Transistor Radio TR 40: InferenceMAX, Video games, Rare earths
62 snips
Oct 10, 2025 The hosts dive into the groundbreaking InferenceMAX, exploring its role in establishing open benchmarks for AI performance. They analyze how historical GPU tactics shaped today’s standards and discuss the significance of performance-per-dollar metrics over simple FLOPS. Transitioning to gaming, they examine EA's recent acquisition and the trend of sovereign funds investing in studios. The intriguing dynamics of rare earths and their geopolitical implications highlight the concentration of processing power in China, raising questions about the future of tech supply chains.
AI Snips
Chapters
Transcript
Episode notes
Living Benchmarks Change Buying Decisions
- Inference Max is a living, nightly benchmark tracking token throughput, cost, and megawatt efficiency across hardware and software.
- The project surfaces real-world trade-offs and helps operators choose GPUs by dollars and megawatts, not just peak FLOPS.
Performance Is A Moving Target
- Inference performance shifts often due to drivers, frameworks, compilers, and model changes.
- Continuous testing captures those moving parts better than annual point-in-time comparisons.
Cost And Power Beat Peak FLOPS
- Throughput-per-dollar and tokens-per-megawatt matter more to AI infra builders than raw peak performance.
- Inference Max shows scenarios where AMD is cost-competitive while Nvidia often leads performance and energy efficiency.