After experiencing a few setbacks, the Dojo ExaPod supercomputer was finally launched last year, marking a significant milestone in the realm of artificial intelligence (AI) computing. Boasting a staggering 1.1 exaflops of computing power specifically optimized for AI tasks, the Dojo supercomputer represents a formidable leap forward in AI research and development.
While the inaugural Dojo supercomputer utilizes custom AI chips crafted by Tesla, CEO Elon Musk has disclosed plans for future iterations to incorporate processors from both Nvidia and AMD. Musk, discussing a new $500 million initiative for the New York Gigafactory, noted that this sum could procure approximately 10,000 H100 AI GPUs from Nvidia. However, Tesla intends to substantially surpass this figure, indicating a commitment to invest “more than that on Nvidia hardware this year.” Musk emphasized that the competitive landscape in AI demands an annual expenditure of several billion dollars.
Regarding potential procurement of AI chips from AMD, Musk simply affirmed with a “Yes,” refraining from delving into specifics. Barring an unexpected development wherein AMD produces custom chips for Tesla—an outcome deemed unlikely but not entirely implausible—the company is poised to acquire AI processors from AMD’s Instinct MI300 lineup, particularly the Instinct MI300X.
The Instinct MI300X, introduced alongside the world’s premier data center APU (Instinct MI300A), mirrors the performance of Nvidia’s H100 in training tasks while exhibiting a 1.6 times acceleration in AI inference scenarios. Unlike Nvidia’s solution, which interfaces primarily with its proprietary networking hardware, the MI300X platform supports a diverse array of Network Interface Controllers (NICs).
Predictably, Tesla intends to leverage the enhanced capabilities of the forthcoming Dojo supercomputers to advance its pursuit of developing safe, fully autonomous vehicles for both individual consumers and large-scale enterprise clients. Enthusiasts seeking insights into the historical evolution of supercomputing may find value in Stephen H. Kaisler’s “The Control Data Corporation’s Supercomputer Systems,” a comprehensive exploration of CDC’s dominance in the market throughout the late 20th century until the mid-1990s.