Big difference with AVX-512 in the new Xeon chipset

In mid-December, Intel released a line of Xeon processors with the new Emerald Rapids architecture. A site focused on Linux Voronix We ran a series of performance tests on the top-of-the-line Xeon Platinum 8592 Plus model, to determine just how useful AVX-512 instructions are.

Our Phoronix test PC was equipped with dual processors with a total of 1TB of DDR5 memory and ran Ubuntu from a 3TB SSD. Tests performed by the site include Embree, Open VKL, Y-Cruncher and above all the Intel Open VINO machine learning test suite.

Processor performance in tests was on average about twice as high with the AVX-512 than without it, with varying degrees of difference in different tests. Open VINO uses AVX-VNNI and BF16 instructions, both of which are common in AI workflows, and in the fastest subtest, advanced vector instructions provide results 10 times faster.

Power consumption when activated by the AVX-512, according to Phoronix testing, is 10 percent higher on average, with a slightly lower average clock frequency.

according to Tom's devices The results can be interpreted to mean that Emerald Rapid's broad support for AVX-512 instructions may make Intel's circuits more viable in competition with AMD's EPYC processors, with speculation that higher performance in specific workflows – particularly AI – may have led to Microsoft's choice of its Sapphire Rapids predecessor. .

The results are also interesting given the vulnerabilities in AVX instructions discovered this summer and whose fixes resulted in severe performance degradation.