Radeon R9700 against RTX 5090: same file, two cards, one price per token

Our production card is the RTX 5090. The card on the other side of the bench is a Radeon AI PRO R9700 with 32 GB that has been sitting in the building since 1 September and is not carrying any traffic. Both cards hold 32 GB, and we fed both of them the same model file, with the same engine and the same command line. That combination is rare in public comparisons, because usually one side runs CUDA while the other runs a different backend, or the two sides are tested with different quants. On 16 October 2026 we worked out what remains once those error sources are removed.

How the measurement came about

The model is dense Qwen3.8-27B in IQ4_XS, one and the same file of 17.03 GB on both cards, llama.cpp with -ngl 99 -fa on, a 262,144 context, one sequence, MTP acceptance on. Measured idle, meaning no other load on the machine (why we still take the VM under Proxmox).

The condition that made the comparison possible is a failed earlier version of the same test. On the 5090 the model first ran at 14 to 18 tokens per second, because the offload into VRAM had not been forced. With forced offload and the next larger file above 24 GB we got cudaMalloc failed: out of memory, and the lane fell into a crash loop (running the 27B model ourselves). With the smaller file and -ngl 99 it landed at 120 to 146 tokens per second. On the R9700, 29.9 of the 32 GB are usable, so the file fits, and the second run was valid.

Two differences in the protocols survive, smaller than before but not gone. The R9700 side has used an 8-bit KV-cache type since 29 September, while the 5090 used the default type in this protocol. And the two machines have different CPUs. Both need a new run to be closed out (bandwidth as the bottleneck).

Midpoints of the measured ranges, 41 to 45 against 120 to 146 tokens per second. Same model, same file, same engine, one card each. (Quelle: Own measurement, protocols assembled on 16 October 2026)

The result in euros per token

We read the prices from Austrian retail on the same day. The R9700 is listed at 2,194.29 euros (XFX, 14 offers, retrieved 16 October 2026), the 5090 at 6,790 euros (ASUS ROG Astral, 2 offers) and at 6,869.99 euros in a single offer from an Austrian merchant (MSI Ventus 3X OC). The ratio is 3.09 to 3.13.

	Radeon R9700	RTX 5090
Measured speed	41 to 45 tokens/s	120 to 146 tokens/s
Midpoints	43	133
Austrian street price	2,194.29 euros	6,790 euros
Euros per token per second	51.03	51.05

Speed 3.09 times higher, price 3.10 times higher, both 51 euros per token per second to three digits. The question which card offers the better value is answered by this measurement, and the answer is: neither. One level down is where the decision lives.

The calculation nobody expects

The product page gives the R9700 300 watts peak board power, the 5090 575 watts total graphics power (AMD product data, NVIDIA product data). Divided by measured speed:

R9700: 300 watts over 43 tokens per second is 6.98 watts per token per second
RTX 5090: 575 watts over 133 tokens per second is 4.32 watts per token per second

The Nvidia card is the more economical machine per unit of work delivered, by a factor of 1.62. Lower power draw is therefore no argument for AMD as long as it is paid for with fewer tokens. Over three years, with 0.20 euros per kilowatt hour as an assumption and nothing measured at the power supply:

Duty	R9700, euros per token per second	RTX 5090, euros per token per second
Purchase alone	51.03	51.05
20 percent load, three years	58.36	55.60
100 percent load, three years	87.70	73.78

At full duty the 5090 comes in 16 percent below the R9700, at a fifth of the load 5 percent below. The harder the card works, the more the parity tips toward the expensive one. The same logic already governed for the running costs of the whole stack (what a model really costs).

Three assumptions sit in that table rather than in a measurement log. The watt figures are datasheet values, and an idle card draws less than its maximum. The 0.20 euros per kilowatt hour is a placeholder rate. And depreciation, rack space, cooling and staff are not included. The calculation still holds, because the direction is far from the measurement margin: even if the R9700 drew only half of what its datasheet states, it would remain the more expensive card per token.

The third-party finding that leaves our number open

In llama.cpp discussion 19890 one user benchmarked both cards on the same engine, Qwen3.5-35B-A3B in Q4_K_XL, commit 244641955. For decoding across several context lengths he arrives at 194 against 127 tokens per second, a factor of 1.52, at a price gap he reports as considerably smaller than the one we see in Austria. His conclusion is dry: 1.52 times the speed is hard to justify on that surcharge, and for a 70B model he recommends two R9700 cards instead of one 5090.

A second observation in the same thread concerns us directly. The same R9700 is said to run at around 140 tokens per second on the Vulkan backend and rarely above 105 on ROCm. Our own two Vulkan attempts ended in August with VK_ERROR_DEVICE_LOST inside the graphics driver and a short GPU reset, and we went back to ROCm. If that report is right, our spare card is running roughly a third below what it can do with this model. It is a test without any purchase, and it is still open.

Capacity is a different axis

The same thread carries a result that appears in no speed table. A single 5090 there would not load a 72B model in Q4_K_M at all, and not a 122B model under any condition. Two R9700 cards with roughly 60 GB usable in common do load a 70B model and bring it to 11.49 tokens per second, with about 600 tokens per second on prompt processing. Slow, but running, and the answer to a capacity question is space rather than speed.

For us that means: 4,388.58 euros for two R9700 buys a model class that 6,790 euros for one 5090 cannot reach, and our current four-card configuration does not reach it either. Anyone who wants a 70B model in the building at better than Q4 density buys memory, not speed (the RTX Pro 5500 with 84 GB, purchase maths).

What we do not know

Card bandwidth separates the two measurements more than we would like. AMD states 640 GB/s peak on a 256-bit bus. NVIDIA lists the 5090 with 32 GB of GDDR7 on a 512-bit bus but gives no bandwidth figure, and our own test used 1,790 GB/s. It is a factor of 2.80, and decoding a dense model runs over that bus. A second measurement with the same KV-cache type on both cards on identical CPUs would make the comparison clean.

Energy stays open as well. One metered working day per card, with a protocol covering idle and fetch, would be the only figure that replaces our datasheet arithmetic. And the question about the R9700 backend is answerable before a single euro is spent.

What we decide with it

Our lane stays where it is: four 5090 cards for the dense model, because bandwidth is the bottleneck there and the card costs less power per delivered token per second. The R9700 stays a spare and does not go into a lane before the backend test has run. A third is larger than anything we would gain by putting it in. Nothing here leads to a purchase. At equal euros per token per second, only the workload decides, and for us that means a 70B model with real requests would be the first thing to justify another card, and then because of its memory.