Will your model run on an ASIC?
A calculator and reference for AI accelerators: Google TPU, AWS Trainium, and Intel Gaudi, measured against the GPUs you already know. It answers two questions, for training and inference: does the model fit, and will the compiler run it. Every number links to a source.
Will it run?
Search a model, pick a workload, and see whether it fits, how many chips it needs, the smallest configuration that holds it, and a cited cost. ASICs first.
Llama-3.1-70B, served in bf16
The same model, sized across six accelerators. Computed live by the engine that powers the calculator.
- TPU7x IronwoodTight192 GiB/chip1 chip
- NVIDIA H200 SXMRuns141 GiB/chip2 chips
- AWS Trainium2Tight96 GiB/chip2 chips
- TPU v5eTight16 GiB/chip12 chips
- Intel Gaudi 3 (HL-325)Runs128 GiB/chip2 chips
- GeForce RTX 4090Won't fit24 GiB/chipneeds 8
~159 GiB total in bf16. Quantize, change the context, or switch to training in the calculator.
What "will it run" actually means
Does it fit?
Weights and KV cache, plus gradients and optimizer state for training. Each is sized per dtype and sharded across chips with ZeRO/FSDP, then turned into a chip count.
Will the stack run it?
A model can fit in memory and still fail to compile. We check the XLA, AWS Neuron, and Intel Gaudi stacks, plus curated, cited notes for popular models.
Can you trust it?
Every spec (memory, bandwidth, compute, price) links to the vendor's own documentation. A test fails the build if any number is uncited.
Google TPU (v1 through Ironwood), AWS Trainium and Inferentia, Intel Gaudi 1 to 3, and NVIDIA GPUs from the RTX 4090 to the B200, all on one comparable schema.