The main difference between Quadro and GTX were amount of memory, clock speed, longevity and drivers.
GTX are faster clocked but relatively less reliable.
GTX are game optimized, Quadro drivers are CAD optimized (extremely so)
GTX are much cheaper, but may have a shorter lifetime (but who runs a render farm on 10 year old cards...)
GTX offer very fast single precision calculations, Quadro single and double. Almost all 3D rendering is done single precision.
It's a spectrum. Not every professional wants and/or can afford the top end cards. Many professionals (i.e. those who make money doing a thing) have been using high-end gaming cards for work pretty much since they existed. There's also the long-running debate as to the actual value of the 'pro' line of video cards for non-mission critical purposes (enough of one that nVidia in their license prohibited the use of gaming cards in servers with the odd exception of crypto mining)
But they are the low-to-mid segment of professionals doing weddings and local business presentations with median income around $50k. The ones doing interesting work can't live without real 10-bit HDR on calibrated 4k screens for realistic printing/video projections, without proper 5k+ RAW cameras and top-end lens etc. and those are extremely expensive.
Every self-declared "professional" I've met spent most of their time unproductively fiddling with their equipment.
While a select few can push the envelope with technology alone, a bit of talent seems to easily compensate for almost any technological limitations. The "latest and greatest" is the easy route to mediocracy.
That's been true for all creative disciplines: from photography to writing to animation. There might even be a mechanism, where inferior (or at least different) tools may be a restriction that nurtures creativity, or at least guarantees results that are easily distinguished from the rest of the market.
>But they are the low-to-mid segment of professionals doing weddings and local business presentations with median income around $50k
Yeah, tell that to the Octane Render community. There's plenty of incredible starving artists using consumer GPUs in their workflow to render top-notch work.
These expensive items are actually useful equipment for video production, but it doesn't mean a video card has the same importance. Who cares about slightly longer rendering times for rarely used special effects?
Isn't this just an implementation of Microsoft's DirectX Raytracing API?
If so, only video games are really going to use that API. I doubt that a software renderer (or CUDA-renderer) would leave raytracing / light calculations to a 3rd party.
There's a rumor that Dinsey bought a bunch of real-time raytracing equipment for their theme parks (the Star Wars section in Disney World). So high-quality realtime computation is needed for high-tech entertainment / theme parks / etc. etc. So there's definitely a market, even if gamers don't buy in.
Do you have a link of one of these raytracing companies explaining how they plan to use NVidia's RTX cards for this sort of thing?
As far as I'm aware, the only API to access these RTX raytracing cores is through the DirectX12 Raytracing API. I did a quick check on Khronos, but it doesn't seem like they have anything ready for OpenGL or Vulkan.
Alternatively, maybe NVidia is opening up a new Raytracing API to access the hardware. But I didn't see any information on that either.
>70-90% price increase... And prices were already inflated by crypto
If you are going off the inflated prices, this is hardly a price increase at all. 70% is the increase from 1080 Ti release price, not from the inflated crypto prices.
I meant that the release price was subsequently inflated by crypto which in the meantime cooled down significantly; if crypto were still hot, then inflated 20x0 prices would be expected; this seems more like an opportunistic move, "residual crypto-wave surfing", and resembles strategy NVidia was doing in pro segment for a while due to no competition.
The "target audience" is the hype train. If you believe Nvidia has a better long term future, that's good for Nvidia. Stuff like this trickles down to the main stream eventually.
TFlops is correllated with perfomance, but this corellation is not that perfect that you can just use TFlops as a stand-in for perfomance. Perfomance is complex. I'd wait for benchmarks.
Where are you getting these TFLOPS numbers? I haven't seen them in anything other than rumors that got a lot of things wrong so I wouldn't be too certain about them.
Unless you are already on a batch size 1 and TensorFlow greets you with a nice OOM message... Try some Wide-ResNet for state-of-art classification and play with its widening parameter.
No crypto currency that I know of has succeeded based on the utility of their proof of work algorithm producing something useful. Only proof of work algorithms in service of the currency itself have made an impact.
I'm thinking more that a blockchain enthusiast will design a new coin that abuses the raytracing hardware into running some new crypto-token just because.
This series tentatively looks like it won't be subject to the insane price hikes and scarcity of the 10 series because it's already expensive for adding more diverse hardware that doesn't immediately help you mine ethereum (or whatever's the GPU coin of choice these days). But you never know...
It doesn't need to _do_ anything as long as you can convince people that they should buy RayCoins because everyone else is going to buy RayCoins and they'd better get in on the ground floor here.
