AMD Gains Ground in AI Chip Performance Race

AMD's MI355X AI accelerator hits 1M tokens/sec in MLPerf benchmarks, matching Nvidia's B200/B300. Strong commercial adoption and a bullish 2026 outlook fuel its competitive push.

Advanced Micro Devices has achieved a significant milestone in the competitive arena of AI accelerators, positioning itself closer to industry leader Nvidia. Recent benchmark results reveal that AMD's latest hardware has surpassed a key performance threshold, though a detailed analysis of the data also highlights areas where the challenger maintains strategic distance.

Commercial Momentum and Financial Outlook

Beyond raw technical specifications, the growing commercial maturity of AMD's platform is noteworthy. For the recent MLPerf benchmarks, nine partner companies—including industry giants Cisco, Dell, and Oracle—submitted their own test results. A novel development was demonstrated by Dell and MangoBoost, who successfully combined multiple chip generations (MI300X, MI325X, and MI355X) within a single system, showcasing the platform's high degree of scalability.

This technological progress, coupled with recent multi-billion dollar deals, has provided stability for AMD's shares. In March, while the S&P 500 declined by five percent amid geopolitical tensions, AMD's stock recorded a gain of 2.5 percent.

The company's management is scheduled to present first-quarter 2026 results on April 28. Market analysts are forecasting revenue of approximately $9.8 billion, representing year-over-year growth of about 32 percent. For the full year, Wall Street anticipates a 59 percent jump in earnings per share to $6.65. This optimism is further supported by the planned introduction of the next-generation MI400 chip family in 2026.

Breaking the Performance Barrier

The centerpiece of AMD's latest achievement is its MI355X system, tested in the most recent MLPerf benchmarks (version 6.0). In a multi-node configuration, the hardware processed over one million tokens per second for the Llama 2 70B language model, marking the first time it has crossed that major performance boundary. Even in single-operation tests, performance showed a steep upward curve. The new chip delivered more than 100,000 tokens per second, a figure that is over three times higher than its predecessor, the MI325X.

These results enable a direct comparison with the dominant rival. For the Llama 2 70B model, AMD's performance is on par with Nvidia's B200 and B300 chips. In interactive inference tasks, the MI355X even surpasses the B200 by 19 percent.

However, AMD opted not to engage in a direct contest in the significantly larger DeepSeek-R1 benchmark, a category where Nvidia has traditionally posted its strongest results. Furthermore, for the text-to-video model Wan-2.2, AMD submitted data in the "open" category, which limits direct comparability with closed systems. These choices indicate areas where the company is still strategically positioning itself against the market leader.