MirrorFrog

AI Compute Card Encyclopedia

A comprehensive catalog of 100+ AI compute cards covering training, inference, and edge computing across all scenarios.

Feature	Description	Link
Full Comparison Table	Side-by-side comparison of 100+ compute cards with filtering	View Table
Future Roadmap	2025–2026 product launch timeline	View Roadmap
Industry News	Latest releases, acquisitions, IPOs, and in-depth analysis	Browse Blog

Type	Description	Representative Products
Datacenter Training	Primary LLM training hardware	H100 / H200 / B200 / MI300X
Datacenter Inference	High-throughput online inference	L4 / L40S / RTX 6000 Ada
Wafer-Scale Training	Single-wafer solutions for massive models	Cerebras WSE-3 / WSE-4
Edge AI	On-device inference acceleration	Jetson Orin / Edge TPU
Innovative Architectures	Non-traditional architecture exploration	Groq LPU / Graphcore IPU

Vendor	Flagship Product Line	Representative Models
NVIDIA	General-purpose GPU + CUDA ecosystem	H100 / B200 / Rubin R200
AMD	Instinct series + ROCm	MI300X / MI400
Intel (Habana)	Gaudi series + OpenVINO	Gaudi 3 / Jaguar Shores
Google	TPU tensor processors	TPU v6e / Ironwood
AWS	Trainium / Inferentia	Trainium 3
Huawei	Ascend series (domestic China)	Ascend 910C / 920
Cambricon	Siyuan MLU series	MLU370 / MLU690

2026-05 — NVIDIA Rubin R200 specs confirmed: 288GB HBM4, 50 PFLOPS FP4
2026-04 — AMD MI400 + Helios rack revealed: 432GB HBM4, 260 TB/s UALink
2026-04 — Google TPU 8t (training) + 8i (inference) split architecture officially announced
2026-04 — Cerebras WSE-4 announced: 1.4 trillion transistors, 125 PFLOPS FP8
2025-11 — Huawei Ascend 920 launched: 4 Tbps inter-chip interconnect, 3× H20 compute

Comprehensive — 100+ compute cards with complete specifications
Continuously Updated — Keeping pace with vendor release cadence, 2025–2026 launches updated promptly
Neutral Comparison — No vendor bias, purely data-driven
Open Source — Contribute on GitHub, PRs welcome

Found an error or want to add a new card? Open an Issue or submit a PR!