Hacker News Reader

Comment by luizfelberti | Hacker News Reader

Is AVX512 really an advantage though?

Intel is infamous for severely downclocking the processor for these and other AVX/SSE family instructions, to a point where sometimes using them makes the program slower than it would be otherwise, especially if you're constantly provoking frequency switches between them and regular instructions.

AMD might not have implemented AVX512 specifically yet (there's nothing legally keeping them from doing so however, they have patent sharing agreements with Intel regarding the entire x86/x64 ISA and extensions), but what they currently DO have is all common SIMD extensions implemented (up to SSE4 and AVX2 if I'm not mistaken) without incurring any frequency penalties on clock speeds for using them.

I can live without AVX512 for now, even though I'd be happier to have it. But I would really rather not have it if it came out in the same crap implementation that Intel has.

View on HN

You can look at binning statistics for non-avx/avx2/avx512 clock speeds: https://siliconlottery.com/pages/statistics

For example, the worst 7980XEs do 4.1/3.8/3.6 GHz for each of these respectively. 0.5 GHz down clock isn't too bad. You can change these settings in your bios however you'd like on the unlocked CPUs (ie, the HEDT lineup + W3175X). I do find those down clocks are necessary; I've passed 100C with a 360mm radiator and roaring fans. AVX2 loads don't get anywhere near that hot.

But for all that, I do see a substantial benefit on many workloads from avx512 -- at least 50% better performance than what I'd get from avx2.

I definitely think it's nice to have, especially if you enjoy vectorizing code and looking at assembly. With much bigger performance wins on the line, it's more rewarding and more fun -- and you have more tools to play with, like vector masking (unfortunately, gather/scatter have been disappointing). Fun or not though, if you offer me avx512 on one hand vs twice the cores with full avx2 for the same price on the other, I'd have a hard time rationalizing avx512.

> I've passed 100C with a 360mm radiator and roaring fans

Damn, which radiator? Was there a GPU under load in your water loop?

Celsius S36. My GPU is on a separate loop and was idle at the time.

I was running benchmarks of Intel MKL's zgemm vs zgemm3m because of a Julia PR that recommended replacing the former with the latter. I don't think anything hits a CPU quite as hard as a good BLAS.

I think my thermal paste may be bad, because the CPU idles hot -- nearly 35 C. I ordered a Direct Die-X and MO-RA-420 radiator, so I'm planning on swapping the AIO for an open loop with way more radiator area and flow through the fins.

Running dgemm, that CPU would hit just a tad below 2 teraflops. I'd like to get it just over that (and run much cooler).

I personally run AMD at home, but in all the benchmarks I have seen, Intel wins by a large margin in AVX enabled tasks - often double the speed. Downclocking did not seem to be a factor here.

I think saying that Intel wins in AVX tasks is absolutely fair.

For example, I had a simulation that had to run on the CPU for reasons, but made use of AVX. Intel was consistently faster on any system I tested.

The 7nm Ryzen parts should have double the avx2 as the older parts. Zen1 has half the throughput per cycle (or twice the reciprocal throughput) when using ymm (256 bit) registers vs xmm (128 bit) in general

https://www.agner.org/optimize/instruction_tables.pdf

If you want to look at `vmov`s or arithmetic like `vadd` or `vmul`. Particularly glaring is that for moves between memory and (xmm vs ymm) Zen1 has a recirpical throughput of (1 vs 2), ie that on average it is able to complete an xmm-memory move once per cycle, and a 256-bit move one every two cycles. Skylake-X instead has 0.5 for xmm/ymm/zmm-memory. That is, it can move up to 512-bits between a register and memory twice per cycle. That is 8-times the throughput.

Arithmetic isn't as bad, but Zen1's reciprocal throughput goes from 0.5 to 1 on xmm to ymm, while Intel stays at 0.5 independent of vector size.

I haven't seen data on the 7nm Ryzen parts, but their marketing claimed it was supposed to have full width avx2, so I imagine things are different now, and that 7nm Ryzen will do just as well for avx workloads per core and clock as all the Intel parts without avx512.

EDIT: Some instructions on Intel get slower with wider vectors, like vdiv, vsqrtpd, vgather...

> sometimes using them makes the program slower than it would be otherwise

oh no, that program will be super fast.

It would make other programs that run concurrently with the AVX one slower.

Reply on news.ycombinator.com