Others have pointed out things like memory bandwidth, but I think there might be another thing: would people want it?
Let's say that the CPU + powerful iGPU cost 95% of what a discrete CPU and GPU cost - but now you can't buy them separately, can't upgrade them separately, etc. You're less likely to get the mix of CPU and GPU that you're looking for since you can't select them independently. Why not just package the RAM with the CPU too? Apple's done that, but I think most people don't love that because it means they can't upgrade their RAM independently.
It also places constraints on how good something could be. Let's say that you produce new GPUs every 18 months and new CPUs every 12 months. Well, now you need to synchronize them. If the new CPU is ready to go, but the new GPU is 3 or 6 or 9 months out, what should your product releases be?
By having them separate, someone can buy the latest AMD CPU even though the next-gen GPU is 6 months out. When the next GPU comes out, they can buy that and upgrade the graphics and CPU on different cycles. Syncing up different product cycles isn't always easy.
I think the reason why is that they don't think there's likely a market. With things like a PlayStation or Xbox, it's going to (pretty much) have one set of capabilities for its 7-year lifecycle. You can integrate the CPU and GPU because there's only one buyer and because the CPU and GPU release have to be synced anyway for the console's release. With PCs, the release doesn't have to be synced and there are many buyers with different priorities.
> Let's say that the CPU + powerful iGPU cost 95% of what a discrete CPU and GPU cost
The main advantage of APUs would be the costs (theoretically). If it ends up being 95% of the cost of a more traditional architecture, what would be the point?
I think the 8xxxG parts are 65 W, which means you have a better chance of being able to listen to a game even if you play without headphones. Also you'll annoy whoever else lives in that house less with the vacuum cleaner noise.
I’ve wondered this too. I would imagine that there’s a significant market for PCs with iGPUs roughly on par with those of consoles… that’s enough horsepower to play all esports titles extremely well as well as most other types of games on medium settings, which is more than good enough for a lot of people.
To work around memory issues, these CPUs would need some onboard memory which would increase costs a bit, but the tradeoff is that it’d make for simpler, cheaper low-end motherboards. One can imagine a mini-ITX board with nothing but a CPU socket and a couple of M.2 slots that’d cost significantly less than current entry-level ITX boards. A full system upgrade could be performed by simply swapping out the CPU which would be great for non-enthusiasts; without a power hungry discrete GPU, power requirements are unlikely to increase meaningfully (and in fact are likely to decrease with upgrades), so upgrading wouldn’t necessitate a PSU change. These hypothetical boxes could easily stay relevant for a decade or more.
Higher end SKUs of motherboards for this type of CPU could have the usual RAM slots (acting as a second tier of slower RAM in place of swap), PCI slots, etc.
Still feels like we need more ram bandwidth. In general I think throughput is the new moat, the new market segmentation; consumer cores now have gobs of CPU and GPU (albiet we seem to have plateau'ed), but limited PCIe and ram bandwidth. We kind of started seeing USB4 compelling some more bandwidth, but I think even Intel is no longer offering USB4 on chip in many upcoming mobile chips, so that's kind of been defeated too.
Anyhow, AMD's next Strix Point is due end of year or there-after, but then their big Strix, Strix Halo has quad channels. That'll be exciting as heck. Folks may finally get their console crushers, perhaps.
I am this close to replacing my desktop (AMD 5500(?) + AMD 7600) with a 7840U (probably in the form of a Minisforum model). Day to day, I do some web browsing and Python programming. It seems silly to have this huge footprint machine sucking down probably 50-100W constantly when I can a book-sized machine that probably idles around 10W for minor performance difference in 99% of cases.
My daily driver is an M1 Max MBP which I’m happy with, but if I were to build a Linux productivity box (no gaming, that’s handled by a different machine), something like a 7840U on an ITX board with a cooler just big enough to practically never be audible would sound pretty great.
All Ryzen 7000 have a small GPU. It is not meant for gaming but enough for productive work. You can get 12 cores at 65W with the 7900. Could be passively cooled for complete silence.
[edit] The 7900 delievers ~89% of the performance of the 7900X at 170W. 7700 & 7600 90% compared to the X counterparts.
I'm the exact same, but moving from a desktop (which I've always had) to a laptop such as the Framework laptop. I'd just like 10-20% better graphics performance than what's available without losing the 13" form factor.
