gpt-oss-20b · 23 GPUs measured first-party · llama.cpp Q4_K_M · Updated October 2026
What GPU Do You Need for gpt-oss-20b?
We ran gpt-oss-20b (Q4_K_M, llama.cpp) on 23 GPUs, from budget cards to datacenter parts, and measured generation speed, prompt processing, power draw and peak VRAM on every one. The fastest was NVIDIA RTX PRO 6000 Blackwell Workstation Edition at 375.1 tok/s.
Benchmarked weights: unsloth/gpt-oss-20b-GGUF
NVIDIA GeForce RTX 5090
415.1 tok/s on gpt-oss-20b, the ceiling. Measured on our bench. 32GB of VRAM, $1,999 at launch.
- 415.1 tok/s on gpt-oss-20b
- 32GB, clears the gpt-oss-20b floor
- 575W board rating
NVIDIA GeForce RTX 4090
286.2 tok/s on gpt-oss-20b, fastest card you can buy at retail. Measured on our bench. 24GB of VRAM, $1,599 at launch.
- 286.2 tok/s on gpt-oss-20b
- 24GB, clears the gpt-oss-20b floor
- 450W board rating
GeForce RTX 5060 Ti
143.3 tok/s on gpt-oss-20b, lowest launch price that still fits. Measured on our bench. 16GB of VRAM, $429 at launch.
- 143.3 tok/s on gpt-oss-20b
- 16GB, clears the gpt-oss-20b floor
- 180W board rating
GeForce RTX 5070 Ti
247.4 tok/s on gpt-oss-20b, most speed per dollar. Measured on our bench. 16GB of VRAM, $749 at launch. That is 330.3 tok/s per $1,000 of launch price.
- 247.4 tok/s on gpt-oss-20b
- 16GB, clears the gpt-oss-20b floor
- 300W board rating
What GPU Do You Need for gpt-oss-20b?, tok/s by GPU
Top 15 shown; 8 more cards in the full table below.
Efficiency: tok/s per 100W drawn
Top 15 shown; 8 more cards in the full table below.
Power is the average pulled during the run, sampled at 1Hz. The fastest card is often not the one here, and for anything left running this is the number that shows up on the bill.
Value: tok/s per $1,000 of MSRP
Top 15 shown; 8 more cards in the full table below.
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
gpt-oss-20b. Measured generation speed by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 | 415.1 | 20022.7 | 2.05 | 202.8 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 375.1 | 17051.6 | 2.46 | 152.5 W |
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 359 | 17392.5 | 2.65 | 135.4 W |
| NVIDIA B300 | 355.6 | 6753.9 | 1.27 | 278.9 W |
| NVIDIA H200 | 354.2 | 10042.7 | 2.19 | 161.9 W |
| NVIDIA B200 | 351.4 | 11467.3 | 1.1 | 318.2 W |
| NVIDIA H100 80GB HBM3 | 346.8 | 10032.6 | 2.1 | 165.1 W |
| NVIDIA H100 NVL | 296.8 | 9189.3 | 1.83 | 162.2 W |
| NVIDIA GeForce RTX 4090 | 286.2 | 12050.9 | 1.86 | 154.3 W |
| GeForce RTX 5080 | 265.4 | 11718.8 | 2.3 | 115.3 W |
| NVIDIA RTX 5880 Ada Generation | 260.6 | 8658.8 | 1.85 | 141.0 W |
| NVIDIA H100 PCIe | 251.1 | 7720.8 | 2.07 | 121.4 W |
| GeForce RTX 5070 Ti | 247.4 | 10534.2 | 2.33 | 106.0 W |
| NVIDIA L40S | 233.8 | 10861.4 | 1.69 | 138.5 W |
| NVIDIA GeForce RTX 3090 | 232 | 6271.4 | 0.91 | 253.7 W |
| NVIDIA GeForce RTX 4080 | 220.7 | 8584.2 | 1.51 | 145.8 W |
| NVIDIA A100 80GB SXM4 | 212.6 | 4701.5 | 1.38 | 154.3 W |
| NVIDIA A100 40GB SXM4 | 205.1 | 4828.9 | 1.69 | 121.1 W |
| NVIDIA A10G | 144.6 | 3691 | 1.47 | 98.6 W |
| GeForce RTX 5060 Ti | 143.3 | 5552.3 | 2.16 | 66.5 W |
| NVIDIA GeForce RTX 4060 Ti 16GB | 102.2 | 3723.5 | 1.23 | 82.8 W |
| NVIDIA L4 | 91.78 | 3333 | 1.69 | 54.3 W |
| NVIDIA T4 | 63.63 | 1373.6 | 1.23 | 51.8 W |
What the numbers show. Across 11 GPUs measured on our own bench, RTX PRO 6000 Blackwell Workstation Edition is fastest at 375 tok/s. Per dollar of launch price, A10G gives the most (51.6 tok/s per $1,000). The fastest card with 16GB or less is T4 at 63.6 tok/s. The measured peak was ~12GB, so it runs on 12GB cards and up. That floor is a property of the model, so it applies to every GPU, measured or not.
