Radeon AI PRO R9700 vs RTX 5090 for local LLMs: speed, ROCm friction and price in November 2026

Both cards have 32 GB, but the RTX 5090 has 1,792 GB/s of memory bandwidth against about 640 GB/s for the Radeon AI PRO R9700, and on the same Qwen3.5-35B-A3B Q4_K_XL file in llama.cpp it generated 194.0 tok/s against 127.4 for the R9700 on Vulkan, with prompt processing 2.6 to 3.4 times faster (llama.cpp discussion #19890, February 2026). The R9700 costs far less: the cheapest in-stock new card on gpuprix was $2,000 on 4 November 2026, against $7,550 for an RTX 5090. It runs on ROCm 10.1 and vLLM, but owners report more setup work, and RDNA 4 has no FP4 matrix support.

Sources: AMD's Computex 2025 release (64 CUs, 32 GB, 256-bit, 300 W), ROCm GPU specs (gfx1201), ROCm precision support (RDNA 4 matrix types), Wccftech on the $1,299 price, retail date and blower cooler; NVIDIA's 5090 page. Bandwidth, MSRP and prices from gpuprix's R9700 US, R9700 Germany, 5090 US and 5090 Germany pages, updated 4 November 2026; the cheapest R9700 was an ASUS Turbo at B&H and the cheapest 5090 an ASUS TUF on Amazon. AMD's own R9700 product page timed out when I tried to read it, so the bandwidth figure here is gpuprix's.

How much faster is the RTX 5090 on the same model?

About 1.5 times on generation and 2.6 to 3.4 times on prompt processing, on the one same-file comparison published. In llama.cpp discussion #19890, opened 25 February 2026, one user ran Qwen3.5-35B-A3B UD-Q4_K_XL (about 19 GB) on both cards with the same llama.cpp commit (244641955), flash attention on and a q8_0 cache. The 5090 ran CUDA 13.0 on Linux with a Ryzen 9 9900X; the R9700 ran Vulkan on Windows with a six-core Ryzen 5 7500F.

tok/s, single stream, llama.cpp llama-bench. The last row mixes sources and builds (Hardware Corner names build 8189) and is only a rough check. The 163 tok/s figure is from user zedbytes, confirmed by the thread's author, who estimated it at about 61% of the R9700's bandwidth.

My reading of why generation is closer than the 2.8x bandwidth gap: this is a mixture-of-experts model with about 3B active parameters: each token reads little, so fixed per-token costs take a larger share and the faster card cannot use all its bandwidth. Prompt processing is compute-bound, and there the gap opens. By my arithmetic, a 32,768-token prompt takes 17.5 seconds on the R9700 and 5.1 seconds on the 5090. For chat that is tolerable; for a coding agent that resends a long context every turn it is the difference that matters most.

On dense models the gap should approach the bandwidth ratio, since every token reads every weight; I found no same-file dense comparison to confirm it. For two cards each on a dense 27B, the closest pair of reports uses different engines: Wendell's two R9700s ran Qwen3.8-27B at MXFP4 in the Radiance engine at 133 tok/s single-stream and 678 tok/s total at 8 streams, with 7-token speculative decoding (Level1Techs, 6 October 2026), while Paralon's two 5090s at 400 W ran the same model at W4A16 in vLLM 0.31.0 at 149 to 156 tok/s and 778 tok/s at 8, with 3-token MTP (Paralon, 12 October 2026). Treat that as "same ballpark with heavy tuning", not as a ranking.

Does ROCm support the R9700 in November 2026?

Officially, yes, on Linux and Windows. AMD's ROCm 10.1.0 compatibility matrix lists the R9700 (gfx1201) on Ubuntu 26.04.1, 24.04.5 and 22.04.5, RHEL 10.2 and 9.8, and Windows 11 25H2, and the 10.1.0 release notes (5 October 2026) list vLLM 0.29.0 and SGLang 0.5.18 for gfx1200 and gfx1201. vLLM's own install docs list the Radeon RX 9000 series (gfx1200/1201) and a vllm/vllm-openai-rocm image built on ROCm 10.0, which vLLM v0.33.0 (2 November 2026) made the default. llama.cpp release b12189 (7 November) ships ROCm 10.0 builds for Linux and Windows alongside Vulkan.

