31–40 of 43 posts
Yeah seems like competing with your customers is a bad idea.
I think the only ARM licensee going for the hyperscaler CPU market is Ampere. Amazon and Microsoft make CPUs for themselves and Nvidia’s are aimed exclusively at AI workloads driving their GPUs.
Yeah seems like competing with your customers is a bad idea.
I think the only ARM licensee going for the hyperscaler CPU market is Ampere. Amazon and Microsoft make CPUs for themselves and Nvidia’s are aimed exclusively at AI workloads driving their GPUs.
Earlier quoted context omitted.
I think the only ARM licensee going for the hyperscaler CPU market is Ampere. Amazon and Microsoft make CPUs for themselves and Nvidia’s are aimed exclusively at AI workloads driving their GPUs.
AWS Graviton and Microsoft Cobalt are arm-based.
It really is a choice for Arm to use their 2023 based mobile X4 cores instead of their current C1 Ultras for this. Hopefully they step up quickly
Validating a core to server standards takes significantly longer. V4 cores should be out this year using X925 and C1 Ultra-based V5 will probably be 2027-2028. I suspect that X4 is already fast enough to beat EPYC in per-core performance when using the whole chip. ARM caught up/passed x86 in IPC all the way back around A77/78 in 2019-2020. They are now much faster per clock and hitting about the same all-core clocksp…
However, Neoverse V3 has a lower die area, so you could implement more cores per socket than with Zen 5, but this has not been done yet, as these new CPUs have only 136 cores per socket versus 192 cores per socket for Zen 5.
For programs that do not use array operations, i.e. which do not use AVX/AVX-512 instructions, Neoverse V3 has better performance per watt than Zen 5. But that changes for programs that benefit from AVX/AVX-512, where Zen 5 has better performance per watt.
Moreover, Zen 5 is already old. By the end of the year there will be Zen 6, which will be the real competitor for these new Arm CPUs, and Zen 6 will have better performance per watt, even more cores per socket and even more performance per core.
However, the claim made by Arm: "the Arm AGI CPU, for agentic AI infrastructure, delivering more than 2x performance per rack compared with x86 platforms" is obviously false.
The new Intel Clearwater Forest Xeon processors use Darkmont cores, which have approximately the same performance per core, the same die area per core and the same power consumption per core as the Neoverse V3, but Intel offers 288 cores per socket and 576 cores per board, in comparison with only 136 cores per socket for Arm.
Therefore there is no chance that these new Arm processors can provide more performance per rack than Intel Clearwater Forest.
For applications that benefit from array operations, the AMD Zen 5 compact cores have much more performance per core than Neoverse V3 and AMD has provided 192 cores per socket for a long time. There is no chance for the new processors to exceed the performance per rack of Zen 5, but for those applications that do not benefit from array operations, these new Arm CPUs should have better performance per watt than Zen 5. But by the end of the year AMD should have Zen 6 Epyc CPUs, with more cores per socket, enhanced performance per core and improved performance per watt, so then there would be even less opportunities for these Arm CPUs to be better at something.
The only way how the claim of Arm can be true is if they have compared their new CPUs with antiquated CPUs like the Intel Granite Rapids Xeon CPUs, instead of comparing with state-of-the-art Intel Clearwater Forest and AMD Zen 5.
Earlier quoted context omitted.
I think the only ARM licensee going for the hyperscaler CPU market is Ampere. Amazon and Microsoft make CPUs for themselves and Nvidia’s are aimed exclusively at AI workloads driving their GPUs.
AWS Graviton and Microsoft Cobalt are arm-based.
Ampere had previously used cores designed by Arm, but their latest CPUs (which do not impress much) use a custom core, like the Apple and Qualcomm CPUs.
Earlier quoted context omitted.
I wonder what the people at Ampere are thinking right now
Agreed. The ARM AGI CPU supports a newer version of the vectorized instructions and has matrix math extensions that the AmpereOne M doesn’t. Also has almost twice the memory bandwidth. One paper at least, the AGI CPU seems like a better choice for AI workloads. Ampere is really pushing the AI workload use cases for the AmpereOne M, so this really makes their lives a lot harder.
Neoverse V3 is the server version of Cortex-X4 and it is an Armv9.2-A CPU, with SVE2, but without SME/SME2.
The matrix math extensions, i.e. SME/SME2, are present only in the latest generation of Arm cores (the C1 cores), which implement the Armv9.3-A ISA, and also in recent Apple cores.
Neoverse V3 is also used in AWS Graviton5 and in a few NVIDIA products, e.g. in Thor, and it is also pretty much equivalent with the Skymont and Darkmont Intel E-cores, which are used in the Lunar Lake, Arrow Lake, Panther Lake and Clearwater Forest CPUs.
These processors will have very decent performance for many applications, very similar to that of AWS Graviton5, but with more cores per socket. However, the claim made by Arm: "the Arm AGI CPU, for agentic AI infrastructure, delivering more than 2x performance per rack compared with x86 platforms" is obviously false. The new Intel Clearwater Forest Xeon processors use Darkmont cores, which have approximately the sam…
IIRC Granite Rapids is also not _that_ old, and either current or a single generation behind. (Has its successor landed yet? IIRC GNR is the same generation as Sierra Forest).