About gpt-oss-20b. gpt-oss-20b: from openai, 21B parameters, on Hugging Face since August 2025, Apache 2.0 licence. 7,254,036 downloads in the last 30 days and 3 community quantizations.
How it compares. H100 80GB HBM3: gpt-oss-20b 346.8 tok/s, DeepSeek-Coder-V2-Lite 309.1 (16B), Qwen2.5-Coder-14B 144.8 (15B), Qwen3 14B 151.2 (15B), Qwen2.5-14B 144.7 (15B). gpt-oss-20b beats all 4 here.
Cost on a rented GPU. 1M generated tokens of gpt-oss-20b: $0.15 on a RTX 3090 ($0.12/hr, 72 min), $0.26 on a RTX 5090 ($0.39/hr, 40 min, 1.8x the cost).
gpt-oss-20b: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA GeForce RTX 3090 | $0.12/hr | 232 | $0.15 |
| GeForce RTX 5070 Ti | $0.15/hr | 247.4 | $0.17 |
| GeForce RTX 5080 | $0.21/hr | 265.4 | $0.22 |
| NVIDIA GeForce RTX 4080 | $0.20/hr | 220.7 | $0.25 |
| NVIDIA GeForce RTX 5090 | $0.39/hr | 415.1 | $0.26 |
| GeForce RTX 5060 Ti | $0.14/hr | 143.3 | $0.26 |
| NVIDIA GeForce RTX 4090 | $0.34/hr | 286.2 | $0.33 |
| NVIDIA RTX 5880 Ada Generation | $0.49/hr | 260.6 | $0.52 |
| NVIDIA T4 | $0.14/hr | 63.63 | $0.59 |
| NVIDIA A100 40GB SXM4 | $0.47/hr | 205.1 | $0.64 |
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | $1.00/hr | 359 | $0.78 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 375.1 | $0.80 |
| NVIDIA L40S | $0.79/hr | 233.8 | $0.94 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 212.6 | $1.24 |
| NVIDIA L4 | $0.44/hr | 91.78 | $1.33 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 346.8 | $1.71 |
| NVIDIA H100 PCIe | $1.94/hr | 251.1 | $2.14 |
| NVIDIA H100 NVL | $2.59/hr | 296.8 | $2.42 |
| NVIDIA H200 | $3.59/hr | 354.2 | $2.82 |
| NVIDIA B200 | $5.98/hr | 351.4 | $4.73 |
| NVIDIA B300 | $6.94/hr | 355.6 | $5.42 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for gpt-oss-20b. 30+ tok/s: 23 (RTX 5090, RTX 4090, RTX 5080). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before gpt-oss-20b writes anything it reads the input: 20022.7 tok/s on the RTX 5090 (0.2s for a 4,000-token prompt), 12050.9 on the RTX 4090 (0.3s), 1373.6 on the T4 (2.9s). Long documents and big code files feel this number more than the generation speed.
VRAM for gpt-oss-20b. Measured peak 10.8GB, so 12GB is the smallest common card size; smallest card it ran on: RTX 5080 (16GB). With long context: Q4_K_M 15GB (tested), Q2_K 15GB, Q3_K_M 15GB, Q5_K_M 16GB, Q6_K 16GB.
Power on gpt-oss-20b. Most efficient: RTX PRO 6000 Blackwell Max-Q Workstation Edition, 135W, 0.10 kWh per 1M generated tokens. Hungriest: B200, 318W, 0.25 kWh. At $0.15/kWh: $0.016 per 1M generated tokens.
Our verdict
Fastest on gpt-oss-20b: NVIDIA GeForce RTX 5090, 415.1 tok/s. Cheapest consumer card that ran it: GeForce RTX 5060 Ti ($429, 143.3 tok/s). Cheapest to rent per job: NVIDIA GeForce RTX 3090, $0.15 per 1M generated tokens.
FAQ
- What GPU do I need to run gpt-oss-20b?
- About 11GB. Cheapest consumer card that ran it: GeForce RTX 5060 Ti (16GB, 143.3 tok/s).
- How fast is gpt-oss-20b on the NVIDIA GeForce RTX 5090?
- 415.1 tok/s at 203W, faster than every datacenter card we ran it on.
- How much does it cost to run gpt-oss-20b in the cloud?
- $0.15 per 1M generated tokens on a NVIDIA GeForce RTX 3090 at $0.12/hr, cheapest of 21 rentable cards we measured.
- Can I run gpt-oss-20b on a 12GB, 16GB or 24GB card?
- It used 10.8GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the RTX 4090 or the RTX 3090 faster for gpt-oss-20b?
- The RTX 4090: 286.2 vs 232.0 tok/s, 23% faster on our bench.
- Should I buy or rent a RTX 5090 for gpt-oss-20b?
- Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 7,680x 1M generated tokens of gpt-oss-20b. Buy only if you'll run more than that.
How we test
llama.cpp llama-bench at Q4_K_M, 512-token prompt and 128 generated tokens, three runs after a warmup, full GPU offload, with power and VRAM sampled throughout. Token generation is memory-bandwidth-bound, so the ranking tracks bandwidth closely, which makes it a fair guide to cards we haven't run yet.