"Which GPU should I buy for local LLMs?" usually gets answered with spec sheets. I wanted an answer from measured numbers, so I took the open RigBase dataset (117 home AI builds, 298 measured results, every row linked to its public source) and divided tokens/sec by what the card costs today.
Short version:
It's free under CC BY 4.0: builds.csv and builds.json. All of these are llama-bench runs in native MXFP4 from the same llama.cpp thread (discussion #15396), so this is the cleanest comparison in the set.
| GPU | tok/s | Price | tok/s per €1k | |
|---|---|---|---|---|
| RTX 5070 Ti | 189.45 | €1,245 new (2026-10-24) | **152.2** | [build](https://getrigbase.com/builds/single-rtx-5070-ti-gpt-oss/) |
| RTX 5060 Ti 16GB | 111.51 | €739 new (2026-10-16) | **150.9** | [build](https://getrigbase.com/builds/single-rtx-5060-ti-16gb-gpt-oss/) |
| RTX 3090 (used) | 161.77 | €1,248 used (eBay sold median) | **129.6** | [build](https://getrigbase.com/builds/single-rtx-3090-gpt-oss/) |
| RTX 5080 | 204.85 | €1,770 new (2026-10-24) | **115.7** | [build](https://getrigbase.com/builds/single-rtx-5080-gpt-oss/) |
| RTX 5090 | 282.51 | €7,140 new (2026-10-24) | **39.6** | [build](https://getrigbase.com/builds/single-rtx-5090-gpt-oss/) |
| RTX Pro 6000 Blackwell | 286.91 | €16,885 new (2026-10-24) | **17.0** | [build](https://getrigbase.com/builds/rtx-pro-6000-96gb/) |
The 5090 is only 1.49× faster than a 5070 Ti here (282.51 vs 189.45), yet it costs 5.7× as much. A 20B MoE in MXFP4 fits comfortably in 16 GB, so the extra VRAM buys you nothing on this model.
Jeff Geerling's ai-benchmarks swaps cards in one Core Ultra 265K system and records peak whole-system power: | GPU | gpt-oss 20B tok/s | Price | tok/s per €1k | Peak system power | tok/s per W |
|---|---|---|---|---|---|
| RTX 3060 12GB | 83.96 | €429 (2026-10-16) | **195.7** | 229.3 W | 0.37 |
| RTX A4000 | 93.55 | €1,050 (2026-10-24) | **89.1** | 199.4 W | 0.47 |
| Radeon AI Pro R9700 | 163.26 | €1,848 (2026-10-24) | 88.4 | 332.9 W | 0.49 |
| RTX 4090 | 279.65 | €3,955 (2026-10-16) | **70.7** | 509.1 W | 0.55 |
Sources: [#40](https://github.com/geerlingguy/ai-benchmarks/issues/40), [#36](https://github.com/geerlingguy/ai-benchmarks/issues/36), [#35](https://github.com/geerlingguy/ai-benchmarks/issues/35), [#43](https://github.com/geerlingguy/ai-benchmarks/issues/43). The quant isn't stated in his summary table.
The cheap card wins on value and the fast card wins on efficiency. The €429 RTX 3060 has the best tok/s per euro in the dataset for this model. One caveat: gpt-oss 20B's weights are about 13 GB, more than the card's 12 GB, so part of the model may have run from system RAM. The 4090 draws more than twice the power, but gets the most tokens per watt.
Here are single-card results on current models, without speculative decoding. Each cell shows measured tok/s → tok/s per €1k.
| Model | RTX 3090 (used, €1,248) | RTX 4090 (€3,955) | RTX 5090 (€7,140) | 5090 ÷ 3090 speed |
|---|---|---|---|---|
| Qwen3.6 35B-A3B | 146 → 117.0 | 160.6 → 40.6 | 258 → 36.1 | 1.77× | | Gemma 4 31B | 38.8 → 31.1 | 43.8 → 11.1 | 71.3 → 10.0 | 1.84× | | Nemotron 3.5 Bolt 74B-A9B (2-bit) | 87.75 → 70.3 | 108 → 27.3 | 154.7 → 21.7 | 1.76× | | Qwen3.8 27B | 40.4 → 32.4 | — | 75.2 → 10.5 | 1.86× |
Sources per row:
On all four models the used 3090 gives 3.1–3.2× more tok/s per euro than the 5090. The 5090 is consistently ~1.8× faster, which tracks its memory bandwidth (1,792 vs 936 GB/s). It also has 8 GB more VRAM, which matters for long context. If raw speed is the goal, buy the 5090. If value is the goal, the used 3090 remains the card to beat in late 2026.
