» Current | 2021 | 2020 | 2019 | Subscribe

Linley Newsletter

Training Competitors Battle on MLPerf

July 20, 2021

Author: Linley Gwennap

For the past few years, myriad startups have promised to deliver chips that would outperform Nvidia’s GPUs in the lucrative AI training market. But none has been willing to face off against the champion in the neutral MLPerf arena—until now. Both Habana Labs (now part of Intel) and Graphcore submitted their first official results as parts of the MLPerf Training 1.0 release, and a Chinese system vendor provided scores for Huawei’s Ascend accelerator. The results, however, underscore the difficulty of competing against Nvidia.

On ResNet-50, both Habana’s Gaudi chip and Graphcore’s latest GC200 achieved only half the performance (twice the training time) of Nvidia’s high-end A100 accelerator in small to midsize configurations. Huawei’s Ascend 910 chip was more impressive but still fell 10% short in the large configurations the company tested. Google’s TPUv4 topped the ResNet charts with 14% better performance across a range of configurations. For the Bert-Large model, however, the A100 pulled ahead of the pack, and the newcomers didn’t even submit results for the other six tests.

Many companies are moving from chips and accelerator cards to developing their own systems, enabling them to better compete against Nvidia’s popular DGX line. Graphcore tested its IPU-Pod systems, which comprise four-chip rackable units. Instead of submitting its own HLS-1 training system, Habana tested a SuperMicro server with eight of its HL-205 modules installed; customers can buy this configuration from SuperMicro. Peng Cheng Labs (PCL) tested a prototype system combining Huawei’s Kunpeng 920 processor, Ascend accelerator, and Mindspore AI framework, demonstrating competitive performance using all-Chinese technology—albeit in the Research category. Google submitted scores for its TPUv4 systems in the Preview category; it says the new chip will be available to its cloud customers later this year.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »

Events

Linley Fall Processor Conference 2021
Coming October 20-21, 2021
Hyatt Regency Hotel, Santa Clara, CA
Linley Spring Processor Conference 2021
April 19 - 23, 2021
Proceedings Available
More Events »