At Advancing AI in San Francisco, OpenAI and AMD announced an extensive collaboration. Among other things, the partnership will strengthen the future MI500 GPU series and driver optimization. Above all, it benefits both OpenAI and AMD to work together to reduce Nvidia’s influence.
The buzzword at Advancing AI is “co-design.” AMD seems primarily focused on ensuring that AI companies get the product they actually want. GPUs need to be flexible, but certain architectural choices can significantly influence whether new chip generations actually achieve their theoretical performance gains. In fact, performance in AI training and inference is by far the most important metric for GPUs in the data center.
Months of access
OpenAI has had access for months to Helios, AMD’s new rack-scale system designed to compete with Nvidia’s Vera Rubin NVL72. Essentially, this is a single AI system with terabytes of memory for running the largest LLMs.
Read also: Helios: AMD scales up to a rack-scale AI system
“I need more computing power,” OpenAI VP of Compute Strategy Sachin Katti told AMD CEO Lisa Su on stage at Advancing AI. But determining exactly what kind of computing power is needed is becoming increasingly complex. According to Katti, CPUs, GPUs, networking, and cooling must all be designed simultaneously to ensure optimal functionality.
Vague but clear
One benefit of the partnership so far is that AMD, working with OpenAI, has been able to detect and patch bugs in Helios as early as possible. The optimization is also likely much more robust than in previous AMD generations.
More importantly, the question is whether this will yield further benefits in the long run. In any case, OpenAI is also turning to AI to improve its AI infrastructure. The goal is clear: to resolve the issues surrounding unstable software that made AI training on AMD systems practically impossible for a long time.
Nevertheless, it remains unclear exactly how the partnership will improve AMD’s position. For now, Nvidia won’t be too concerned, as it has already sold every GPU months or even years in advance. In the longer term, it is in the interest of both AI model developers and AMD to become a formidable competitor.