Garage · Updated 2026-10-01
Threadripper with a mismatched pair
by Need for VRAM @needforvram Verified owner · YouTube channel ↗
- Cards
- 1 × NVIDIA RTX 3080 Ti (12 GB), 1 × NVIDIA RTX 3060 (12 GB)
- CPU
- AMD Ryzen Threadripper 2990WX (32 cores, 64 threads)
Lap times
Generation speed, tokens/sNo measured runs yet. They appear here as soon as the owner adds them.
The machine behind this site. Two cards that were never meant to work together: a 3080 Ti with fast memory and a 3060 with slow memory, 12 GB each.
Why it is interesting
24 GB in total, the same as a single 3090, spread over two cards with very different speeds. The 3080 Ti moves about 912 GB/s, the 3060 about 360 GB/s. When llama.cpp splits a model by layers, each token has to pass through both cards, so the slow card sets much of the pace.
That makes it a good test bench for the questions people ask about mixed setups. How to split the layers, which card should hold the context, and how much speed you give up compared to one 24 GB card.
What comes next
Wall power, cost and measured tokens per second, model by model. They will show up on this page as soon as they are measured, with the exact command lines used.
This page belongs to Need for VRAM. Only they can change it, through a pull request on the site repository. Something wrong on this page? Report it.