The math just isn't close to a 40% per-core lead in literally any single test let alone that being representative across the board.
Real-world, AMD is still losing in per-core performance on average. Close, sometimes tying, sometimes leading, but not leading by 40%.
Again: we already know what Zen2 performance looks like.
Cascade Lake-X is basically Skylake-X+, a new stepping with some tweaks and (much) higher clocks. These are not new products when it comes to expected performance.
It's straightforward to see that 40% per-core lead is purely wishful thinking from the AMD set here. Don't get your hopes up, you'll just crash the hype train yet again.
I don't know why the hype train is such a thing for the AMD crowd, this spring it was supposed 5.1 GHz and 6C12T for $99 and that crashed into the brick wall of reality hard too.
>this spring it was supposed 5.1 GHz and 6C12T for $99
Did a reasonable person actually claim this, or are you citing the most extreme of the most extreme AMD fanboys? That group isn't anywhere near representative of normal AMD customers.
Obviously his reputation is in the toilet ever since (also due to him frankly melting down about how this was really AMD's fault, and that they must have changed the lineup since he published his leak, and if his leak wasn't right they should have told him). But at the time he was extremely revered among the AMD fanbase for breaking a fair few scoops, like the IO Die design for Zen2 server and desktop chips, and he claimed this was a very well-sourced leak that he was confident in.
People also leaned on statements from Kyle at HardOCP, another long-time analyst as well as statements from Der8auer (an overclocker who works at CaseKing) that 5 GHz was quote "very realistic".
There has been a concerted effort to retcon that and pretend that he was just some crazy who nobody ever believed, but this shit completely took over hardware discussion for the month before CES, and had a strong contingent of believers all the way up to the actual launch.
Intel badly needed good competition, because they grew arrogant, lazy and stopped evolving (ECC, PCIE 4, more PCI lanes anybody?). What they got is even better, and they are OK with admitting current superiority of AMD on desktops in most use cases.
Is the situation still bad with Intel regarding side-channel attacks? The fixes for those used to drop performance for up to 30%, which render them pathetic compared to AMD which doesn't suffer from those issues. This is real performance, not some synthetic benchmarks.
Cascade Lake-X has a new set of hardware mitigations for side-channel attacks (including some which are not hardware-mitigated on AMD architectures, specifically Spectre v2 - AMD claims they are "difficult to exploit" but still enables software mitigations).
Regardless, Firestrike Physics is not a benchmark that is substantially affected by mitigations. End-user tasks in general are not affected very strongly, generally ~1-3%, up to 5% in some cases.
Phoronix tested across a spectrum of workstation tasks (not gaming) and found a geomean of about 12% performance impact for Intel and 4.5% for AMD. Linked earlier, here again:
Generally, server tasks are hit hardest because they involve lots of context switching. Gaming is typically not affected much at all. Workstation tasks fall somewhere in the middle.
I agree, but it could be that it's just a subjective feeling.
In any case if I remember correctly AMD was the first using a ~64bit arch in consumer CPUs (which I think is why the arch is called "AMD64") and the same about offering multicore CPUs ( https://www.pcworld.com/article/117654/article.html ), but Intel was then always able to quickly catch up (and, in the end, present better products).
In August 2012, Jim Keller returned to AMD, where his primary task was to design a new generation microarchitecture called Zen. After years of being unable to compete with Intel in the high-end CPU market, the new generation of Zen processors is hoped to restore AMD's position in the high-end x86-64 processor market. On September 18, 2015, Keller departed from AMD to pursue other opportunities, ending his three-year employment at AMD. In January 2016, Keller joined Tesla, Inc. as Vice President of Autopilot Hardware Engineering. In April 2018, Keller joined Intel
That doesn't give me a good feeling about the future of AMD chips, but on the other hand AMD's CEO (Lisa Su) gives me the impression of being a "no bull*hit"-person, so I do still have hope that they won't mess up things in future revisions of the architecture :)
Intel's hyperthreading nets you an extra 20%, depending on workload. AMD's Zen1 SMT, in comparison, was more like 40%. Add in something like the branch prediction working better on this problem pattern, or the cache being too small / wrong shape on the Intel, and another 20% is easy.
Honestly, it looks like you're arguing with data at this point. It really does do 40% better per core.
> Intel's hyperthreading nets you an extra 20%, depending on workload. AMD's Zen1 SMT, in comparison, was more like 40%. Add in something like the branch prediction working better on this problem pattern, or the cache being too small / wrong shape on the Intel, and another 20% is easy.
This is already measured in the existing benchmarks. You don't take the benchmarks and then add 60% to them arbitrarily.
Broadly speaking, Zen2 is slightly slower per-core to Coffee Lake (eg 3700X/3800X vs 9900K). Including whatever architectural features you care to name - that's built into the result. Skylake-X is slightly slower per-core than Coffee Lake due to the mesh architecture, let's say 5%, so perhaps roughly on par with Zen2. Nowhere near 40% different
And Fire Strike, specifically, is not one of the things that Intel suffers on (not to mention this architecture has hardware mitigations for most of the vulnerabilities). FireStrike is supposed to resemble physics processing for a game, there is not a lot of context switching (which is what Intel suffers on).
--
> Honestly, it looks like you're arguing with data at this point. It really does do 40% better per core.
Well, I'm not the one arguing that we need to be taking the benchmarks and adding 60% so it matches an anomalous data point ;)
Yes, this is absolutely an outlier or anomaly. No, it's not "arguing with data" to point out when a data point bucks a larger trend in the data as a whole.
We already know the relative per-core performance of Zen2 and Skylake-X, the difference is not 40%.
I'm not saying not to buy it. I'm just saying, this thread is going to look real embarrassing in a month or two when the 3950X doesn't have a magic 40% per-core performance gain over the existing Zen2 chips. 10980XE will be more expensive, and more power hungry... but it will be faster. Marginally.
