AI compute
Revision as of 09:08, 8 September 2025 by KevinYager (talk | contribs)
Contents
Analysis
- 2024-03: AI and Memory Wall
Cloud GPU
Cloud Training Compute
Cloud LLM Routers & Inference Providers
- OpenRouter (open and closed models, no Enterprise tier)
- LiteLLM (closed models, Enterprise tier)
- Cent ML (open models, Enterprise tier)
- Fireworks AI (open models, Enterprise tier)
- Abacus AI (open and closed models, Enterprise tier)
- Portkey (open? and closed models, Enterprise tier)
- Together AI (open models, Enterprise tier)
- Hyperbolic AI (open models, Enterprise tier)
- Huggingface Inference Providers Hub
- AskSage
Multi-model with Model Selection
Multi-model Web Chat Interfaces
Multi-model Web Playground Interfaces
Local Router
Acceleration Hardware
- Nvidia GPUs
- Google TPU
- Etched: Transformer ASICs
- Cerebras
- Untether AI
- Graphcore
- SambaNova Systems
- Groq
- Tesla Dojo
- Deep Silicon: Combined hardware/software solution for accelerated AI (e.g. ternary math)
Energy Use
- 2021-04: Carbon Emissions and Large Neural Network Training
- 2023-10: From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference
- 2024-01: Electricity 2024: Analysis and forecast to 2026
- 2024-02: The carbon emissions of writing and illustrating are lower for AI than for humans
- 2025-04: Why using ChatGPT is not bad for the environment - a cheat sheet
- A single LLM response uses only ~3 Wh = 11 kJ (~10 Google searches; examples of 3 Wh energy usage)
- Reading an LLM-generated response (computer running for a few minutes) typically uses more energy than the LLM generation of the text.
- 2025-07: Mistral: Our contribution to a global environmental standard for AI
- 2025-08: Measuring the environmental impact of delivering AI at Google Scale (blog)