Whether it's substantially better for deep learning depends on whether they chose to neuter fp16 like they did in all other consumer GPUs. The main bottleneck is not actually the size of the model per se (sizes of the models have been mostly going _down_ as of late), it's the size of the batch during training. Both batch size and model size determine the size of the main memory hog: activations that you need to keep to do training. You don't need to keep activations to do inference: you can discard them after they're fed into the next layer. So it's not entirely accurate to say that this is useless for DL. What could make it totally kickass for AI would be a proper, non-neutered implementation of fp16, similar to what you'd find in P100, V100 and (surprisingly) Jetson TX2, which would effectively double the possible batch size, which in turn leads to quicker convergence. It could also halve the model size as an added bonus. This is NVIDIA, however, so I would bet good money fp16 is still neutered in "consumer" SKUs.
Of course, for inferencing assuming >100 TFlops on float 16 and >400 TOPS in INT4 as demoed by Jenson would be awesome for already trained/pruned models. Though for training I am not so sure, even V100 have rough edges in doing that with TensorFlow. And for model size - I guess it all depends on area in which are you are focusing; I run out of memory all the time these days even with batches of size 1 :(
If it's data size in GPU RAM you are concerned about, couldn't you store fp16 and cast into fp32 just in the kernel? In OpenCL, you would do this with vload_halfN and vstore_halfN to convert during load and store.
You won't get double throughput compared to fp32, but you shouldn't fall back to some terribly slow path either.
I haven't looked into it myself, but it could be that due to all this massaging you'd lose more on throughput than you gain on memory use. It's similar to doing 8 bit quantized stuff on general purpose CPUs. It's very hard to make it any faster than float32 due to all the futzing that has to be done before and after the (rather small) computation. In comparison, no futzing at all is needed for float32: load 4/8/16 things at a time (depending on the ISA), do your stuff, store 4/8/16 things at a time. Simple.
Right, you won't exceed the fp32 calculation rate, unless perhaps it was bandwidth-starved accessing the slowest memory tier. You are doing all fp32 operations after all. What you can do is fit twice the number of array elements into which ever GPU memory layer you decide to store as fp16, so potentially work on half as many blocks in a decomposed problem or twice the practical problem size on a non-decomposed problem.
You can separately decide which layer of cache to target for your loads and stores which convert, and then use fp32 with normalized 0-1 range math in that inner tier. You only have to introduce some saturation or rounding if your math heavily depends on the fp16 representation. The load and store routines are vectorized. You load one fp32 SIMD vector from one fp16 SIMD vector of the same size (edit: I mean same number of elements) and vice versa.
I have used fp16 buffers frequently on NVIDIA GPUs with OpenGL in generations ranging from GTX 760 (GK104) to Titan X (GM200 and GP102) as well as mobile GPUs like GT 730M (GK208M). I do this for things like ray-casting volume rendering, where the dynamic range, precision, and space tradeoffs are very important.
My custom shaders performing texture-mapping and blending are implicitly performing the same underlying half-load and half-store operations to work with these stored formats. The OpenGL shader model is that you are working in normalized fp math and the storage format is hidden, controlled independently with format flags during buffer allocation, so the format conversions are implicit during the loads and stores on the buffers. The shaders on fp16 data perform very well and this is non-sequential access patterns where individual 3 or 4-wide vectors are being loaded and stored for individual multi-channel voxels and pixels.
If I remember correctly, I only found one bad case where the OpenGL stack seemed to fail to do this well, and it was something like a 2-channel fp16 buffer where performance would fall off a cliff. Using 1, 3, or 4-channel buffers (even with padding) would perform pretty consistently with either uint8, uint16, fp16, or fp32 storage formats. It's possible they just don't have a properly tuned 2-channel texture sampling routine in their driver, and I've never had a need to explore 2-wide vector access in OpenCL.
IIRC all recent NVIDIA GPUs support fp16, it's just that the perf of fp16 is severely hampered on "consumer" grade hardware, so the memory/perf tradeoff is not viable. I mean, on the one hand I can see why that is: fp16 is not terribly useful in games. But on the other, it's probably nearly the exact same die with a few things disabled here and there to differentiate it from $7K Tesla GPUs, which as an engineer I find super tacky, much like MS deliberately disabling features in Windows Home that don't really need to be disabled.
For the money they're charging, they could have at least thrown us a bone by including HDMI 2.1 VRR, or ideally even adaptive sync.
Nvidia users still have to pay an Nvidia-tax to access expensive G-sync monitors, which are far fewer in availability and selection than Freesync (which also works with Xbox).
The cost to settle a transaction is prohibitively high on GPUs, so is mining. Crypto might be fine with ASICs after this downward movement, perhaps, but GPUs are quickly going out of the loop due to cost/performance ratio.
Exchange rates fluctuate, often it’s been ~1.05:1 which is a long way from covering the 20% VAT. Add some uncertainty, stronger consumer protection laws, and smaller markets and the price difference is relatively small.
Where in reality there shouldn't be any, since sending stuff across the pond costs nothing as seen in products like bananas. At least I live in place which has only 8% VAT, so prices look a bit more like US (but still higher for no good reason)
VAT I get - everybody pays it, but if company imports stuff (especially local branch/official distributor), they should do import paperwork - are there any additional fees included?
