News Arm Sues Qualcomm and Nuvia for Breaking License Agreement

[bit_user]

Yes, but if you would rip everything out of an AMD CPU that arm can't do either would that difference still be that big?

SPECbench is meant to be a diverse, real-world workload that's more focused on server & professional applications (i.e. as opposed to something like PC Mark or Geek Bench). So, probably most features & functions in a Zen 3 core are contributing in some way to its result.

You can't say "the A15 is running that benchmark just as well as X only that it is running a completely different benchmark... "

Are you claiming that the Apple A15 core didn't complete all of the individual benchmark tests? Where does it say that?

 

[TerryLaze]

SPECbench is meant to be a diverse, real-world workload that's more focused on server & professional applications (i.e. as opposed to something like PC Mark or Geek Bench). So, probably most features & functions in a Zen 3 core are contributing in some way to its result.


Are you claiming that the Apple A15 core didn't complete all of the individual benchmark tests? Where does it say that?

What do you mean where does it say that?!
What do you think that this means?
"we’re skipping these components entirely for mobile devices."
" What this means also, is that the total aggregate scores presented here are not comparable to the full suite scores on other platforms,"

One continuing issue with SPEC CPU 2017 is the Fortran subtests; due to a lacking compiler infrastructure both on iOS and Android, we’re skipping these components entirely for mobile devices. What this means also, is that the total aggregate scores presented here are not comparable to the full suite scores on other platforms, denoted by the (C/C++) subscript in the score descriptions.

 

[ikjadoon]

You can't say "the A15 is running that benchmark just as well as X only that it is running a completely different benchmark... "

Do you know how to read those charts? Did you really think AnandTech placed scores from separate tests into one chart? Your out-of-context qualifier is referring to the previous charts on that page measuring SPEC_base_rate-1, which thus only shows iOS to iOS benchmarks.

The chart I linked has ALL the CPUs only tested with the C/C++ subset. The Ryzen 5950X scores 9.60 on the full suite, but in the C/C++ subset, it's 7.29, while the A15 on the C/C++ subset scores 7.28.

Those scores are 100% comparable: it is the same test. Let's read to the bottom of the page first before "finding problems".

Yes, but if you would rip everything out of an AMD CPU that arm can't do either would that difference still be that big?

"Arm can't do"? Arm is an ISA, not a CPU. Feature blocks are dropped & added to any CPU generation. A vague "everything that's missing" isn't much of a leg to stand on.

Or is your insinuation x86 as an ISA is handicapping AMD's peak performance? In that vein, AMD can ship more Arm CPUs: who's stopping them? AMD already has an ARM license (e.g., Opteron A1100) and AMD can ship Jim Keller's Arm chip design.

While AMD and Intel are still struggling to pull up their IPC, even half-a-decade-old Arm server CPUs have solid 1T performance based on their significantly higher IPC, even with quite moderate clocks (2.4 GHz to 3.3 GHz). Turns out, plenty of CPU workloads are not at all intrinsically tied to AMD's cores, Intel's cores, nor the x86 ISA.

//

More benchmarks, here from the Ampere Altra Q80-33 (80x cores; 3.3 GHz). That Ampere CPU uses a server-centric Cortex-A76 core called the Neoverse-N1; the A76 was released in 2018. The updated Neoverse-V1 has a server-centric Cortex-X1 core, which is still a 2020 core (Cortex-X3-based SoCs will release in 3-4 months).

 

[bit_user]

What do you mean where does it say that?!
What do you think that this means?
"we’re skipping these components entirely for mobile devices."

It would've helped if you'd explained that in your prior post.

The scores they're comparing appear to be "apples to apples", in that the composite scores for non-mobile don't include those tests either.

Your contention that the few Fortran tests would bias the results in favor of x86 seems a stretch, but we can explore that. If you actually look at the tests and the Ryzen 5950X review, there's only one test missing from the SPECint set run on the A15 - 548.exchange2_r. Comparing that test with the other SPECint tests, the distribution of results looks fairly similar. It seems rather far-fetched that its inclusion would significantly skew the results towards the x86 set.

In the SPECfp benchmarks, they omitted 5/12 of the tests for mobile, but then @ikjadoon didn't quote the SPECfp composite, either.

" What this means also, is that the total aggregate scores presented here are not comparable to the full suite scores on other platforms,"

I think that was meant to caution people against comparing the raw numbers between different articles, rather than to say that they analysis within that article was flawed. If they thought their comparison methodology was so flawed, why even go to the trouble of doing it?

 

[bit_user]

AMD can ship Jim Keller's Arm chip design.

Thanks for that link. I was unaware of that interview.

I knew of the K12 (and Opeteron A1100) for a long time, so that part wasn't news to me. There are two aspects of the K12 cancellation that I think a lot of people miss.

1. AMD was fighting to stave off bankruptcy, at the time. Lisa Su pulled a rabbit out of a hat with the Hygon Dhyana licensing deal. IMO, AMD was too short on both the funds and the people to bring both CPUs to market.
2. The ARM server market wasn't where it is today. AMD would've struggled to get market traction with this CPU even more than they did with the first-gen EPYC, and even that didn't seem to have very much takeup. It wasn't until EPYC 7002 (Rome) that they saw a significant ramp. Keep in mind that server design lead times are long and the entire server market is fairly conservative - from the OEMs to the customers and ISVs.

The way I see the A1100 was primarily as a way to test market interest. If it had been well-received, that would've told AMD the ARM server market was ready for something like the K12. But, as the article points out, AMD didn't even bother following on with an already-planned 16-core A1100. That tells me the market reception was weak, indeed.

I have been expecting AMD to jump (back) into the ARM server market, at some point. However, I think they need to sense that their market penetration with EPYC is peaking. Because, if you think about the signals it sends to the market, announcing an ARM-based server product is really going to damage confidence in their x86-based EPYC. From the moment they announce an ARM-based option, EPYC is going to seem like a dead-end product that AMD has stopped investing in. That means if you're not already an EPYC customer, you're much less likely to consider switching to it.