Anthropic bets on multi-chips and computing multipliers

During its IPO roadshow, Anthropic detailed its approach to increasing its "total effective compute" and claims an advantage built on compute multipliers. The company also described a porting of Claude to AMD GPUs completed in three days with almost no human intervention.

Compute Multipliers Presented as a Decisive Asset

Anthropic presents its computational advantage as largely relying on the pace of discovery of compute multipliers. The company explains that it focuses on an indicator of "total effective compute," defined as the product of physical capacity and identified multipliers. It claims that each new multiplier increases this compute pool. To illustrate this approach, it cites a case where a model requiring 100 compute units can, with a 4x multiplier, be trained with just 25 units. Another example provided is mobilizing 100 physical units to train a model that previously required 400. According to Anthropic, these multipliers stack and provide a hard-to-replicate advantage, with each new discovery reinforcing the effects of the previous ones. The company adds that this process could accelerate as it approaches self-recursive improvement, estimating that Claude will be able to discover multipliers on its own, which would increase their number at an ever-quickening pace.

Accelerated AMD Porting by Mythos and Claude's Contribution

Anthropic indicates that it began working with AMD GPUs for the first time this year. In the past, integrating a new chip platform required months of work and involved dozens of engineers. This time, the company entrusted the trial to a system named Mythos. According to them, Claude was up and running on AMD in three days, with virtually no human intervention. Anthropic specifies that Claude ported or wrote almost 100% of the kernels during this integration.

Acknowledged Multi-Chip Execution and a Precedent on TPU

Anthropic is now training and deploying Claude on AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. The company attributes this capability to an early bet on a multi-chip strategy, in a context where, according to them, many doubted the value of operating models on different chips. It emphasizes that each chip family imposes its own software and inference stack, which are costly and complex layers, and claims to have been working within this framework for years. In September, a post-mortem of three recent issues already provided insight into this approach, reporting bugs that only affected its TPU deployments. These elements were presented in the video of its roadshow launched yesterday, in a segment hosted by co-founder Tom Brown.