And Then We Have Lamini: Leveraging AMD for AI Leadership

Lamini co-founder Gregory Diamos first drew attention by announcing on LinkedIn they had secured a significant stock of GPUs and were prioritizing promising AI startups. With wait times for NVIDIA's H100 stretching over a year, this was a lifeline for many firms. What was less known was that Lamini had achieved this due entirely to their work optimizing for AMD's Instinct line.

By leveraging the flexibility of AMD's open source ROCm software and large HBM capacities of cards like the MI250X, Lamini was able to match the capabilities of NVIDIA's top models for tasks like fine-tuning massive models such as Meta's Llama 2. Through optimizations handling data isolation and model switching, Lamini could serve over 10,000 unique language models on a single server.

These advances allowed Lamini to launch their LLM Superstation earlier this year, boasting the power of 128 AMD Instinct GPUs in a customized supercomputing cluster. Customers could now book access to state-of-the-art models without the traditional wait or AWS prices that were often 10 times higher. Successful trials saw adopters like AI safety startup Anthropic transition entirely to AMD hardware via the Superstation.

AI leaders from Microsoft and Meta had already signaled adoption in their internal data centers, touting AMD's competitiveness. With independent benchmarks verifying parity between AMD and NVIDIA's training performance, plus AMD gaining an inference advantage, it's unsurprising more enterprises are exploring the ROCm ecosystem through partners like Lamini.

For AI, 2024 may be the year AMD becomes a significant player.

Sources:

https://investorplace.com/2023/10/amd-vs-nvidia-investing-in-the-ai-chip-showdown/