Here are ten before/after pairs where the same build ran the same model with the same engine and quant, and the only change was turning on a draft method (MTP, DSpark):
| Setup | Model | Plain tok/s | With drafting | Speedup |
|---|---|---|---|---|
| RTX 3090, llama.cpp Q4_K_M, MTP 2 drafts (code prompt) | Qwen3.8 27B | 40.4 | 72.5 | 1.79× |
| RTX 3090, llama.cpp Q4_K_M, MTP 2 drafts (reasoning prompt) | Qwen3.8 27B | 40.4 | 50.9 | 1.26× |
| Strix Halo, llama.cpp Vulkan Q6_K, MTP n=2 | Qwen3.8 27B | 9.59 | 18.01 | 1.88× |
| Strix Halo, llama.cpp ROCm Q8_0, MTP n=3 | Qwen3.8 27B | 7.76 | 15.67 | 2.02× |
| Strix Halo, llama.cpp Vulkan Q8_0, MTP n=3 | Qwen3.8 27B | 7.77 | 13.19 | 1.70× |
| DGX Spark, vLLM, MTP k=2 | Qwen3.8 27B | 11 | 19 | 1.73× |
| DGX Spark, vLLM, MTP drafter k=3 | Gemma 4 26B-A4B | 30.8 | 56.5 | 1.83× |
| DGX Spark, vLLM NVFP4, MTP k=2 | Qwen3.6 35B-A3B | 78.1 | 102.3 | 1.31× |
| DGX Spark, vLLM, DSpark k=3 | Nemotron 3.5 Lightning 30B-A3B | 70.3 | 94.4 | 1.34× |
| 2× Tesla V100, bnk Q8_0, MTP | Qwen3.8 27B | 42.4 | 104 | 2.45× |
Sources: OpenJet, Sleeping Robots, howquant.dev, Cristian Tala, bnk. The speedups range from 1.26× to 2.45×, with a median of 1.76×. That's roughly the gap between a 3090 and a 5090, and it costs nothing. Two patterns show up:
The catch is that drafting numbers are not comparable with plain decode across sources. That's why RigBase flags them ⚡ and lets you hide them.
For ~100B+ models, the hardware is mostly unified-memory boxes (Mac Studio, Strix Halo, DGX Spark), and I don't have comparable price data for them yet. So here are the raw measured speeds for gpt-oss 120B: | Hardware | tok/s | Engine | |
|---|---|---|---|
| RTX Pro 6000 Blackwell (96 GB) | 196.31 | llama.cpp MXFP4 | [source](https://github.com/ggml-org/llama.cpp/discussions/15396) |
| Apple M2 Ultra 76-core (192 GB) | 79.68 | llama.cpp MXFP4 | [source](https://github.com/ggml-org/llama.cpp/discussions/4167) |
| Ryzen AI Max+ PRO 495 (Radeon 8065S) | 55.8 | llama.cpp Vulkan MXFP4 | [source](https://www.hommie.net/posts/ai-box.html) |
| Ryzen AI Max+ PRO 395 (Radeon 8060S) | 44.7 | llama.cpp Vulkan MXFP4 | [source](https://petronellatech.com/blog/dgx-spark-vs-strix-halo-vs-gorgon-halo-local-llm-benchmarks/) |
| DGX Spark (GB10) | 44.4 | vLLM, 128K ctx | [source](https://ctala.github.io/local-llm-agentic-workflows/benchmarks/) |
A €16,885 card is 2.5× faster than a 2023 Mac Studio, and the Strix Halo boxes match the DGX Spark. If you know reliable current prices for these systems, I'd like to add them.
Data snapshot: RigBase builds.json, 24 Oct 2026. Prices: billiger.de 16–24 Oct 2026; used RTX 3090 from PCPrice.Watch, 24 Oct 2026. All ratios are arithmetic on the sourced numbers. Corrections welcome.