Is the AMD R9700 a great GPU now?

The Radeon AI PRO R9700 is a great GPU for one job: putting a lot of new VRAM in a normal workstation. It is not a great GPU if you want the fastest single card you can buy.

Each card has 32GB of GDDR6, a blower that exhausts out the back, and a 300W rating, on the full Navi 48 chip. Bandwidth is 640 GB/s. AMD set a $1,299 retail price in October 2025. Shortage pricing had it around $1,800 by October 2026. ROCm officially supports it, and llama.cpp and vLLM both run on it. Plan on Linux.

What you actually get

Capacity, not speed. By our estimate, one R9700 runs Qwen3.8 27B at Q4_K_M at about 24 tok/s. A used RTX 3090 does the same model at about 39 tok/s, with 8GB less memory and more bandwidth (936 GB/s against 640). At Q6_K the R9700 is about 18 tok/s estimated. The extra memory is the point: a 27B model at a higher quant fits with room for context, which is the case the 3090 cannot make.

Measured llama.cpp numbers tell the same story. On Llama 2 7B Q4_0, a community ROCm run posted in September 2026 had the R9700 at 112.72 tok/s generating and 4,665 tok/s on a 512-token prompt, with flash attention on at 121.62 tok/s. That is a real card doing real work. It is not an RTX 5090.

Four R9700s are 128GB in a case that can take blower cards, at about $7,200 and about $56 per gigabyte. An RTX 5090 was about $4,300 for 32GB in October 2026, about $134 per gigabyte, and it is much faster per card. Four used 3090s are about 96GB and about $4,200, draw far more power, and usually want an open frame. The GPU buying guide has the full pick.

The catches

640 GB/s is modest for 32GB. An RX 9070 XT has the same bandwidth and the same RDNA 4 matrix hardware for about $800, with 16GB instead of 32. You are paying for memory and the blower, not for a faster chip.

CUDA-only tools will not run. Prompt processing on AMD still trails a comparable NVIDIA card. And the street price is well above the $1,299 MSRP, so run the per-gigabyte math again before you buy four.

The Arc Pro B70 is the other new 32GB card, about $1,300 with 608 GB/s. It is cheaper and the software still needs more setup. If the requirement is “32GB, supported, and I can put four in a tower,” the R9700 is the one I would buy. If the requirement is “one fast CUDA card,” it is not. That is still the 3090, or a 5090 if the budget is there.