feat(calc): accelerator schema — add powerCapW (TDP) + tier classification

[thinkingfish]

Summary

Two SKU-characteristic fields added to the accelerator schema (bundled — both touch every entry):

• `powerCapW?: number` on `AcceleratorVariant` — per-variant TDP in watts. Optional (skipped where vendor doesn't publish per-chip TDP). Placed on the variant level because SXM/PCIe/OAM form factors of the same chip ship with materially different thermal envelopes (e.g., H100 SXM = 700W vs PCIe = 350W).
• `tier: 'datacenter' | 'pro' | 'consumer'` on `AcceleratorSpec` — required. Drives downstream UI work (optgroup rendering in pickers, Info-tab catalog filter) that ships in a follow-up PR.

This PR is schema + data only. UI hookup deferred to keep review surface tight.

Tier breakdown — 38 accelerators classified

• datacenter (16): h100, h200, a100, l4, b100, b200, gb200, mi300x, mi325x, gaudi-2, gaudi-3, tpu-v5p, tpu-trillium, trainium-2, inferentia-2, cerebras-wse3
• pro (5): l40s, rtx-pro-6000, radeon-ai-pro-r9700, radeon-pro-w7900, radeon-pro-w7800
• consumer (17): rtx-5090, rtx-5080, rtx-4090, rtx-4080, rx-9070-xt, rx-9070, rx-7900-xtx, rx-7900-xt, rx-7900-gre, rx-7800-xt, rx-7700-xt, m3-pro, m3-ultra, m4-pro, m4-max, m5-pro, m5-max

TDP backfill — 33 of 52 variants

Vendor-direct citations in inline comments. Dropped 19 variants where TDP isn't publicly per-chip:

• Apple (14) — SoC thermals vary by chassis; no per-SoC TDP comparable to discrete GPU TBP
• TPU v5p, Trillium — Google publishes compute/HBM but withholds TDP
• Trainium 2, Inferentia 2 — AWS publishes instance-level aggregates only
• Cerebras WSE-3 — quotes ~23 kW per system; no meaningful per-chip TDP

Judgment calls to verify

1. B200 = 1000W per GPU — picked the HGX B200 SXM module spec (Lenovo's product name literally is "HGX B200 180GB 1000W"). The 1200W figure applies to GB200 NVL72 — placed there.
2. GB200 per-GPU = 1200W — cross-source consensus (Tweaktown / Wccftech / SemiAnalysis), not vendor-direct.
3. B100 = 700W — SemiAnalysis Blackwell deep dive notes B100 sized to fit HGX H100 thermals.
4. H100 PCIe 94GB at 350W — H100 NVL is configurable 350-400W; used 350W for the single-card PCIe entry, 400W for the NVL pair. Flippable.
5. Gaudi 3 = 900W — air-cooled OAM variant. Liquid-cooled is 1200W (separate variant).

Boundary tier decisions

• L40S → pro — display outputs, Ada-Pro silicon, but heavily datacenter-deployed in inference. Either-way defensible.
• L4 → datacenter — sold at nvidia.com/data-center/l4, no display outputs, 72W low-profile inference card.
• Apple line → consumer — Mac is consumer/prosumer; SoCs ship across MacBook Pro/Mac Studio/Mac mini.
• Cerebras WSE-3 → datacenter — wafer-scale only in datacenter form; placed there for lack of a better tier.

Test plan

• `npm test` — 395/395 passing (+2 new schema-validation tests in `test/data/accelerators-schema.test.ts`)
• `npm run check` — 0 errors, 0 warnings.
• `npm run build` — clean.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	llm-calc	33cffa0	Commit Preview URL	Jun 16 2026, 01:03 AM