Epoch AI chart: AMD MI525X doubles Nvidia GB300 inference performance per chip
AccBalanced · x · 2026-11-15
The reposted Epoch AI chart (Nov 8, 2025) ranks modeled inference performance per chip in GB300-equivalents: AMD's MI525X leads at 2.9, roughly 2x Nvidia's GB300 (1.4), with Google TPU v7 at 2.2, Nvidia VB200 1.5, and Google TPU v8 0.8; most other parts (H200, B200, GB200, etc.) sit below 1.0.
The original poster argues Nvidia under Jensen Huang prioritized shipments to restricted markets over engineering execution, and that falling behind on design wins is making its position increasingly desperate — a pointed take on the Nvidia-vs-AMD race backed by the benchmark chart.
More from Infra
- Report: Top 4 Chinese AI Labs Trained New Models on Export-Banned Nvidia Blackwell Chips — Traditional-Chip8339 · 2026-11-15
- Alibaba's T-Head files for Hong Kong IPO, spinning off data-center chip business — pstAsiatech · 2026-11-15
- Vercel: decision models took 26% of gateway requests for 0.53% of spend — nukebuddy · 2026-11-15
- Bitcoin miner MARA to sell large amounts of BTC to fund AI data center expansion — Polymarket · 2026-11-15
- J.P. Morgan: Nvidia's share of future frontier lab compute deals falls to 39% from 80%+ — Gregory_C_Allen · 2026-11-15
- Dev rewrites OpenCode Browse in Rust for $60, cutting memory use by orders of magnitude — AccBalanced · 2026-11-15