NVIDIA will have new supercomputer hardware soon after Ponte Vecchio arrives, but since the MI200 is just coming out I doubt AMD will. Intel might claim they are “targeting both ” but they are unlikely to get much traction in AI with this first generation of GPUs (in fact they are unlikely to get much traction in HPC, either).
Neither Intel nor AMD have the software stack for it. Neither Ponte Vecchio nor MI200 will be used extensively outside supercomputers. “By the time PVC comes to market we might even have new GPUs from NVIDIA/AMD as competitors.” I think the FP64 matrix operation is an extremely niche operation at the moment. I see mention of TF32, FP16, BF16, and INT8 for the matrix engines. Besides, I could be wrong, but I don’t think Intel has talked about FP64 matrix operations for Ponte Vecchio. So Ponte Veccho should have about 40 teraflops of FP64, which is close to the throughput of the MI200 accelerators.
The RPeak has always referred to vector arithmetic and I think it’s more likely that’s what they mean. I believe you are making an assumption that Aurora’s 2 exaflops is referring to matrix arithmetic. PVC would reach ~40TF matrix FP64, while MI200 reaches ~90TF matrix FP64.”
“Basic math is already showing us that PVC at 54k+ GPUs is not really competitive with MI200 series for simple FP64 matrix multiplication (aka simple HPC FEM computation). Now, we just need to get to 2022 so we can see these chips in the wild since their platforms are already being shown off! My sense is that Intel probably has some sense of what its competition is offering customers, so this may be a big opportunity for Ponte Vecchio. Time will tell a bit on that one, but the other key on this question is that Jeff noted that he expected PV to be competitive in both realms. Jeff said that he expects Ponte Vecchio to target both workloads. I teed up the question noting that NVIDIA seems keen to position its data center accelerators as AI accelerators while the AMD Instinct MI200 seems to be focused on HPC.
That makes perfect sense, but it is also a good confirmation.Īnother item I asked Jeff about was how to view Ponte Vecchio. It seems like Intel is expecting higher volumes on the non-HBM models. This is expected, but we have had some readers assume that Sapphire Rapids’s popular SKUs would be the HBM2e version. As such, it seems like Intel is expecting these to be very different segments. As one would expect, we did not give exact pricing, but he did indicate that the HBM2e equipped processors would carry a significant price premium. One of the big questions that I asked Jeff McVeigh VP and GM of HPC at Intel about was about HBM versus non-HBM Sapphire Rapids Xeon mix in the market. Still, it is good to get new disclosures. Intel is releasing information about its new chips at an excruciatingly slow pace. In the final words a quick recap of a few questions.
SC21 Intel Aurora 2 Exaflop Supercomputer That is something we knew about previously.
Intel also said that Aurora has over 54,000 Ponte Vecchio GPUs and 18,000 Sapphire Rapids processors. This is similar to how NVIDIA is selling its Redstone and HGX A100 8x GPU platforms. Intel is productizing 4, 6, and 8 GPU configurations. On the Ponte Vecchio side we get 408MB of L2 cache and lots of performance. SC21 Intel Xeon Sapphire Rapids 64GB HBM2e Given the product timing we get HBM2e as we see on accelerators like the NVIDIA A100 80GB today. We just saw HBM3 last week in our Wild Hardware of OCP Summit 2021 Speed Run. Perhaps the big new disclosure is that Intel is using 64GB of HBM2e on Sapphire Rapids. We even saw Samsung showing Intel Sapphire Rapids PCIe Gen5 Server at OCP Summit 2021 and a Flex Bodega Bay Intel Xeon Sapphire Rapids System. We have covered Intel Details Sapphire Rapids several times now. SC21 Intel Sapphire Rapids Xeon And Ponte Vecchio SC21 Intel 2022 2023 Roadmap Including Sapphire Rapids HBM And Ponte Vecchioįor 2022 though, the big ones are Sapphire Rapids for the Xeon side and Ponte Vecchio for the GPU. As no surprise, Intel expects to continue making new chips. Above that in terms of capability and integration there is Ponte Vecchio Intel’s high-end HPC GPU/ accelerator.
Intel Sapphire Rapids Update at SC21įirst, Intel has its Sapphire Rapids Xeon and the HBM model. Intel also is highlighting the Frontier supercomputer at SC21. We had the opportunity to chat with Intel prior to SC21 and have a bit more market context around the new chips. At SC21 we got additional details around Intel Sapphire Rapids, the next generation Intel Xeon platform.