It's the 2-year warranty, translation of manuals and a need to have office in all major EU markets due to different legal stuff that makes it all pricier.
You dont need an office in every EU market, thats the thing about the EU, thats it one big singular (or connected) market, and if you can sell your stuff in one EU country, you can sell it in all of them. (very few restricstions apply)
Most probable explanation. 19% in Germany. So, $1199 in US, ~$1443 in Europe. 1199 * 1.19 = ~$1427. $1439 with Austrian VAT of 20%, but price is still 1259 Euro, so it seems nice. Seems about right.
Yeah; it's illegal to list prices for consumer products ex. VAT. If you're b2b you're allowed list ex-VAT but you are required to be very clear about what you're doing.
Which things exactly? Manuals translations is SO DAMN EXPENSIVE. They have to sell, what a dozen or so cards to get those losses? Or 7 european (EU) offices of which one is also a dev center vs 15 US offices (of which one is HQ/dev)? Or stock which comes to both from Taiwan?
In any case, it seems VAT is the explanation and in that case price is without premium, even great compared. Still expensive.
The EU does enforce a "2-year warranty", but it's not what you think. (Speaking as a German:) When you want to replace a broken product under the mandated warranty, then:
- In the first 6 months after purchase, the merchant must replace the product unless they can prove the defect was not present at purchase.
- After 6 months, the burden of proof reverses, and the customer must prove that the defect in question was already present at purchase.
In practice, whatever party has the burden of proof usually doesn't bother. So in effect, "6-month warranty" is a much more realistic description of this 2-year warranty.
(The fine print: Many vendors offer their own voluntary warranty on top of the mandated one. And I don't know if the rules are different in other EU countries.)
> In practice, whatever party has the burden of proof usually doesn't bother.
In practice, I've never had to prove anything within 2 years of purchase. Might be a difference between Germany and other EU countries, but somehow I doubt that.
This is probably how it is across the EU. Why antagonize your customers unnecessarily by forcing them to jump through hoops when your product fails in less than 2 years? That just leads to bad PR and reduced customer satisfaction and thus lower trust and sales.
They benefit in the perceived sense of reliability of a 2 year warranty. Why buy a product if it's going to fail in 6 months and you can't get it replaced, when the competitor is more likely to treat you fairly.
Stuff very rarely breaks in the second year. I'd bet the vast majority of warranty claims are within the first year, which is legally mandatory in every jurisdiction of note.
At NVidia's level, nearly so. They ship on the order of 50 million desktop GPUs each year. Legal hours, real estate, or translators may seem expensive, but when you count the number of graphics cards that would be required to pay for them it's not even in the ballpark of 50 million.
The warranties do cost something because they add significant risk/cost against each incremental unit, I will grant you that.
Regulations like these have a disproportionate effect relative to volume. NVidia would probably have most of those things even in the absence of the regulations, a couple guys in a garage would certainly not.
NVidia's lesson from the Crypto-boom seems to be: "Some gamers are willing to pay >$1000 for their cards".
EDIT: To be fair, NVidia is still on 14nm or 12nm class lithography (I forgot which one). So the increased die size of these new chips will naturally be more expensive than the 10xx series. Bigger chips cost more to produce after all. So if you fail to shrink the die, the economics demand that you increase the price instead.
Still, we all know that NVidia has fat margins. We also know that they overproduced the 1080 series during the Cryptoboom, and that they still want to sell all of their old cards. If they push the prices down too much, then no one will buy the old stuff.
NVidia doesn't make many cards. They mostly make chips. The "Founders edition" are an exception, but the mass market products are made by EVGA, MSI, and other such companies.
The fire-sales on EVGA 1080 Ti chips make it darn clear that there's too many 1080Ti and 1080 cards out there.
Second: these RTX 2080 chips have been in the rumor mill since June, maybe earlier. The fact that NVidia delayed until now is proof enough. NVidia has been stalling on the release of the RTX series.
>> Nvidia previously had forecast sales for cryptocurrency chips for the fiscal second quarter ended July 29 of about $100 million. On Thursday it reported actual revenue of only $18 million.
So its not entirely clear who overproduced things per se, but what we DO know is that NVidia was expecting $100 Million cards to be sold to cryptominers between April and July. Only $18 million were sold.
In any case, it is clear that there's a lot of 10xx cards laying around right now. And NVidia clearly wants to extract as much value from current stock as possible. Pricing the 20xx series very high above the 10xx series is one way to achieve what they want.
Could you cite this? I've heard the opposite: that Nvidia didn't chase after the crypto market /because/ a crash would cause problems (if they overproduced). Besides, they make plenty of money everywhere else. Furthermore, Intel has charged (edit: consumers) $1000+ for a chip before. The market will bear it, crypto or no crypto.
>> Nvidia previously had forecast sales for cryptocurrency chips for the fiscal second quarter ended July 29 of about $100 million. On Thursday it reported actual revenue of only $18 million.
