Cerebras says AMD team-up delivers 5× the tokens per megawatt of CS-3 alone
Cerebras says AMD GPUs handle prompts while its CS-3 systems generate responses, splitting AI inference duties.
TLDR
Different chips, different chores: Cerebras says AMD Instinct GPUs process prompts while its CS-3 systems generate responses. The company reports five times the throughput per megawatt of CS-3 alone on its GPT-OSS-120B production endpoint, without slower generation. It says AWS, Fireworks AI, Gimlet Labs and General Compute have early access, with broader deployment planned by the end of 2027. Cerebras cautions that performance can vary by workload and configuration.
Combined views
—
3 Sources, first seen ago

Comments