An APU with a decent amount of fast on-package memory is going to be expensive. $1000 maybe? It'd be odd to pair a $1000 CPU with a $100 mobo. Maybe it's a good idea, but the market would find it confusing.
I'd expect that you could probably get okay yields at okay costs if you are running a process that's a rev or two behind and making smaller chiplets that are then wired together after testing -- like the pentium pro's cache but for main memory to get 2 / 4 / 8gb ram all on "chip"
It'd probably cost more than a normal CPU, but the trade off is much more speed and the actual computer / motherboard at that point would just be a couple USB devices. You (the CPU maker) would grab a lot more of the per-unit profit.
Ah -- that's why they don't do it; no vendor would want their milkshake drunk...
AMD does seem to be trying new stuff with Strix Point Halo, but in-package RAM seems like it'd add new physical challenges that probably need a defined market before they swing for the fences?
Because it's not just the memory chip, also the interposer it's stacked on top of, which now needs to be bigger, which means you need to find more room in the PCB (today) or a bigger silicon interposer (likely in the near future) which reduces yields, and so on and so forth. If you wanted to have more than the very limited SoC RAM, you'd also then be looking at having multiple DRAM controllers, which also adds to surface area, and so on and so forth.
Such a simple motherboard could even just put the power supply onboard, with a barrel plug or something, similar to how routers like this are configured:
So much peripheral functionality has been pushed inside the CPU that a modern desktop motherboard is from one point of view essentially a fancy DC-DC converter that happens to also handle I/O port breakout. The motherboard has to translate 12 volts at 10-ish amps from the PSU, into a highly stable 1.2-ish volts at 100-ish amps for the CPU power rail.
Many motherboards have 12V or 19V DC input barrels, I think thin-itx requires it.
Generally the keyword you want to search for is “industrial motherboard”, some server motherboards have them too. Asrock Rack and Asrock Industrial and Supermicro are great.
(I fucking love asrock rack’s design team, absolute madlads and some seriously impressive density etc. they’re out there making the designs people don’t know they want, romed8-2T and genoad8x-2T are fantastic.)
Mini-box themselves sell dc-dc converters and “picoPSUs” that do this as well, although idk if they sell one with pcie power plugs. The M350 is a surprisingly high quality case for the price, and while my first picoPSU failed almost immediately (first shutdown iirc) they warrantied it no problem. A very funny trip through some oracle branded ordering/invoicing framework. Good people, good products.
My one criticism is that there is very obviously a lot of psu noise. I had a dell laptop and my pc desktop and the picopsu on an audio push-button switch and I kept getting a ton of ground loop and couldn’t figure it out etc and finally noticed it stopped and then came back as I plugged the M350 from the audio switch. Iirc it also showed through usb dac as well. I had the DIN plug brick and I think it has a lot of noise.
Intel NUCs are also extremely high quality implementations and have a great aftermarket with brands like akasa and hdplex etc.
"Some onboard memory" for a socketed CPU is impractical. They're not going to make a die with a wide DRAM controller for on-package memory and then a second narrower DRAM controller for memory slots, especially if the latter was only going to be used for high"end systems. It'll be soldered CPU and DRAM or the usual sockets/slots; I don't see how a mixed approach would make economic sense.
On-package memory is coming to x86 laptop CPU's, it's a question of when rather than if, IMO. Apple MacBooks are killing x86 laptops, and a large part of the reason is the on-package memory. I'm sure Dell et al are screaming at Intel and AMD asking for competitive chips.
You're right that they likely won't make 2 different dies. Desktop AM5 chips will just get a package with some of the memory controller pins unconnected. The big question is whether they'll also package the full width laptop chip with on-package memory in a package for desktop that's incompatible with AM5.
If they don't, somebody is going to solder that monster laptop chip into an ITX motherboard. People will grumble about a motherboard that can't upgrade either the CPU or the memory, but if the performance is there they'll still buy it.
Intel already did on-package memory last year in a custom part for Asus that ran the memory at a higher clock speed than any other Raptor Lake part. It didn't accomplish much because the memory interface was still just 128 bits, it shipped very late, and now there are systems with Intel or AMD processors running at the same memory frequencies using memory that's off package and merely soldered on the motherboard next to the CPU as is traditional for LPDDR.
