Technology newsroom
AMD opens the AI battle! 🚀 sends Helios server to challenge NVIDIA, claims 30% more power per token
AMD launches Helios, a new AI server that boasts 30% better performance per token than NVIDIA Vera Rubin with full specs of 72 GPUs per cabinet and has already received response from major cloud service providers.
📌 Summary of important points:
- AMD launches Helios, a fully integrated AI server in a rack specifically for large-scale AI models.
- Highlights performance-to-price that is 30% better than competitors like NVIDIA Vera Rubin NVL72 per token.
- Packed with the latest technology, including the 6th generation EPYC CPU, Instinct MI455X GPU, and Pensando network chips, totaling 72 GPUs per cabinet.
At AMD Advancing AI 2026, AMD shook up the AI industry with the launch of AMD Helios, a series of servers designed as a ready-made rack solution, or what AMD calls a “Token Factory,” built specifically to run large-scale AI models.
The most interesting point is that AMD compared the Helios' performance to its direct competitor, the NVIDIA Vera Rubin NVL72, stating that the Helios offers better performance for the price. (Performance-per-dollar) up to 30% better in token processing This is an important indicator in the world of Generative AI.
⚙️ Deep dive into the internal specs
Helios combines the latest hardware from AMD with 6th Gen AMD EPYC CPUs, 5th Gen AMD Instinct MI455X AI graphics, and AMD Pensando networking chips. The rack is divided into 18 Compute Trays, each with 4 GPUs, making up to 72 GPUs in a single enclosure, all connected via high-speed UALink over Ethernet (UALoE) technology.
In terms of processing power, Helios can deliver up to 2.9 exaflops in FP4 and 1.4 exaflops in FP8. It comes with a massive 31TB of memory and a total memory bandwidth of 1.7 petabytes per second. MI455X also supports multiple data formats including MXFP4, MXFP6, MXFP8 and FP8, competing with NVIDIA's NVFP4 data center.
☁️ Design and feedback from giants
Helios server cabinets are designed to be liquid cooled. The launch was also received by major cloud service providers, with Microsoft, Meta, OpenAI and Oracle announcing their deployment of Helios servers, confirming the performance and confidence in AMD's technology.
Helios is now in full production, with Tom Brown, co-founder of Anthropic (a famous AI development company), revealing that he has previously tested Helios. And found that customizing the model to work on AMD's platform was not as difficult as thought. A single engineer can use AI to optimize the system until it reaches its full potential in no time.
💬 How much do you think the launch of Helios will shake NVIDIA's throne in the AI market?