inference/research
verifiedhigh · 0.85updated 2026-10-11

cadamcat: A second RX 7900 XT buys 1.1x at decode and 1.9x at prefill

benchmark

Fields

id
sources/cadamcat-dual-radeon-a100-vs-two-radeons
URL
https://cadamcat.github.io/dual-radeon-vllm/articles/a100-vs-two-radeons.html
Publisher
cadamcat (dual-radeon-vllm)
Kind
benchmark
Accessed
2026-10-11
notes
Direct instrumentation of tensor parallel scaling across 2x RX 7900 XT under vLLM, decomposing decode KV head starvation vs prefill attention parallel scaling.

Notes

Direct instrumentation of tensor parallel scaling across 2x RX 7900 XT under vLLM, decomposing decode KV head starvation vs prefill attention parallel scaling.

Cross-references

This record references no other records.