Reaction to AMD's presentation: Intel corrects AMD's Turin benchmarks for Emerald Rapids

Standard correction a few days after publication? This only happened with AMD vs. Nvidia, and now the second part follows: AMD vs. Intel. The reason is that AMD's values ​​in Turin are a bit optimistic compared to Intel's

Shortly after the first test results of AMD's Epyc 9005 alias Turin were published at Computex 2024, voices emerged via Twitter/X that AMD was potentially discrediting Intel's solutions in benchmarks through software.

As a rule, these tests are conducted within a framework that always shows the advantages of one party's products rather than the other's best, and it is rare for the other party to react. AMD has now been hit for the second time. In the winter, Nvidia did not miss the opportunity to respond to AMD's information about MI300 for Nvidia Hopper. A minor exchange of blows followed, as Nvidia saw itself exploited by AMD. AMD, in turn, responded and said that everything will become faster over time. The original statements were ultimately true here too, but they had limitations.

Does AMD represent the Intel Xeon processor so poorly? Intel thinks yes

The theme now seems to be recurring quite similarly with wizards. It's about the LLM test series of AMD's Turin processor against the Intel Xeon Platinum 8592+, the flagship of the Emerald Rapids family, which was demonstrated on stage at Computex 2024 by AMD CEO Lisa Su. ComputerBase was on site and captured this in the photo.

Intel has now chosen the middle LLM test, the chatbot, which shows 5.4x higher performance than AMD, and recreated it and found some discrepancies. This explains:

Last week at Computex 2024, AMD (~54:30) demonstrated “Turin” 128C benchmarks against a 5th Gen Xeon, but the results do not reflect performance that can be achieved on a 5th Gen Xeon with improved open source software. AMD has not revealed which software package it uses to measure performance. Thanks to publicly available software, 5th generation Xeon processors outperform future AMD CPUs.

Let's take a deeper look at the Llama2-7B Chatbot use case, one of the most relevant use cases for which AMD claims 5th Gen Xeon performance is significantly poor. They chose to compare their future CPU with Intel's generally available 5th generation Xeon, which launched late last year. Figure 1 shows the performance of the updated Llama2-7B on a 5th generation Xeon, taking advantage of the claims made for “Turin”. Since there is no proof of software and SLAs used in these claims, we measured performance using publicly available software (PyTorch with Intel Extension of PyTorch) and applied a P99 latency constraint to reflect more stringent customer requirements.

5th Generation Intel Xeon delivers a real performance lead over competing next-generation CPUs in the Chatbot scenario. This is an amazing result that is 5.4 times better than AMD claims. The improved software also helps deliver a significant improvement in the inaccurate representation of 5G Xeon performance in summary and translation scenarios, 2.3x and 1.2x, respectively.

Intel Corporation

Intel also explains that some of the data on the Torino is of course still missing, as the CPU has not been officially introduced, so a daily model was chosen for the chatbot. In this regard, this does not completely refute AMD's statements, but the competitor makes it clear that the world is not black and white – and this is exactly what was the case with AMD versus Nvidia in the winter. Software is extremely important in the AI ​​environment; There are always huge jumps in performance.

Intel is confidently heading into Q3 with this knowledge, because after that Intel Granite Rapids will be released as the Xeon 6P, the “CPU for artificial intelligence.” This is also assumed to be the case from our own tests against AMD products. These matters can of course also be looked at critically.

We can't wait to show you the Xeon 6's continued CPU leadership position in AI and LLM inference performance in Q3 2024.

Intel Corporation

ComputerBase previously received information about this announcement from Intel under a non-disclosure agreement. The only requirement was the earliest possible publication date.