3 ms·

> unlike with Intel processors, there should be no worry of slowing down the clock for the whole processor when using these instructions with larger vectors be

> unlike with Intel processors, there should be no worry of slowing down the clock for the whole processor when using these instructions with larger vectors because they don't implement "true" 512bit registers for AVX-512. This is basically to support instruction set rather than SIMD performance boost.

doikor 4y ago
They do still get around 60% speedup up according to phoronix https://www.phoronix.com/review/amd-zen4-avx512/6 https://www.phoronix.com/review/amd-zen4-avx512/6 > On average for the tested AVX-512 workloads, making use of the AVX-512 instructions led to around 59% higher performance compared to when artificially limiting the Ryzen 9 7950X to AVX2 / no-AVX512.
- fomine3 4y ago
  Wow it looks great than expected.
adrian_b 4y ago
Actually, Zen 4 implements true 512-bit registers. See the other thread on HN, which points to https://www.mersenneforum.org/showthread.php?p=614191 https://www.mersenneforum.org/showthread.php?p=614191 It appears that Zen 4 implements around 192 registers of 512 bits. However, the latest Intel model (Sapphire Rapids) implements more 512-bit registers, around 332 such registers. For most operations, Zen 4 has the same execution units as Zen 3, it has not expanded them to 512 bits, but the execution units of Zen 3 are enough to match the 512-bit throughput of the Intel CPUs. The only exception is that Zen 4 can do one 512-bit FMA and one 512-bit FADD per cycle, while the most expensive of the server and workstation Intel CPUs have a second 512-bit FMA unit, so they can do 1 FMA + 1 FMA, instead of 1 FMA + 1 FADD. The cheaper Intel CPUs with AVX-512 are worse than Zen 4, because they can do only 1 FMA, not also an FADD, like Zen 4. On the other hand Zen 4 implements a true 512-bit shuffle unit, while the current Intel CPUs have only 256-bit shuffle units, slowing down the shuffles that must cross the register halves. As tested at that link, while there are a few cases when Zen 4 has slower implementations for some 512-bit instructions, in many cases the AVX-512 implementation in Zen 4 is better than in the current Intel CPUs.

doikor 4y ago

They do still get around 60% speedup up according to phoronix https://www.phoronix.com/review/amd-zen4-avx512/6 https://www.phoronix.com/review/amd-zen4-avx512/6 > On average for the tested AVX-512 workloads, making use of the AVX-512 instructions led to around 59% higher performance compared to when artificially limiting the Ryzen 9 7950X to AVX2 / no-AVX512.

fomine3 4y ago
Wow it looks great than expected.

adrian_b 4y ago

Actually, Zen 4 implements true 512-bit registers. See the other thread on HN, which points to https://www.mersenneforum.org/showthread.php?p=614191 https://www.mersenneforum.org/showthread.php?p=614191 It appears that Zen 4 implements around 192 registers of 512 bits. However, the latest Intel model (Sapphire Rapids) implements more 512-bit registers, around 332 such registers. For most operations, Zen 4 has the same execution units as Zen 3, it has not expanded them to 512 bits, but the execution units of Zen 3 are enough to match the 512-bit throughput of the Intel CPUs. The only exception is that Zen 4 can do one 512-bit FMA and one 512-bit FADD per cycle, while the most expensive of the server and workstation Intel CPUs have a second 512-bit FMA unit, so they can do 1 FMA + 1 FMA, instead of 1 FMA + 1 FADD. The cheaper Intel CPUs with AVX-512 are worse than Zen 4, because they can do only 1 FMA, not also an FADD, like Zen 4. On the other hand Zen 4 implements a true 512-bit shuffle unit, while the current Intel CPUs have only 256-bit shuffle units, slowing down the shuffles that must cross the register halves. As tested at that link, while there are a few cases when Zen 4 has slower implementations for some 512-bit instructions, in many cases the AVX-512 implementation in Zen 4 is better than in the current Intel CPUs.