4 ms·
It kind of feels like this is just Intel's marketing machine. The chip is less impressive than the article makes it sound, Nvidia shipped the A100 in 2020. This
by rubyn00bie 6y ago
It kind of feels like this is just Intel's marketing machine. The chip is less impressive than the article makes it sound, Nvidia shipped the A100 in 2020. This Intel chip doesn't even exist in a production system... and the A100 is already pretty damn close, hitting (w/ FP16) 624 TF with sparsity according to Nvidia's documentation which is at least as accurate as an unreleased data center chip from Intel:
https://www.nvidia.com/en-us/data-center/a100/ https://www.nvidia.com/en-us/data-center/a100/
I'd guess, by the time Intel actually ships anything useful, Nvidia will have it made it mostly moot.
- my123 6y agoGiven the _actual_ performance metrics that they gave about its Xe HP cousin (which Intel didn't publish any indication on it having FP64 at all), I'm inclined to believe that the 1PF number is indeed some very ML-specific stuff. https://cdn.mos.cms.futurecdn.net/BUsZ5EdKUcP8mWRKypTNB4-970-80.jpg https://cdn.mos.cms.futurecdn.net/BUsZ5EdKUcP8mWRKypTNB4-970... When excluding ML... (because that's what Intel gave actual metrics on for Xe-HP) 41Tflops FP32 with 4 dies. For comparison, an RTX 3090 (arch whitepaper at https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/a...) has 35.6Tflops FP32, with a single die.
- Veedrac 6y agoTFLOPS-equivalent-with-sparsity is not real TFLOPS. The article compared to A100's 312 TFLOPS which is much more reasonable.
- my123 6y agoIntel didn't say what the TFlops number was about at all, we just don't know anything other than the headline.
- dragontamer 6y agoWe know the performance targets of the Aurora supercomputer though. The only way Intel reaches those performance targets is to outdo the current crop of GPUs: MI100 (from AMD) and A100 (NVidia). Not that that is a guarantee that we got a winner here, but we know the goal for what Intel is shooting for at least.
- arcanus 6y agoThe performance targets for Exascale are based on FP64, not ML-flops.
- xbmcuser 6y agoIt is I was amazed how with the new ceo bombastic statement for some reason magically all intel problems went away and their statements about 7nm and 10nm were believed. They are going back to tick tock cadence?. Intel are such masters of marketing in 2016 they already knew 10nm fabs were not going to work so they announced stopping tick tock.
- refulgentis 6y agoI don't think that's the case - I haven't run into many people who think all of Intel's problems went away.
- shaklee3 6y agoNvidia shipped the A100 in 2019. I was using one in August.