Hardware index

Primarily used for ranking and recommendation (R&R) inference and training.

Google does not officially state TPUv7's TDP. In the blog post announcing TPUv7 , Google disclosed that TPUv7 is 29.3x more efficient than the TPUv2 in…

Individual GPUs are sometimes called Blackwell Ultra, but systems are referred to as GB300. GB300 is only available in systems containing two or more GPUs. Performance…

Blackwell Ultra in a single accelerator. Its 288 GB of fast HBM targets reasoning models, long-context inference and the largest training jobs.

The air-cooled CDNA 4 option, pairing 288 GB of HBM3e with a lower platform power envelope than MI355X.

AMD’s highest-power CDNA 4 accelerator, built for dense liquid-cooled systems and memory-heavy frontier AI workloads.

The most powerful desktop GPU ever made. 96 GB of ECC GDDR7 swallows models a 5090 cannot, which makes it the workstation card for serious local…

The rack that started the AI factory era. 72 Blackwell GPUs and 36 Grace CPUs in one NVLink domain, the deployment unit behind most 2024 to…

The humanoid robot brain. Blackwell compute and 128 GB in an edge module, already adopted by the robotics companies that matter.

The Blackwell Ultra baseboard OEMs build flagship servers around. Eight GPUs and 2.3 TB of HBM3e per system, the building block of current-gen clusters.

The current top of the stack. One liquid-cooled rack acts as a single giant GPU: 72 Blackwell Ultra GPUs and 36 Grace CPUs on one NVLink…

The card every local LLM build is measured against. 32 GB of GDDR7 runs 70B-class quantized models on a desk, no datacenter required.

A petaflop of AI compute that fits in one hand. 128 GB of unified memory runs 200B-parameter models locally, and two units pair for 405B class.

Amazon's answer to GPU scarcity. Anthropic trains on clusters of these; you rent them as Trn2 instances rather than buying cards.