Intel and AMD Battle in AI Workload Processing
Intel refuted AMD's claim that its 5th-generation EPYC 'Turin' processor is stronger than Intel's 5th-generation Xeon processor in handling AI workload tasks in a blog post yesterday. (June 15, TapTechNews).
Intel said that its 64-core Xeon Platinum 8529+ processor can outperform AMD's latest 128-core data center CPU in the same workload after proper optimization.
AMD released the 5th-generation EPYC processor with up to 192 Zen5 cores at Computex 2024, and the new product will be listed and sold as the Epyc 9005 series, which is designed specifically for computing, cloud, telecommunications and edge workloads.
Although these chips are expected to hit the market later this year, benchmarks shown by AMD indicate that these processors will be faster than Intel's EmeraldRapids series in AI throughput workloads.
AMD said that in running a Llama2-based chatbot, the performance of the dual-channel EPYC 'Turin' processor is 5.4 times that of Intel's Xeon Platinum 8592+ CPU.
According to Intel, AMD did not disclose the software configuration used in these benchmarks, and Intel's own tests show that Xeon chips are faster than EPYC processors when performing the same tasks.
AMD said that in Llama2-7B, the performance of the fifth-generation EPYC CPU in a dual-channel configuration (with 128 cores each) is 671 tokens per second, while the fifth-generation Xeon Platinum 8592+ chip in a dual-channel configuration (with 64 cores each) is 125 tokens per second.
TapTechNews attaches the following pictures for the news manuscript: Intel uses its Intel Extension for PyTorch (P99 Latency), and the result is also 5.4 times, but the result is exactly the opposite. The Xeon processor outputs 686 tokens per second, while AMD is 125.
Intel also claimed that its Xeon chips improve the performance in translation and summarization workloads by 1.2 times and 2.3 times respectively, compared to the benchmarks demonstrated by AMD at Computex.