I don't think putting the memory in the package helps much with practically achievable clock speed or bus widths compared to just soldering the memory nearby. (Consoles and discrete GPUs aren't doing on-package memory despite running GDDR at significantly higher frequencies than LPDDR.) And given that, there's even less reason to expect a messy hybrid configuration with half the memory controller connected to on-package LPDDR5x and half routed to DIMM slots, whether or not it uses the AM5 socket.
I wasn't suggesting hybrid, or at least didn't mean to. Memory with 2 different speeds is not well supported by the OS unless you dedicate one chunk to the GPU and the other to the CPU. So I don't think it's a good idea for AMD or Intel to go there.
What I would suggest (and my suggestions are worth about as much as I'm getting paid to write this):
AMD should create an I/O Die (IOD) with 512 bit DRAM memory width. Then create an AM5 variant with 128 of those connected to the socket and the other 384 tied off. Then create variants with 256 bit and 512 bits connected to on-package memory and no off-package memory pins. Sell those to laptop manufacturers.
A big motivation on laptops is that slow & wide memory uses
a lot less power than fast & narrow; on-package also uses a lot less power than driving the signal between chips.
Then in early 2026 introduce AM6 with 4 channel 256 bit memory support. Sell AM5 & AM6 in parallel for a while.
Their existing chiplet approach is already horrible for laptops. Making the IO die bigger and more power hungry is not an option unless they switch to much more expensive packaging to get interconnect power down to reasonable levels.
> They're not going to make a die with a wide DRAM controller for on-package memory and then a second narrower DRAM controller for memory slots
Why not? They already do the equivalent for SRAM. The big cost of a wider memory bus is routing it through the socket and the system board, which you're not doing since only the narrower bus goes through there. The wider bus is solely within the APU package.
You could also take advantage of the additional channels -- have e.g. 8 memory channels within the package and two more routed through the socket, for a total of 10. Now if you have 8GB on the package and 16GB off of it, you have 10GB striped across all 10 channels and another 14GB striped across two.
Continuing to have memory slots also allows you to sell chips with and without on-package DRAM and use the same socket. High bandwidth memory only makes much sense if you have a strong iGPU, since CPUs are rarely memory bandwidth bound. But low end systems and high end systems wouldn't have that, only mid-tier ones would. The low end system has a small iGPU where HBM is both unnecessary and too expensive. The high end system has the same small iGPU, or none at all, because it uses a big discrete GPU.
HBM is also more expensive than ordinary memory, but Windows will sit there eating several GB of RAM while doing nothing, so having some HBM and some DDR should lower costs compared with having the combined total in HBM.
HBM would be having it bandwidth reduced significantly to sync with the on system. Thats ignoring how the timings for everything would be ruined which is a pretty important and hard to manage thing for RAM.
You use HBM or GDDR if you want very high bandwidth, like top-end discrete GPUs would get. But then the memory itself is more expensive and you want the external channels to reduce cost, so the OS bloat can go in the cheap memory and preserve the limited amount of high cost memory for what needs it.
This is notably not what Apple does -- they're just using ordinary LPDDR5 with a wide bus, equivalent to having a lot of memory channels. It gets them several hundred GB/s worth of bandwidth, similar to a midrange discrete GPU. If you were going to do that, you could put most of the channels within the package and still have two of them outside of it.
That sort of configuration would allow some flexibility. The on-package memory might have lower latency (if they're both just ordinary DDR this isn't going to be much difference if any), but if you configured the system to only interleave between the on-package memory channels then the "close" memory could achieve that lower latency. Interleaving the external channels into the same pool would have a small latency hit but increase bandwidth by e.g. 25%. Which could be configured in UEFI based on your expected workload.
The G chips are also pretty seriously crippled by the cache reduction from 32MB to 16MB. It hurts their compute performance so much they behave similar to equivalent chips from a few generations before as well. I suspect its done to hit a power target but its an unfortunate trade off making these chips a bit of a let down.
I don't think they are reluctant, their APU have started low end and been steadily [0] moving up to the performance spectrum. We're about on schedule for some powerful ones to hit the market. The 8000G is the first that has entered the conversation but I doubt it'll be the last.
Mhm, but it kinda hurts to see, that AMD is able to push out APUs powering the likes of a Playstation 5 and everything on a single chip, while on desktop you need to buy the cpu and a chunky gpu seperately.
As others have pointed out, you can't fit enough memory bandwidth though the AM5 socket to feed a powerful iGPU.
