Comments on: Intel Downplays Hybrid CPU-GPU Engines, Merges NNP Into GPU https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/ In-depth coverage of high-end computing at large enterprises, supercomputing centers, hyperscale data centers, and public clouds. Sat, 18 Nov 2023 16:43:12 +0000 hourly 1 https://wordpress.org/?v=6.7.1 By: John IL https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-216446 Sat, 18 Nov 2023 16:43:12 +0000 https://www.nextplatform.com/?p=142500#comment-216446 Intel builds chips for two very different platforms. One is the desktop PC that cares less about efficiency and more with performance.
Then you have laptops which need better efficiency and less performance gains. But Intel to save money tries to marry both into one design. That does not even work well for Apple M series as the SoC is fine for low powered laptops, but with desktops that need much better performance Apple has to marry SoC’s together. AMD seems to know where the market is and it has a CPU design that transitions well from low power to high power needs. A very all around design that Intel only wishes it has. Best thing Intel could do is keep hybrid model for laptops and go back to P only cores for desktops.

]]> By: Timothy Prickett Morgan https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-211329 Wed, 19 Jul 2023 02:51:39 +0000 https://www.nextplatform.com/?p=142500#comment-211329 In reply to JayN.

Well, it is only now shipping to Lawrence Livermore, and it is far more likely to see MLPerf on Genoa+MI 300X.

]]> By: JayN https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-211328 Wed, 19 Jul 2023 01:43:39 +0000 https://www.nextplatform.com/?p=142500#comment-211328 In reply to 8^p.

“…the unbeatable MI300A…”
Still no mlperf entry??

]]> By: 8^p https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-209888 Mon, 12 Jun 2023 12:54:15 +0000 https://www.nextplatform.com/?p=142500#comment-209888 I agree with McVeigh on this. AMD has a proven, field-tested CPU (EPYC), and proven field-tested GPU (instinct), both of which are known to be top-notch, and so it makes sense to join them at the hip in an integrated package (the unbeatable MI300A). It is best to wait for Ponte Vecchio’s successor (Falcon Shores) to be properly field-tested, and known to be top-notch, before pairing it, in-package, with high-performing CPUs, that it could (otherwise) just ball-and-chain drag down into both a low perf and low eff mosquito-infested swamp of doom and malaria. For nVidia, the “world’s” expectations for the pairing with Grace is not truly as much (in my mind) as that for an Intel-Intel CPU-GPU pairing (long in coming, since Xeon Phis and the likes …), because, as long as the H100 part is super good “we”‘re all rather satisfied that they did well (very different expectations, in my mind).

Still, I think I’d buy some Gaudi arch’ed chips, just for the inspirational name, even if they are never truly completed (eg. Sagrada Familia –> 2026?).

]]> By: Matt https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-209850 Sun, 11 Jun 2023 17:53:05 +0000 https://www.nextplatform.com/?p=142500#comment-209850 In reply to JayN.

I assume you are being sarcastic? If so, I disagree. He did support the conclusion. He said that Intel was likely canceling future independent NNPs and folding them into their GPU line. He said the future of the Gaudi line is in doubt for this reason. It’s also in doubt whether developing on Gaudi today will transition easily to developing on Intel’s future GPUs. The scheduling and control systems are likely entirely different. The roadmap shows the Gaudi line ending and something being called a ‘GPU’ continuing in its stead. And Intel has left customers hanging in the past, e.g., the Xeon Phi line. So with such an environment of uncertainty over the future of the platform, it follows that “And the only companies that will buy Gaudi2 and Gaudi3 are those who are desperate…” is a reasonable conclusion.

Frankly, my point of view is that the line that Falcon Shores will “bring together the best from our Gaudi products, which includes standard Ethernet switching” and “I/O designed for scale” gives me pause in believing the suggestion that the systolic array from Gaudi will be replacing the Xe matrix engines. Perhaps actually it’s just the systems-level architecture from Gaudi that will contribute to Falcon Shores. From a software standpoint switching out Xe’s matrix engines seems like starting over from square one when it comes to Intel’s GPUs. Can Intel really afford to lose another 2 years in software ecosystem development? They’re already behind AMD let alone Nvidia. I don’t know if the mantra “it’s early days” can bear much weight any more.

]]> By: JayN https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-209758 Fri, 09 Jun 2023 06:14:49 +0000 https://www.nextplatform.com/?p=142500#comment-209758 “And the only companies that will buy Gaudi2 and Gaudi3 are those who are desperate …”

What a greatly supported conclusion.

]]> By: Eric Olson https://www.nextplatform.com/2023/06/08/intel-downplays-hybrid-cpu-gpu-compute-engines-merges-nnp-into-gpu/#comment-209755 Fri, 09 Jun 2023 05:18:15 +0000 https://www.nextplatform.com/?p=142500#comment-209755 From what I can tell, seventeen years of consistently improving their ecosystem in a compatible way allowed Nvidia to build CUDA to the point where it’s the standard for GPU compute. Since CUDA applications written 10 years ago are possible to run on modern hardware and tune, there is every reason to believe such software will have a path forward for another 10 years. Though I may be missing something, Intel’s roadmap does not look to me like it includes the consistent accelerator platforms needed for software to evolve. The swim lane for Gaudi comes to a quick end, the Rialto Bridge was recently cancelled and not long ago the Knights Corner and related MIC architectures just disappeared.

If I just bought a US$ 500 million supercomputer, I’d want the software written specifically for that system to be useful with the next supercomputer at least one generation.

]]>