That suggests that NVidia has $82+ million worth of 10xx series GPUs laying around somewhere.
Rumor is that AMD's Navi is but a minor update next year. "Next Generation" is 2020 and beyond for AMD.
So unfortunately, NVidia can bet on a lack of competition for the near future. NVidia can always drop prices when Navi comes out (if it happens to be competitive). But it seems like they're betting that Navi won't be competitive, at least with this pricing structure.
I dunno. We know Intel's roadmap: Icelake next year at 10nm (with AVX Instructions), Tiger Lake (10nm optimization), Sapphire Rapids (7nm) in 2021, etc. etc.
It seems like if you want people to buy your products, letting them know about them and the features they'll support (ex: AVX512) so the hype can build is a good thing.
About 4 years ago Nvidia also used to publish a "roadmap" that showed a somewhat fake performance versus architecture plot. They stopped doing that after Volta.
There’s a huge difference between saying “we will have something in the future” (duh) and saying “we have absolutely nothing for the next year and a half.”
The latter gives your competitor the freedom to ask any price the market will accept without having to worry about a competitor undercutting this price in some near future.
Balance between area, yields and tooling - a mature process with established tooling and strong yields can offset some of the additional wafer costs required by larger area.
Most of it is dramatically lit renderings of fans and vacuous marketing-speak of course, but there's a tidbit near the bottom of the page about NVLink I find interesting.
"GeForce RTX™ NVLINK™ Bridge
The GeForce RTX™ NVLink™ bridge connects two NVLink SLI-ready graphics cards with 50X the transfer bandwidth of previous technologies."
I guess NVLink is finally becoming relevant as a consumer technology then?
Do you think this will be the GPU generation when consumer motherboards capable of pushing things to the GPU by NVLink will appear as well?
"NVLink is a wire-based communications protocol for near-range semiconductor communications developed by Nvidia that can be used for data and control code transfers in processor systems between CPUs and GPUs and solely between GPUs"
Sounds like you're assuming operational mode A when this is going to be operational mode B.
What are the approximate gains from the last gen, i.e. 1080->2080 or 1070->2070, for a non RTX enabled game? Would be speculation at this point I assume, but just based on clock freqs, mem bandwidth and the number of units on the chip?
Why do you need major performance increases for old looking games that are already maxing out your monitor's refresh rate? Next you're gonna say programmable shaders are overkill.
The prices are obscene. The '70 tier has usually been my tier of choice, but they're sliding that up from the $400 range to the $600 range.
I actually just scored a 1080 in the realm of $400, presumably because crypto demand dived, and they had a lot of stock coming into the new generation announcement. I'm glad I got it, because it's going to be a while before I touch cards at these price points. Presumably when they filter down to a 2060 or something it'll be affordable?
I think I spent $300 for a 970 and I've been keeping an eye out for something that has 2x the performance at the same price. I guess the replacement will be the 2050?
The prices are absolutely insane. I was prepared to pay £599 for the 2080, but £749 for a non-Ti card is a joke. It feels to me like Nvidia knows there's no competition from AMD so they can ask for literally any price and people will pay it.
for a non ti card?
that's just a name.
you have a very wide array of cards, at a wide array of s prices, that can do various amounts of maths per second. pick the one you want.
Is anything known about the number of tensor cores in the new 2080 Ti? What about the float16 performance using the Tensor cores? I couldn't find this on the spec page [1].
A couple of sites have said 384 TC's, which would be approximately ~50% of the v100/new high-end Quadro card. Would like to see a confirmed source as well.
Yeah, it seems at the moment that there's no confirmation on to what degree they're actually enabled either. It may well be that NVidia will make them available for inference via DirectX so they can do AI based denoising and anti aliasing in games but artificially limit support in frameworks for training so as not to eat into the Quadro / Tesla markets. I'm holding off on ordering one until I see more info on that.
> It may well be that NVidia will make them available for inference via DirectX so they can do AI based denoising and anti aliasing in games but artificially limit support in frameworks for training so as not to eat into the Quadro / Tesla markets. I'm holding off on ordering one until I see more info on that.
Good point, that would not be cool. I'll wait with ordering until more is known about this.
Using DNN for generating higher resolution images[1] in realtime at fraction of the performance cost is quite brilliant. Is there any info it if this is nvidia proprietary technology or general feature of DirectX Raytracing that will eventually be available also from AMD and Intel?
[1] Though showing CT/medical imaging as one of it's applications is a big no-no
> Is there any info it if this is nvidia proprietary technology or general feature of DirectX Raytracing that will eventually be available also from AMD and Intel?
This is a general feature of their RTX 'pipeline', not just DirectX, so I imagine Vulkan will introduce support eventually. Whether AMD will be able to put it into GPUs is a different story, but it appears they are doing some of their own work in the area of Tensorflow so they may provide their own equivalent.
> Though showing CT/medical imaging as one of it's applications is a big no-no
Why is that? They weren't showing it as an application of their higher res image DNNs, that was just an example of where they have used DNNs in general.
