Garage · Updated 2026-10-10
A Radeon AI PRO R9700 next to a 3060
by Need for VRAM @needforvram Verified owner
- Cards
- 1 × AMD Radeon AI PRO R9700 (32 GB), 1 × NVIDIA RTX 3060 (12 GB)
- CPU
- AMD Ryzen 9 5950X (16 cores, 32 threads)
- System RAM
- 30 GB
- Engines
- llama.cpp (ROCm/HIP), llama.cpp (CUDA), ComfyUI
Lap times
Generation speed, tokens/sTwo vendors in one case, and that is the point of this rig: the same machine answers an AMD question and an NVIDIA question without a rebuild. The R9700 carries the model that needs 32 GB, the 3060 carries what fits in 12 GB, and the two never pool their memory — llama.cpp splits a model across cards by layers, it does not add the VRAM up.
Why it is interesting
The R9700 is the card that makes consumer gfx1201 (RDNA4) measurable at all: most published numbers for it come from a benchmark harness rather than from a model actually served, and the floors on this site are per-vendor ladders, so a card whose backend is ROCm has to be tested with ROCm. The 3060 beside it is the cheap 12 GB baseline many people already own, which makes it the fair comparison point: what a second-hand NVIDIA card does with a quant the AMD card has room for at a wider one.
What comes next
Wall power under load, a temperature log under sustained load (the AMD card’s junction temperature is the number to watch, not the edge), and the same model on both cards at the same quant, to show the ROCm/CUDA gap on identical weights.
The number that decides
44 GB of VRAM in total, but 32 GB usable for one model: the card a model lands on is the limit, and that is why the per-card rows below are kept separate.
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.