Net-Net Conclusions: AMD Is Catching Up

Although AMD has yet to meet investor expectations, and its products still lag behind those of Nvidia, AMD remains committed to its annual accelerator roadmap. The MI350 delivers nearly four times better performance compared to the previous generation. This progress could help AMD close the gap with Nvidia in GPU performance and maintain an advantage over Nvidia in memory capacity and bandwidth, even though Nvidia continues to lead in networking, system design, AI software, and ecosystem.

However, AMD has improved its networking capabilities this year with UltraEthernet support, and plans for UALink support next year for scale-out and scale-up respectively. For the first time, AMD presented a 2026/27 roadmap featuring the "Helios" rack-scale AI system, which provides some competition to Nvidia's NVL72 and upcoming Kyber rack-scale system. At least now AMD is competing on the same field.

Oracle announced it is building a 27,000 GPU cluster using AMD Instinct GPUs on Oracle Cloud Compute Infrastructure, indicating that AMD is gaining momentum. Additionally, AMD launched ROCm 7.0 and the AMD Developer Cloud Access Program, both contributing to a stronger and more robust AI ecosystem.

The AMD MI350 Series GPUs

Despite struggles to catch up with Nvidia, customers appreciate AMD’s price/performance ratio and openness. AMD claims to offer 40% more tokens per dollar, and seven out of the ten largest AI companies are now using AMD GPUs, among more than 60 named customers.

AMD highlights its larger memory footprint as a key differentiator, reaching 288 GB of HBM3 memory with the MI350. This allows today’s large models—up to 520B parameters—to fit on a single node, offering 60% more memory than competing solutions. That translates into lower TCO for many models. The MI350 also offers twice the 64-bit floating point performance compared to Nvidia, which is crucial for HPC workloads.

The MI355 uses the same silicon as the MI300 but is binned for higher speeds and power consumption, making it AMD’s flagship data center GPU. Both GPUs are available on industry-standard UBB8 boards in air- and liquid-cooled variants.

AMD asserts—and has demonstrated via MLPerf benchmarks—that the MI355 is approximately three times faster than the MI300, and even comparable to Nvidia’s B200 GPU. However, keep in mind that Nvidia maintains leadership in AI thanks to NVLink, InfiniBand, system design, ecosystem, and software. Meanwhile, the B300 is expected to begin shipping soon.

AMD’s GPU Roadmap Becomes More Clear

AMD provided further details about its upcoming MI400 series as well. Sam Altman appeared on stage and praised the MI450, emphasizing his company’s role in shaping AMD’s engineering priorities.

The MI400 will incorporate HBM4 at 423GB per GPU and support 300GB/s UltraEthernet through Pensando NICs.

To illustrate the expected performance leap of the MI400, refer to the hockey stick growth curve below. It echoes a similar slide from Jensen Huang at GTC, showing AMD is clearly moving in the right direction.

Networking: AMD’s Missing Link

While much of the attention during AMD’s Advancing AI event focused on the MI350/355 GPUs and their roadmap, the networking advancements were arguably more exciting and impactful.

For large-scale AI, AMD is a founding member of the UALink consortium and will implement UALink support starting with the MI400 series. While the slide below makes UALink look impressive, it’s important to note that Nvidia may be shipping NVLink 6.0 around the same time—or earlier.

AMD ROCm Might Actually Start to Rock!

Let’s give ROCm its due credit. Since last year’s harsh critique of AMD’s AI software stack by Silicon Analysis, the development team has worked hard to improve performance and gain ecosystem adoption.

To showcase these improvements, AMD demonstrated over three times the inference performance using ROCm 7. Part of this improvement comes from enhancements in the open AI stack, such as OpenAI’s Triton, creating a trend that keeps Nvidia on alert.

The above is the detailed content of AMD Announces MI350 GPU And Future Roadmap Details. For more information, please follow other related articles on the PHP Chinese website!