I feel like there is a hard limit on rendering tech right now. We saw 4k upgrades to the PS4/xbox and next gen consoles might be a way off.
I thought the gaming industry was waiting for hardware that would let you render Marvel movie style graphics. Things that just barely hit that uncanny divide; that would give the player comic book film style immersions.
The demo video they posted does look pretty impressive. Is this the generation of video cards that will take us there, or if not, what is the next big innovation in gaming/graphics going to be?
One of the major challenges when trying to generate realistic, believable graphics is correctly modeling ambient light. Without that, adding higher resolution displays or finer polygon meshes won't get you there.
The most efficient and accurate methods of modeling ambient light are fundamentally based on ray tracing. Path tracing is probably the best known and most popular, photon mapping is another. They are both based on random sampling, so the more rays you can trace per frame, the more accurate the lighting calculations are. (Inaccuracies tend to manifest as graininess.)
That's about what I'd expect. If they're doing it sensibly, though, it should only affect the resolution of ambient light effects, not the portion of the rendering that comes from primary rays. So, the geometry would have sharp edges, but the caustics might be blurry. Seems like a good tradeoff if it drastically reduces the overall computation.
4k gaming is still very much GPU limited. The PS4 Pro and Xbox One X often do not render at the full 3840x2160, do not run at 60FPS, and have individual graphics settings turned down compared to a PC for example.
I am currently running a 1080TI with an 8700k and at 4k resolutions I still need to turn down some graphics settings to maintain 60FPS and <100% GPU usage while my CPU sits at 40-50% usage.
If the 2080TI is 50% faster than a 1080TI than perhaps it will be the card that is able to push 4k 144FPS for the faster refresh 4k monitors that are starting to come out.
Art direction has more influence on realism than hardware capability. Good visual artists can get much closer to "real" than poor artists simply by choosing what to depict.
It's not either / or. A given artist can get better results with less manual effort with path tracing. That's the main reason almost all offline rendering has moved to path tracing despite it being much more computationally intensive. Artist time is more valuable to them than CPU time.
Pixar thinks its artists have better ways to spend their time than placing lots of fake bounce lights to approximate true GI. It's not that they couldn't get some very good results given enough time and effort, it's that they can get better results leaving that stuff to the renderer and using their time elsewhere.
VR is still very limited by available GPUs. Current headsets suck (low res, small screens, low refresh rate) and current GPUs aren't enough. If/when VR headsets get better, GPUs will fall behind even more.
Realtime ray tracing will take us there and this is the first hardware generation to offer dedicated hardware acceleration for that which makes it practical for games for the first time. Don't expect things to change over night though, it will be a while before this hardware is mainstream and a while before rendering engines are designed around it. We'll see hybrid rendering approaches first and there's no upper limit on the horizon for the amount of compute that can be burned to improve visuals.
Triangles and fragments per second for rasterization grew exponentially because of growth in circuit performance . Ray tracing needs to grow say 3 orders of magnitude in performance now (at least) but do so without (that) massive advances in semiconductors. That will be a challenge.
Depends what you mean by 'needs'. The combination of dedicated ray tracing acceleration hardware and APIs to drive it with some remarkable advances in denoising, in part driven by deep learning plus associated dedicated hardware, means that we are at a threshold where hybrid techniques that mix rasterization with ray tracing are viable for games and can advance the state of the art in real time rendered image quality.
That certainly doesn't mean real time rendering is 'solved'. We could certainly use another 3 orders of magnitude compute performance (and more) if we had it to get better visuals. That's pretty much always been true for graphics though.
Yes, I'm only talking about "non-hybrid" GI - I.e. what is needed to play a game looking the same in terms of detail, sharpness etc. as a rasterized game does today, at 4k/60 or 2k/144, but using path tracing.
Hybrid (shadows, reflections only) already works today and looks pretty impressive. But just a rough guestimate of how many more rays/pixel we need to play a AAA game with PT only? I'm thinking 1000 times.
I think we'll get there with less than a 1000x increase in raw compute due to a combination of factors.
For one, ray tracing in hardware is new. We've had many years of clever techniques to improve efficiency of rasterization in hardware so overall performance has improved faster than raw compute. I think we can expect similar improvements for ray tracing.
We're also likely to see new algorithmic advances that let us get better results with fewer rays. There was a whole course at SIGGRAPH on applying machine learning to light transport problems, things like learning probability distributions for irradiance to do more effective importance sampling with promising results. Combined with the advances in denoising there's a lot of potential for improved results without more rays there. Lots of scope to take more advantage of temporal coherence I think too.
On the hardware side there is still room to scale wider for ray tracing. It's getting hard to make transistors smaller or faster bit there's still some room for putting more of them on a chip it seems.
Combine all these and I think we're a lot closer than a 1000x increase in raw compute to real time path traced games. Primary rays might still be rasterized though, it still has coherence advantages.