One inversion is worth knowing: vLLM's gpt-oss recipe lists the Radeon AI PRO R9700 as a supported GPU and the RTX 5090 as "in progress". For the models people actually run on a 5090, CUDA support is rarely the problem; the planned guide to vLLM on an RTX 5090 covers what works on sm_120.

Where RDNA 4 falls short is FP4. ROCm's precision table gives RDNA 4 FP8 (E4M3 and E5M2), INT8 and BF16 matrix support but no FP4, so NVFP4 and MXFP4 checkpoints must be dequantized on the R9700, while the 5090's tensor cores multiply in FP4 directly. FP8 is the format to use on the R9700, and the quantization notes explain the trade-offs between them.

What software friction do R9700 owners report?

Mostly in vLLM and multi-GPU setups; single-card llama.cpp is close to plug-and-play. The reports from 2026, with dates:

• Missing tuned kernels. vLLM ships pre-tuned Triton configs per GPU, and an August 2026 upgrade to AMD's ROCm 7.14 image (vLLM 0.23.1) had none for the R9700 at the shapes Qwen3.6-35B-A3B-FP8 needs. Single-stream speed on two R9700s fell from about 77 to about 39 tok/s, a 47 to 49% regression, while Gemma 4 26B-A4B was unaffected at about 95 to 98 tok/s. Hand-tuned configs recovered 7 to 9% on the older image but only 2 to 3 points on the new one, and the MoE sweep alone took about 8.5 hours on an idle card (Level1Techs, 2 August 2026).
• Unmerged FP8 tuning. vLLM issue #28649, opened 13 November 2025, reported Qwen3-30B FP8 decode rising from about 52 to 85 tok/s with community kernel configs for gfx1201 (the poster later credited part of the gain to a nightly build). It was still open, without a merged fix, in the comments I could read (to July 2026).
• Multi-GPU deadlocks. Wendell's May 2026 guide found tensor parallelism of 2 or more deadlocks on gfx1201 without NCCL_PROTO=Simple, because of an RCCL tuning bug, and that the vLLM v0.21.0 image hung outright (Level1Techs, 22 May 2026).
• Backend choice. In discussion #19890, Vulkan beat ROCm on generation in two separate reports (about 183 against 150 tok/s on Qwen3-30B-A3B; about 140 against at most 105 on Qwen3-Coder-30B-A3B), and a card in a chipset x4 slot dropped from 85 to 67 tok/s, which another user later attributed, unverified, to a Mesa RADV issue.

The working multi-GPU recipe from Wendell's guide, for reference:

# kernel command line
amd_iommu=on iommu=pt pcie_aspm=off

# vLLM ROCm container, tensor parallel 2
docker run --device /dev/kfd --device /dev/dri --group-add video \
  --ipc=host -e NCCL_PROTO=Simple \
  vllm/vllm-openai-rocm:v0.20.2 \
  --model Qwen/Qwen3.6-35B-A3B-FP8 --tensor-parallel-size 2

None of this is unusual for a new GPU family, and the R9700 is a year old. But on a 5090 the same jobs mostly run from the default images, which is part of what its price buys. The DGX Spark against Strix Halo comparison makes the same CUDA-against-ROCm point for AMD's 128 GB mini-PCs.

Which is better value: R9700 or RTX 5090?

The R9700, by a wide margin, at November 2026 prices. The calculation uses the same-file generation speeds from discussion #19890 and gpuprix's US prices of 4 November.