Fixing this problem kind of requires abandoning the current desktop form factor and switching to a unified module with both the CPU and soldered-on memory. Though at that point, the motherboard is doing little more than power regulation and breaking out all the IO to various ports.
At that point, does it still count as a desktop form factor?
> Fixing this problem kind of requires abandoning the current desktop form factor and switching to a unified module with both the CPU and soldered-on memory.
Does it though? What stops you from putting some HBM onto the APU package and still installing it into the AM5 socket? It wouldn't even preclude you from continuing to use the memory slots, that memory would just be slower than the on-package memory.
Such a design would be possible, but not really commercially viable.
HBM memory is expensive, it requires a huge amount of extra IO on the die and an expensive silicon interposer. And you still need to keep around the old IO for the DDR5 memory. All that drives up costs, for what would still be a mid-range GPU.
Also, current software and games doesn't know how to deal with two pools of memory that have different performance characteristics, so the hardware would be underutilised.
The design which abandons the AM5 socket and switches to using much simpler and cheaper soldered-on gddr6 memory just ends up being cheaper and avoids the thermal and area limitations of the AM5 socket, so it could probably compete with high-end GPUs. It's just a better product direction.
AMD will either stick with their current strategy of APUs with their low-end GPUs because they would rather sell both a CPU and a dedicated GPU, or they will skip straight to a new form factor. The middle ground of trying to add more memory bandwidth to an AM5 style package just doesn't make any sense.
> HBM memory is expensive, it requires a huge amount of extra IO on the die and an expensive silicon interposer. And you still need to keep around the old IO for the DDR5 memory. All that drives up costs, for what would still be a mid-range GPU.
It's a product that replaces both a midrange CPU and a midrange GPU, which together have not only all of those costs but also the cost of needing two separate packages -- a CPU package for e.g. AM5 and a PCIe package for the GPU. Putting them together costs less.
> Also, current software and games doesn't know how to deal with two pools of memory that have different performance characteristics, so the hardware would be underutilised.
Except that's exactly what they know how to do, and they do it already. They expect there to be a slower pool of memory on the CPU and a separate faster one on a GPU. The system could expose them to existing applications in the same way -- the fast memory via the iGPU and the slow memory via the CPU. That's just software.
You could even do better if you have e.g. 16GB of fast memory and your game only needs 8GB of VRAM, because then the other 8GB can be used as L4 cache for the CPU and could plausibly fit the entire working set of the game in it while the only thing in DDR5 is the OS and idle background apps.
Meanwhile newer code which is aware of the configuration could make better use of it, e.g. by not having to worry about the cost of "copying" between GPU memory and CPU memory via PCIe since they're actually both directly connected.
> The design which abandons the AM5 socket and switches to using much simpler and cheaper soldered-on gddr6 memory just ends up being cheaper and avoids the thermal and area limitations of the AM5 socket, so it could probably compete with high-end GPUs.
That isn't cheaper, because then 100% of your system memory would have to be GDDR6. In a mid to high end system that's going to be dramatically more expensive than continuing to have a socket with DDR5 memory channels, because you'd have to replace e.g. 64GB of DDR5 and 16GB of GDDR6 with 80GB of GDDR6.
Meanwhile AM5 supports TDPs up to 170W and sTR5 up to 350W. 170W is reasonably sufficient for the combination of a midrange CPU and midrange GPU -- a midrange CPU is typically ~65W and a midrange GPU ~150W, which together hypothetically exceeds 170W, but a workload that simultaneously maxes out the GPU and all cores of the CPU is uncommon. In that rare case you would simply clock them slightly lower. Knocking the TDP of a 65W desktop CPU down to that of a laptop in that rare circumstance would have a relatively minor performance impact:
350W would be sufficient at the high end for the same reason.
And if they were going to design a new socket (as happens from time to time anyway), they could give "AM6" a larger footprint and higher TDP without omitting the valuable external memory channels.
But creating a new CPU interface is expensive -- all the OEMs have to design new boards. So if your concern is cost then it makes more sense to wait until the next time you were going to do it anyway, e.g. when DDR6 becomes a thing, and do something else in the meantime.
Could be that they are selling all the capacity they could hire. In that case, they'll aim to produce just the products with the highest margins (meaning the more specialized ones).
Yeah, I know the 8x00G exists, but it's kinda too little too late.