> But just a rough guestimate of how many more rays/pixel we need to play a AAA game with PT only? I'm thinking 1000 times.
Depends on the LOD and number of lights you want. You can get a better image than raster using only one or two shadow/reflection bounces and several lights. At 10 Giga rays, the 2080TI could do 166,666,666 rays per frame at 60fps. 4k screens have 8,294,400 pixels, so for every pixel you could have one initial ray with at maximum 19 shadow rays. Of course when you introduce a few reflection bounces etc. it becomes significantly less rays/pixel.
This would absolutely max out the RT cores to their theoretical limit, so I imagine we will stick with hybrid for a while but I think 1000x is an extreme estimate.
Yes, well I tried to extrapolate from the demos we see now, using AI denoising. It certainly does look promising, and even if there are artifacts e.g. at rapid movement, one could easily imagine just dropping to a simpler rasterizer when moving rapidly (to avoid dragging) and then refine.
These will make a good card to experiment with. I am going to purchase one when I can. More RAM which is nice for larger models, but you can do a lot with 8GB of ram, especially if you are just getting going. Do not let worrying about having the latest and greatest hardware keep you from getting started. Time is the most important resource to have, not spec sheets.
These cards look like they will have Tensor Cores, which should give a significant boost (e.g. ~50% of P100/V100 performance for certain operations at ~10% of the cost) over the 1080 series.
The Quadro's have a higher spec ram (think ECC vs. DDR) which is important for a workstation GPU, but the price difference goes up in a hurry. My suggestion would be to learn on a RTX and then you can rent a Quadro if you need more power to play with.
Why would we want raytracing as for 'gaming' as opposed to the rasterizers we have now? Is this a gimmick or are there algorithms for post processing and the like that could take advantage of this?
I believe there are a lot of “tricks” engineers have developed to accomplish rendering goals that would be more natural with ray tracing. The added simplicity would lower the barrier of entry for more complex effects.
Rasterization will always win for primary rays. It's the secondary rays that are interesting: reflections, soft shadows, ambient occlusion.
There are hacks to do the same things with rasterization and screen space image tricks but they need to be customized with the art. E.g. shadow maps can provide fake soft shadows but all shadow mapping techniques need fine tuning to look good in a certain scene.
As the other reply said, currently lighting and shadows in games is a huge mixed bag of tricks. Raycasting would make everything simpler and more robust... and slower.
This is exactly what im wondering, with the next gen consoles confirmed for AMD CPU/GPU I doubt there will be much Ray Tracing tech in the next generation, which means we're looking at 3-5 years before this tech can be mainstream.
These GPUs will be antiquated by the time Ray Tracing can gain any real traction...
I believe that lots of games already do ray tracing/casting when they render (especially for things like lighting and reflections), so there can be benefits even before games switch to 100% ray tracing.
For the most part it really isn't "ray tracing", it's _close_ but almost everybody "cheats" because of the performance gains.
Even if some games do using ray tracing for some things, the developers would have to update their game to support the features of a currently very expensive and uncommon card. Not to mention most gamers have a GTX 1060 or less according to steam to give you an idea of what the current "mainstream" is at.
I agree. It won't become mainstream until the next gen of consoles are able to take advantage, because developers will still be required handle the current generation.
However, the demand from from developers themselves could be enough for console manufacturers to make sure it is available in their next iteration.
Where have you seen that next gen consoles are confirmed for AMD?
Also, if there is Nvidia tech here that AMD can't replicate, it doesn't mean that AMD can't license it.
There is a new API. It’s currently pretty opaque. There’s a comment from a former ex-Nvidia employee I can’t find that explains that the new chips have hardware support for ray-triangle intersection tests and some sort of bounding-volume-hierarchy traversal built in.
my understanding could be wrong, but I believe doing raytracing well requires better branching architecture, which traditionally gpus are pretty bad at.
Anyone know if the raytracing cores on these things can be used to accelerate Shadertoy style raymarching stuff? This technique has been threatening to break through into the mainstream for a while now.
This is basically a rewrite of iq's primitive shadertoy using DXR, with additional, non-SDF objects like metaballs. https://www.shadertoy.com/view/Xds3zN
I'm not sure how much of a speed up you'd get from using this API though. Most of the complex shadertoy shaders are unoptimized and brute-force given the limitations of the environment, and you can render similar scenes without a raytracing API using similar techniques in more efficient ways. The talk on Claybook at GDC 2018 or Alex Evan's Siggraph 2015 talk are good examples of ways to scale procedural sign distance field rendering. They both assume dynamic scenes or user edited content, there are probably other tricks that could be used to accelerate mostly static content.
It could, but to the best of my knowledge the OpenGL(/es) Vulkan extensions for raytracing don't exist yet as standards or implementation. But most of the Shadertoy raymarching scenes could be implemented using the raytracing API and custom intersecting functions.
