The AI Hardware Market: AMD and Nvidia in the Race for Training and Inference supremacy
The 125-member MLCommons organization has resumed alternating training and inference benchmarks every three months, but this time the focus is entirely on training benchmarks. Unlike previous years, where inference has driven the most significant growth in the industry, training is now the largest market, though not by as much as inference. As a result, Nvidialove still holds the distinction of leading in this segment, often with significant earnings from their Association for Computing Machinery (ACM) shipments.
In the training world, this year marks the first time that AMD has joined the competition, while Nvidia continues to win the series for submissions in the lastlude. Each company has emphasized its strengths: AMD has demonstrated a larger memory capacity (HBM) and can run entire medium-sized models on a single chip, providing greater flexibility. Nvidia, on the other hand, has leveraged its arm and gpu-based superchip (GB200) and NVLink connections, positioning itself as the market leader for both training and inference benchmarks.
The bottom line is that AMD, by hurdles because it cannot compete with the second-highest model (Blackwell) run by Nvidia. "I’m not sure if I would even choose AMD for an 8% improvement," AMD executive said. "But if you offer a similar model but with more affordable hardware, it could be a good steal." The real question, however, remains how much better the next version (MI350) will be when it launches in early 2026, especially considering the need for higher performance and scheduling for larger clusters.
The Network Edge:.Create strengthens Position
AMD is also addressing a critical gap in its market: high-speed networking for scale-up. The company has shown great promise, particularly with its_roa model on the MI325 platform. The first version of AMDSome Low Power (sumple) runs on 16GB of HBM memory. While well-supported, there are expectations for more efficient and faster network back door support when the MI350 release is scheduled in early 2026.
At the same time, Nvidia is claiming a lead with superior performance in inference processing,农强 Environmental Products spokesperson mentioned that the GB200 NVL72 "can outperform Hopper by 30x for inference processing," but only about 2.5X "in XT burning time." This is still a significant efficiency gain, but training efficiency is relatively lower due to precision costs.
Nvidia’s achievements are further boosted by having already accepted up to nearly 2500 GPUs in its submissions, which is a haven for firms seeking large-scale devices. However, Nvidia’s reliance on-arm and AVX-512 for inference processing, especially on the older GB200 (predecessor to the latest Kyber-based NVL576), could delay these claims.
Data rows: The Trade-offs
AMD and Nvidia both claim to dominate their respective markets, but their trade-offs make it challenging to determine which is in the winners’ circle. While AMD can now outperform the top €200 model (H200) of Nvidia, it will not gain a competitive edge until the MI350 launch in late 2026 or earlier. This presents a "preemptive" opportunity, but with significant potential laggards.
AMD’s focus on training hardware is currently leading, offering better performance with lower costs, albeit at the cost of slower connectivity for scalability. Nvidia’s strength in inference processing is a vir机床 to this making the two companies stand in opposition. While training hardware can still face growing adoption pressure, Nvidia’s achievements highlight its ability to innovate in inference processing, which is critical for large-scale machine learning at scale. This means that competitors in the training market are still far behind, but as precisely as those in inference processing can bring their chips to market, new Alternatives could emerge faster.
Looking Ahead: The Counterbalance of Network Back Door
The MI325 report highlights AMD’s readiness to take the next big step to compete with Nvidia’s H200, particularly with the added networking support expected by the MI350 release. To do so, AMD will need to overcome the hurdle of adopting ROCm, which was already receiving significant support with up to fifty benchmark submissions to the market.
Nvidia, meanwhile, expects its new Kyber-based NVL576 with NVLink7, Vera CPU, and improved Rubin GPU to supplant the Hopper CMS_edge when it launches next year. This could bring new and higher-end production-grade capabilities, but it may delay all of the AMD-competitions until late 2026.
In the end, both companies are choosing between two dominant technologies: training with one edge and inference with another. As the field continues to evolve, the ability to address both training and inference challenges in a balanced way will be key to maintaining competitiveness. Whether AMD is ready to take over in time for next year or if the industry will still lean heavily in Nvidia’s favor, the path to winning the future depends on both companies and the broader AI hardware landscape.
