4 ms·

Xeons / AMD EPYCs are measured in (low single-digit) TFlops, not GFlops. AVX512 gets places, especially with Fused Multiply-accumulate (FMA) instructions and m

Xeons / AMD EPYCs are measured in (low single-digit) TFlops, not GFlops. AVX512 gets places, especially with Fused Multiply-accumulate (FMA) instructions and multiple pipelines. That the thing though: TFlops is still "weak" because GPU market crushes Linpack so hard its just not comparable. But I'd assume that your typical Xeon D will run circles around these other cores watt-for-watt. Its the GPUs that Xeon D is competing with and struggling with. But if we're talking about a bunch of CPU-style compute (non-Linpack workloads), or even the occassional matrix-multiply, its hard to beat Xeon or EPYC in cost and power-efficiency. And you certainly ain't gonna beat a Xeon or EPYC with tiny ARM chips right now.

justinclift 3y ago
> And you certainly ain't gonna beat a Xeon or EPYC with tiny ARM chips right now. Do you include Apple's M family in that list of tiny ARM chips?
- dragontamer 3y ago
  That's an edge case. But I'm mostly talking about RK1 or Rasp. Pi 5. I'm well aware of "large" ARM chips like Fujitsu A64FX ARM that did well (when it came out) in terms of Linpack performance, thanks to hugely efficient designs like HBM and 512-bit SVE. I'm not very well aware of what Apple M1 offers to the table, I know some people are experimenting with it but its a far larger chip than RK1 or Rasp. Pi. But its probably smaller than A64FX. I know that Apple M1 has only 128-bit vectors though, so that's a big penalty vs AVX512 or even the older Fujitsu A64FX. I'd expect them to be bad at Linpack / Matrix multiplication as this is where SIMD shines exceptionally well... and AVX512 on Xeon is a very good SIMD implementation. I recognize Apple M1 has multiple 128-bit pipelines that operate in parallel per core, so its better than it looks, but there's a huge power-efficiency advantage to 512-bit sized vector units in the Linpack-style matrix-multiplication code.
  - tubs 3y ago
    M chips has some matrix multipliers in hardware as well.
geerlingguy 3y ago
I thought we were talking about Xeon D (1700 series specifically called out)—the typical CPUs in that range are in the low 100s of Gflops, similar to the performance of this 4 node cluster of RK3588s. Obviously, throw a 64, 96, or 128+ core EPYC or the latest Xeons at it and that will make it cry (while gobbling up hundreds of watts doing so ;) Only the Xeon D-2700 series starts touching above 500 Gflops in that generation, and those CPUs (just the CPUs, mind you) are in the $1000-2000 range, over 100W TDP.
- dragontamer 3y ago
  Xeon D 2700 comes in 16-core flavors. At 2GHz AVX512 with 1x FMA per clock tick, that's at least 1TFlop, especially because other pipelines are still available to contribute to the processing. 2GHz x 16 SIMD (32-bits per lane) x 16 cores x 2 ops per instruction (Multiply+Accumulate) == 1TFlop or there-abouts. > Obviously, throw a 64, 96, or 128+ core EPYC or the latest Xeons at it and that will make it cry (while gobbling up hundreds of watts doing so ;) And those are closer to 4 to 10 TFlops. The bigger Xeons are 2x FMA or more instructions per clock tick IIRC. So not only are there more cores, each core can do double the work. We are talking about the "little" Xeon D with relatively low clocks. But 512-bit FMA really gets going. --------- GPUs are well into 20TFlop to 50TFlop regions. ------- Although I'm calculating ~32-bit Flops, you may have done 64-bit Flops. Its a bit ambiguous but we might not be apples-to-apples yet.

justinclift 3y ago

> And you certainly ain't gonna beat a Xeon or EPYC with tiny ARM chips right now. Do you include Apple's M family in that list of tiny ARM chips?

dragontamer 3y ago
That's an edge case. But I'm mostly talking about RK1 or Rasp. Pi 5. I'm well aware of "large" ARM chips like Fujitsu A64FX ARM that did well (when it came out) in terms of Linpack performance, thanks to hugely efficient designs like HBM and 512-bit SVE. I'm not very well aware of what Apple M1 offers to the table, I know some people are experimenting with it but its a far larger chip than RK1 or Rasp. Pi. But its probably smaller than A64FX. I know that Apple M1 has only 128-bit vectors though, so that's a big penalty vs AVX512 or even the older Fujitsu A64FX. I'd expect them to be bad at Linpack / Matrix multiplication as this is where SIMD shines exceptionally well... and AVX512 on Xeon is a very good SIMD implementation. I recognize Apple M1 has multiple 128-bit pipelines that operate in parallel per core, so its better than it looks, but there's a huge power-efficiency advantage to 512-bit sized vector units in the Linpack-style matrix-multiplication code.
- tubs 3y ago
  M chips has some matrix multipliers in hardware as well.

geerlingguy 3y ago

I thought we were talking about Xeon D (1700 series specifically called out)—the typical CPUs in that range are in the low 100s of Gflops, similar to the performance of this 4 node cluster of RK3588s. Obviously, throw a 64, 96, or 128+ core EPYC or the latest Xeons at it and that will make it cry (while gobbling up hundreds of watts doing so ;) Only the Xeon D-2700 series starts touching above 500 Gflops in that generation, and those CPUs (just the CPUs, mind you) are in the $1000-2000 range, over 100W TDP.

dragontamer 3y ago
Xeon D 2700 comes in 16-core flavors. At 2GHz AVX512 with 1x FMA per clock tick, that's at least 1TFlop, especially because other pipelines are still available to contribute to the processing. 2GHz x 16 SIMD (32-bits per lane) x 16 cores x 2 ops per instruction (Multiply+Accumulate) == 1TFlop or there-abouts. > Obviously, throw a 64, 96, or 128+ core EPYC or the latest Xeons at it and that will make it cry (while gobbling up hundreds of watts doing so ;) And those are closer to 4 to 10 TFlops. The bigger Xeons are 2x FMA or more instructions per clock tick IIRC. So not only are there more cores, each core can do double the work. We are talking about the "little" Xeon D with relatively low clocks. But 512-bit FMA really gets going. --------- GPUs are well into 20TFlop to 50TFlop regions. ------- Although I'm calculating ~32-bit Flops, you may have done 64-bit Flops. Its a bit ambiguous but we might not be apples-to-apples yet.