I'm personally curious about the performance on custom intersections with the new hardware. Shadertoy demos have a lot of flexibility and tend more towards procedural geometry, since polygonal ray intersection is hard. The new hardware raytracing will be highly optimized for polygonal geometry though, since that's the bread and butter of current realtime engines.
Guessing that you'll provide a simple bounding primitive for it to do the ray test against, and then it will fall back on more or less identical shader code to resolve the custom ray intersection.
Are these Ray Tracing Extensions and Hardware going to be exclusive to Nvidia, at least for the time being? Assuming AMD didn't know about it until now and had to work and implement it, it would be at least 3 years before a Ray Tracing hardware are included in their GPU.
And in terms of efficiency, how much better is this Hybrid Ray Tracing (HRT) compared to current Shadow Maps and reflection? I assume the latter needs a lot of fine tuning but are faster while HRT is a lot simpler and slower?
I think right now the "highest" graphics quality isn't bottleneck or most important part at all, ( Whether majority of gamers were able to go to the highest graphics quality is a different question ) we sort of reach that stage some time after Crysis. It is how to reduce the time needed for designer and developers to reach that fidelity. I assume HRT will help designer to get what they want quicker?
Because somewhere along the line, most game developers only wants to give me the best graphics looking gaming possible, and the fun and storyline part seems to have faded. Apart from Nintendo.
P.S To those calling this expensive, the $999 RTX 1080ti gets you a silicon Die Size of 775mm2, along with very fast GDDR6 memory. Even your top of the line CPU don't come close to this die size, let alone the inclusion of memory.
For the software side: All of the games demoed were using DXR (DirectX Raytracing) for which RTX is just a backend for. DXR works on any recent AMD GPU (albeit slowly due to the lack of dedicated hardware) and AMD was a hardware partner in development of DXR.
For the hardware side: AMD has not said anything about specialized hardware for ray tracing. They almost certainly have known about it for quite a while but nobody (publicly at least) knows if they ever started on this.
Is ray-tracing a very large improvement for rendering quality or do you have diminishing returns now that "conventional" rendering techniques already employ such good heuristics and approximations to make the scenes look good?
Just curious since I don't know much about the state of the art of rendering and the implication of ray-tracing (other than the fact that it's, somewhat trivially by how it works, the most accurate method).
A lot of the clever tricks employed to get effects like reflections, ambient occlusion, soft shadows and GI effects using rasterization are 'hacks' with a lot of edge cases and artifacts. They're not terribly robust. Ray tracing let's you get better quality with fewer artifacts because there's less
'cheating' going on. You can get better quality in more situations with less manual tweaking, if you can afford the performance cost.
Artifacts around the contact points of shadows and reflections and artifacts for reflections and ambient occlusion of off screen objects are places where ray tracing can give much better visual results.
Ray tracing replaces parts of the pipeline that are using heuristics and approximations that are at their limits with the "correct" solutions (because ray tracing is physically correct). The problem is that there probably aren't better techniques for these approximations that aren't marginal improvements, we've started to hit a limit of what we can achieve (with physically based rendering being the target; stylized rendering still has plenty to explore) with classical rendering. Ray tracing has always been the future of graphics, we are slowly getting there.
So it should provide a standard "generation" worth of increase in graphical quality.
Good point, but what about the comparison with the old 10-series Titan Xp (released at $1200)? The 20-series improved peak GFLOPS perf by a meager 10%. I still would have expected more from upgrading from 16nm to 12nm. I guess floating point perf wasn't a priority for Nvidia in this microarchitecture refresh...
Putting AI/ML aside, I have a question for graphics experts or/and gamers: what is the practical [1] ceiling of the current champion, the 1080ti? I was under impression that that card already was much more than most people needed. In other words, is the new generation only useful for 4K gaming?
[1] By "practical" I mean on 1440p@144Hz monitors.
VR, for one, is going to (eventually) have insane resolution and performance requirements: a flat 90 to 120fps with perhaps 4K resolution for each eye.
Even my regular 1080 has trouble with a few current-gen games.
It doesn't have to be that bad. If they can get foveated rendering to work (rendering only the part you're looking at in the highest resolution), it will dramatically cut the performance requirements.
I'm currently running a 1070ti on a 1440p@144Hz monitor and I can maintain a solid 110-120 in AAA games that aren't pieces of shit (looking at you, PUBG) while on medium-high settings.
The 1080ti is still definitely more than you need for 1440p@144hz, as long as you're willing to mess with the settings. There are so many settings that eat performance for very little visual benefit, so if you're dead set on running every graphical setting at the highest, then you probably won't always hit 144hz at 1440p with a 1080ti.
IMHO you're correct, the new generation is right now only useful for 4k gaming. In a year or two after the full 20xx range is out, newer games with better graphics will be out and my 1070ti will probably no longer give the performance I want at 1440p.
Current cards can't cope with 4K@144Hz. May be new ones will be able to.
I'm waiting for AMD to release newer cards and for upstream to support FreeSync on Linux. No point in buying high refresh rate / high resolution monitor until then.