My arithmetic: 127.4 / 2.0, 194.0 / 7.55, 127.4 / 1.7, 194.0 / 3.33, 32 / 2.0, 32 / 7.55, 127.4 / 300 W, 194.0 / 575 W, 1,877 / 2.0 and 6,461 / 7.55. Rated board power, not measured draw. Using the R9700's later 163 tok/s raises its figures by 28%.

Even prompt processing per dollar favors the R9700 slightly at in-stock prices (939 against 856), although the 5090 is 3.4 times faster per card. Two R9700s give 64 GB for about $4,000, a little over half the price of one in-stock 5090, which puts models in reach that a single 32 GB card cannot load. Four of them, in Wendell's guide, ran Qwen3.5-122B-A10B at GPTQ-Int4 at 53.8 tok/s decode with tensor parallelism of 4.

Which should you buy?

For a personal llama.cpp machine at today's prices, I would buy two R9700s before one RTX 5090: twice the memory for about half the money, and generation speed that is fine for one person. Use the Vulkan build first and compare against ROCm on your model.

Buy the 5090 if your work is prompt-heavy (long documents, agents), if you serve with vLLM and want the default images to work, or if you depend on NVFP4 checkpoints or CUDA-only tools. Its single-card speed is in another class, and the sibling roundup of RTX 5090 LLM benchmarks collects the published numbers. For a 24 GB option from the third vendor, see the sibling piece on the Intel Arc Pro B60, and for the general method, choosing a GPU for open-source LLMs.

GPUwerk rents a dedicated RTX 5090 at $0.89 an hour but has published no measurements of its own on it; every number on this page is third-party. An hour on the card with your own model and prompts is a cheap way to see whether its speed is worth three to four R9700s' worth of money. If the model does not fit 32 GB at all, the RTX 5090 against DGX Spark comparison and the memory calculator cover the next size up.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• RTX 5090 (CUDA) vs Radeon AI PRO R9700 (Vulkan): Qwen3.5-35B-A3B llama-bench results (discussion #19890), ggml-org/llama.cpp on GitHub, 2026-02-25
• AMD introduces new Radeon graphics cards and Ryzen Threadripper processors at COMPUTEX 2025, AMD, 2025-05-20
• AMD officially prices Radeon AI PRO R9700 at $1299, Wccftech, 2025-10-23
• Compatibility matrix (ROCm 10.1.0), AMD ROCm documentation
• ROCm 10.1.0 release notes, AMD ROCm documentation, 2026-10-05
• Data types and precision support, AMD ROCm documentation
• AMD GPU specifications, AMD ROCm documentation
• Radeon AI PRO R9700 price in United States, gpuprix, 2026-11-04
• Radeon AI PRO R9700 price in Germany, gpuprix, 2026-11-04
• GeForce RTX 5090 price in United States, gpuprix, 2026-11-04
• GeForce RTX 5090 price in Germany, gpuprix, 2026-11-04
• GeForce RTX 5090 specifications, NVIDIA
• RTX 5090 local LLM benchmarks, Hardware Corner, 2026-03
• vLLM 0.23.x / ROCm 7.14 upgrade on 4x Radeon AI PRO R9700, Level1Techs Forums, 2026-08-02
• AMD Radeon R9700 Linux TP2 and TP4: the definitive guide, Level1Techs Forums, 2026-05-22
• Ultimate 2x R9700 guide for Qwen 3.8 27b MXFP4, Level1Techs Forums, 2026-10-06
• Someone please upstream this gfx1201/RDNA4 FP8 patch into vllm-rocm (issue #28649), vllm-project/vllm on GitHub, 2025-11-13
• GPU installation, vLLM documentation
• GPT-OSS recipe, vLLM recipes
• vLLM releases, vllm-project/vllm on GitHub, 2026-11-02
• llama.cpp releases, ggml-org/llama.cpp on GitHub, 2026-11-07
• Two RTX 5090s on a 27B model: 150 tok/s alone, 1,300 at 32, Paralon Cloud, 2026-10-12