Intel claims AMD misrepresented 7nm Epyc performance over Xeon

Lisa Su, CEO of AMD, revealed the new details of the company's new 7-nm Epyc processor line, named Rome, to Computex. AMD said Rome would offer a substantial improvement in performance over first-generation Epyc, thanks to the new 256-bit AVX2 registers and its higher core count. On stage, Su presented baseline information showing that two AMD 64 core Rome processors exceeded two Intel Xeon Platinum 8280s. The Xeon Platinum 8280 is a Cascade Lake processor with 28 cores (Cascade Lake). Not surprisingly, AMD's 128 core configuration has more than doubled the performance of the Intel system. The Intel platform was able to bend 9.68 ns / day, while the AMD platform was 19.6 ns / day.

Intel, however, was an exception to these benchmark results and released its own data involving AMD that did a poor comparison here.

Intel has therefore made two separate claims, in the order. First, Intel claims that AMD made a mistake by not using the appropriate optimizations for its own platform and that, if that had been the case, the performance on the Intel benchmark platform improves by 30%, from 9.68 to 12.65 ns / day. It's not enough to equal Rome, but it's still a big gain.

Second, Intel claims that its Cascade Lake-AP processors will offer much more competitive performances against Rome. The Xeon Platinum 9242 (48C / 96T clock, 2.3GHz – 3.8GHz) scores 19.9 points against 19.6 for AMD, while the Xeon Platinum 9282 (56C / 112T, 2.6 GHz – 3.8 GHz) gets a gain of 24.16, 1.23 times faster Rome platform.

Positioning the CPU

We will first discuss the question of positioning the CPU. AMD did not reveal the prices of its new Epyc processors, but it would be extremely difficult to compare them to Intel prices, even if they did. Intel does not sell Cascade Lake-AP processors separately. They can only be purchased as part of a complete Intel system, sold directly by the company. Prior to Cascade Lake-AP, rumors had suggested that some of these processors were costing $ 20,000 each. They also consume much more energy than the expected consumption of Epyc. According to rumors, the TDP classification for AMD's new 7 nm chip is 240W, which seems perfectly plausible based on the TDP distribution of the Ryzen 3000. family. Intel's Cascade Lake-AP processors, on the other hand, consume 350W and 400W in basic clock.

AMD most likely chose the Platinum 8280 as a point of comparison because it wanted to argue a price argument. We have no idea what price it will have for its future 64-core processors, but an Epyc 7601 (at 32 cores) currently costs $ 4464 at Newegg. It's not crazy to think that AMD could market a 64-core processor between $ 8,000 and $ 10,000, making the 8280 a solid choice of competitors. Intel's Cascade Lake AP may be faster core-core than Rome once optimized, but it will also cost a lot more and consume a lot more power.

Optimization claims

The other advanced claim is that AMD misrepresented Intel 's performance by not performing certain optimizations before running the test. The Intel website provides the specific build indicators to call for this performance test and also suggests other performance optimizations. AMD may not have done that.

However, in the Intel slide, a small footnote minimizes the company's argument. The first sentence reads: "Software and workloads used in performance testing may have been optimized only for performance on Intel microprocessors.

A little later, there is another important disclosure:

Intel compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not specific to Intel microprocessors. These optimizations include the SSE2, SSE3 and SSSE3 instruction sets and other optimizations. Intel does not warrant the availability, functionality, or efficiency of any optimization on non-Intel microprocessors. The microprocessor-dependent optimizations in this product are designed for use with Intel microprocessors.

If you remember the antitrust lawsuit between AMD and Intel from the mid-2000s, you may be laughing right now. In summary, the Intel compiler refused to generate SSE or SSE2 code running on AMD processors, while AMD was paying for a license to use SSE and SSE2 technology. This fact was not well known and Intel had some difficulty presenting its own compiler as the fastest x86 solution without necessarily revealing its refusal to produce code that could work optimally on non-Intel chips.

Intel's legal statements point out that Intel has absolutely no legal obligation to ensure that its products work properly on competitive hardware. And that seems to matter when we discuss the obligations of one company vis-à-vis another when it comes to optimizing a comparative analysis. .

There is a difference – a significant difference – between harm The performance of a person in a way that makes his score lower than it otherwise would be different from the tests in a default state. And according to THG, who broke the story, there is no evidence that AMD attempted to interfere with Intel's test suite or that it used other compiler flags or parameters that would look bad. to the Intel processor. Yes, AMD used Epyc as a different compiler than Intel (for example) – but look at the usual legal language and you'll immediately understand why. Intel makes no explicit promise that its compiler will produce good code for a non-Intel processor.

Intel's argument is not very strong, especially given the long-term history of compiler optimizations and its competitive advantage. Of course, there are people who would say that the manufacturers' performance comparisons should Compare the best features of the competition, but this kind of mind is something you usually only see when a company knows it has a product that will win the competition.

Always take the benchmarks provided by manufacturers with a grain of salt no matter who they come from.

Now read: