will-it-asic

Can you run that model on that AI ASIC? A vendor-neutral, open-source reference and calculator for AI accelerators — Google TPU, AWS Trainium / Inferentia, Intel Gaudi — compared apples-to-apples against NVIDIA GPUs, for both training and inference.

Most "will it run" tools assume a GPU and treat the question as pure VRAM math. ASICs break that assumption: a model can fit in memory yet still fail to run because the vendor compiler stack (XLA for TPU, AWS Neuron for Trainium, the Intel Gaudi software suite) doesn't support its architecture. will-it-asic combines memory math and compiler/framework compatibility, and every hardware number is grounded in an official source with a citation.

⚠️ Estimates, not guarantees. Memory figures are first-order estimates from published specs and documented formulas, with explicit, tunable assumptions. Always validate against a real run.

What's inside

Surface	What it does
Calculator (/calculator)	Enter a Hugging Face model id (or specs by hand), pick a chip + mode, get a RUNS / TIGHT / WON'T FIT / UNSUPPORTED / UNKNOWN verdict with a full memory breakdown, chips-needed, and cited reasons.
Compare (/compare)	Sortable spec matrix across every accelerator. Every cell cites its source.
Explainer (/explainer)	How ASICs differ from GPUs for training/inference, and what actually matters.
Swipe (/swipe)	"Should you use ASICs?" — swipe through decision factors for a quick recommendation.
CLI (will-it-asic)	The same engine in your terminal and CI.

Monorepo layout

packages/
  core/   @will-it-asic/core   zod schemas + the PURE verdict engine (memory, sharding, compatibility)
  data/   @will-it-asic/data   one typed + cited data file per accelerator, + curated support notes
  hf/     @will-it-asic/hf     Hugging Face Hub fetch + parse -> ModelSpec
apps/
  web/    Next.js 15 App Router site
  cli/    the `will-it-asic` command-line tool

The engine in core is pure (no React, Next, or filesystem) so it runs identically on the server, in the browser, and in the CLI — and is trivially unit-testable.

Develop

pnpm install
pnpm test         # vitest across all packages
pnpm typecheck    # tsc --noEmit across the workspace
pnpm lint         # biome
pnpm dev          # run the web app (and watchers)

Run the CLI locally:

pnpm --filter @will-it-asic/cli dev meta-llama/Llama-3.1-70B --accelerator tpu-v5e --mode inference

Adding an accelerator

The dataset is designed so a new chip is one new file. See CONTRIBUTING.md — the schema requires a source URL for every spec, and a test fails the build if any spec is uncited.

License

MIT. Data is compiled from public vendor documentation; each value links to its source.