AMD takes on Nvidia with MI300 AI chips

For the past year, anyone building large AI models had one massive bottleneck: getting their hands on Nvidia's elusive H100 GPUs. AMD is finally making its move to break that monopoly.

AMD officially unveiled its Instinct MI300 accelerator lineup, featuring the MI300X GPU and the MI300A APU, alongside its new Ryzen 8040 mobile chips for AI-enabled laptops. The announcement makes it clear that the AI chip arms race is no longer a one-horse show.

Beating the H100 at its own game

The spotlight was firmly on the Instinct MI300X. AMD claims the accelerator packs 1.5 times more memory capacity than its predecessor, the MI250X, with substantially better energy efficiency.

AMD CEO Lisa Su declared the MI300X the highest-performing accelerator in the world. While Su claimed it matches Nvidia's H100 in training workloads, she emphasized that it shines brightest in inference. Running Meta's 70-billion-parameter Llama 2, the MI300X reportedly outperformed the H100 by 1.4 times.

Big tech buyers didn't wait to jump on board. Microsoft CTO Kevin Scott joined the stage to announce that Azure ND MI300X virtual machines are now available in preview, while Meta confirmed it will deploy MI300 hardware across its data centers.

APUs in supercomputers and AI in laptops

AMD also rolled out the MI300A, an APU designed for data centers that fuses CPUs and GPUs with unified memory, eliminating the need to shuttle data across separate chips. AMD says it offers 1.6 times the memory capacity of an H100 and a 30-fold leap in energy efficiency.

The MI300A is already in production and will power Hewlett Packard Enterprise's El Capitan supercomputer at Lawrence Livermore National Laboratory, which is expected to surpass two exaflops. Su projected AMD's total addressable data center market to reach $45 billion, though pricing has not been disclosed.

On the personal computing side, AMD announced the Ryzen 8040 processor family for laptops, featuring integrated NPUs that deliver 1.6 times faster AI processing than previous models. Companies like Dell, HP, Lenovo, and Asus will ship laptops with the chip in early 2024. AMD also opened up its Ryzen AI Software Platform for models like Whisper and Llama 2, while teasing its next-gen Strix Point NPUs for 2024.

Why it matters

For developers and enterprises, this is the first credible escape hatch from Nvidia's pricing power and supply shortages. Being able to spin up MI300X instances on Azure means teams deploying large models like Llama 2 can actually compare costs and performance rather than waiting on backordered H100s.

For the broader industry, Meta and Microsoft backing AMD signals a crucial turning point. The cloud giants do not want a single supplier controlling the infrastructure layer of generative AI.

We think AMD's strategy here is remarkably sharp. Rather than picking an unwinnable fight strictly on model training where CUDA dominates, AMD targeted memory capacity and inference speed—the exact pain points enterprises face as models enter production. That said, raw hardware specs are only half the battle. AMD still has to convince developers that its software stack is stable enough to justify switching away from Nvidia.