I run far cry 5, wolfenstein 2, battlefield 1 and overwatch ultra-maxed out at 4k 60fps pretty consistently with a 4.8ghz 7700k, 1080ti thats also a bit overclocked as well.
I think the purpose of this card is less to allow for an upgrade to 4k (the boost in speed for raster stuff is pretty trivial), but rather to enable the switch to ray tracing as a rendering tech. So you should see a visual improvement in your gaming at the same resolution/hz, rather than an increase in res/hz.
For me, it's around ~45-50fps@4k for a good set of games, and often ~100-120fps@144hz for me.
1000$ is a steep price for the 2080TI, although depending on the 2080 performance it might be the only card which can and will continue to reliably hit 60fps@4k.
Not upping ram with such a huge price increase is absolutely absurd.
You can easily push the 1080 Ti to its limit with 4K or 21:9 1440p (3440x1440). It's an incredible card, but it only feels overpowered at 1080p or maybe 16:9 1440p.
Here's how a VR enthusiast should think about this news. Make an estimate for when 90Hz 1440P VR can be supported by the average laptop. I think that would be a good benchmark for the time a year before the widespread adoption of VR/AR by the general public.
I wonder how long it'll take to get these cards into a "Nano"/ITX/small form factor? Is it even going to be possible after a certain point? IMO, it's the ideal size for eGPU builds, but so far, you can only go as high as a 1080.
(Here's my hacky 1080 eGPU build over TB2: http://archagon.net/blog/2018/07/25/egpu-redux/ And just to assuage any fears, there are attractive, pre-built options of the same size for TB3! I had to go this route because TB3->TB2 adapters don't work with my model of laptop.)
I wish there was more information about the acceleration structures and how they're generated.
When I worked on hardware raytracing, my conclusion was that the actual tracing was not the difficult part. That can be very well optimized, requires very little silicon to do and is only really limited by memory bandwidth.
The difficult part for real-time raytracing of animated scenes is generating efficient acceleration structures fast enough.
It makes sense that the API for this is opaque, but it would still be nice to know what goes on behind the scenes.
They're using deep learning as fancy Anti-Aliasing mechanism, to fill gaps which GPU didn't have time to properly raytrace or to upscale parts of images. As I understood from what I seen ML there mostly for AA.
In their Siggraph presentation last week they mentioned using AI for both denoising and antialisasing (but they didn't provide anymore details about it).
It's likely mostly for denoising, I highly doubt it can have enough power to perform both denoising and antialiasing. The ray tracing presented would be sampling and then use the model to fill the gaps.
You're effectively training the network for both when training for denoising as your ground truth images are both denoised and anti aliased. With ray tracing both problems are brute forced by the same approach of throwing more samples per pixel at the problem.
Could anyone comment on the likely state of Linux drivers for these new cards?
I've anecdotally, previously been recommended to use either AMD or Intel GPUs, even when out performed by Nvidia, because their open source drivers will usually easier to deal with.
Do you have an ethical requirement to only use open source drivers? Because there are official closed source nvidia drivers for linux. That is what I have always used and they have worked well. Games work, dual monitors with weird aspect ratios work, etc.
Stats that matter for deep learning: 616 GB/s Memory Bandwidth, 11GB memory. About 27% improvement over 1080Ti, assuming mem b/w is bottleneck. 11GB is too low - business decision to segment gaming and deep learning markets.
It was about time... The old series is, well, pretty old by now, but (to me) never really seemed worth the upgrade from my r9 290 (used for Vive/VR gaming).
Maybe mid-2019 will see me dropping some money on a new GPU, once prices settled.
I'm right there with you. I'm still running a 290 as well, and haven't really been able to justify an upgrade, especially given I really only play Overwatch and Rocket League.
Can anyone suggest a 4 GPU build for a non-GPU budget of USD 1500 to 2000 (i.e., not including the cost of the 4 GPUs, which will vary depending on which GPU is chosen)?
I will add to this the cost of 4 GPUs (like GTX 1080 Ti or higher).
This is my big question too. It's pretty standard marketing: if your new product isn't impressively better than the old one on commonly used benchmarks, you invent a new benchmark where your new product is impressively better. I wonder how much of the "look how much better these cards are at ray-tracing!" actually reflects real-world improvements users will see, and how much is there to create a benchmark that shows a big improvement over existing GPUs.
I don't think it's that bad. The RTX is a whole new architecture and will have significantly more cores. The bigger question is probably just how the performance is compared to the price and power consumption.
For fp16, int8 and int4, the RTX blows the GTX out of the water. For fp32 it's a tiny increase. The memory bandwidth is about 40% higher as well. You'll always find people complaining about each release, but overall I think this is a solid card.
I am usually able to keep a card for several generations before I can't get 60fps at 1080p in games anymore. Unless you're a professional gamer I don't get the appeal of throwing that kind of money at a card.
Unless somebody invented RTX-based coin of course, then this is the minimal price...