3 ms·

AMD enabled AVX-512 without increasing their power consumption. To do that their AVX-512 runs only half of max speed compared Intel when using 512 bit registers

by gmokki 3y ago

AMD enabled AVX-512 without increasing their power consumption. To do that their AVX-512 runs only half of max speed compared Intel when using 512 bit registers (aka max flops is same as with AVX2 256 bit registers). Still, because the new instructions in AVX-512 are beneficial in many scenarios the actual speed up is still often 10-20% in code that benefits from the new instructions. Then 4 years later AMD has the option to bump to full-speed AVX-512 if it is really needed. This is the same they did with AVX2 initially.

adrian_b 3y ago
No, this (AMD running at half speed) is a myth. For most instructions AMD runs the 512-bit instructions at exactly the same speed as Intel. Both Intel and AMD execute either two 512-bit instructions per cycle (most simpler instructions) or one 512-bit instruction per cycle (a few of the more complex instructions). The reason is that AMD executes four 256-bit instructions per cycle (and Intel only three, with a fourth 256-bit section that can be used only by the AVX-512 instructions). Both Intel and AMD pair 256-bit sections to execute 512-bit operations at half rate, but equal throughput (the effective throughput actually increases due to lower overhead). The main exception is that some of the models of Intel Sapphire Rapids, only the most expensive of them, i.e. Xeon Platinum and most of the Xeon Gold models, have a second FMA unit, so only these models are able to execute two 512-bit FMA instructions per cycle, while Zen 4 can execute only one 512-bit FMA instruction + one 512-bit FADD instruction per cycle. The second advantage of Intel is that it has double bandwidth towards the L1 cache, which is absolutely necessary for Intel, because otherwise it would have been impossible to provide the operands for two FMA instructions per cycle. While the more expensive Intel models have this FMA throughput advantage, there are also many AVX-512 instructions where AMD has lower latencies than Intel. AMD has also higher clock frequencies when all the cores are busy, and also higher memory throughput, so the result is that for many AVX-512 workloads a Zen 4 beats easily a Sapphire Rapids. The main applications where Intel wins are those that depend mainly on FMA, i.e. linear algebra operations on dense arrays, e.g. DGEMM or LINPACK.

by gmokki 3y ago

adrian_b 3y ago

No, this (AMD running at half speed) is a myth. For most instructions AMD runs the 512-bit instructions at exactly the same speed as Intel. Both Intel and AMD execute either two 512-bit instructions per cycle (most simpler instructions) or one 512-bit instruction per cycle (a few of the more complex instructions). The reason is that AMD executes four 256-bit instructions per cycle (and Intel only three, with a fourth 256-bit section that can be used only by the AVX-512 instructions). Both Intel and AMD pair 256-bit sections to execute 512-bit operations at half rate, but equal throughput (the effective throughput actually increases due to lower overhead). The main exception is that some of the models of Intel Sapphire Rapids, only the most expensive of them, i.e. Xeon Platinum and most of the Xeon Gold models, have a second FMA unit, so only these models are able to execute two 512-bit FMA instructions per cycle, while Zen 4 can execute only one 512-bit FMA instruction + one 512-bit FADD instruction per cycle. The second advantage of Intel is that it has double bandwidth towards the L1 cache, which is absolutely necessary for Intel, because otherwise it would have been impossible to provide the operands for two FMA instructions per cycle. While the more expensive Intel models have this FMA throughput advantage, there are also many AVX-512 instructions where AMD has lower latencies than Intel. AMD has also higher clock frequencies when all the cores are busy, and also higher memory throughput, so the result is that for many AVX-512 workloads a Zen 4 beats easily a Sapphire Rapids. The main applications where Intel wins are those that depend mainly on FMA, i.e. linear algebra operations on dense arrays, e.g. DGEMM or LINPACK.