Gemma 4 31B · 11 GPUs measured first-party · llama.cpp Q4_K_M · Updated October 2026
What GPU Do You Need for Gemma 4 31B?
We ran Gemma 4 31B (Q4_K_M, llama.cpp) on 11 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 H100 80GB HBM3 at 76.59 tok/s.
Benchmarked weights: unsloth/gemma-4-31B-it-GGUF
NVIDIA H100 80GB HBM3
76.59 tok/s on Gemma 4 31B, the ceiling. Measured on our bench. 80GB of VRAM, $30,000 at launch.
- 76.59 tok/s on Gemma 4 31B
- 80GB, clears the Gemma 4 31B floor
- Rentable by the hour rather than bought
- 700W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA GeForce RTX 5090
71.42 tok/s on Gemma 4 31B, fastest card you can buy at retail. Measured on our bench. 32GB of VRAM, $1,999 at launch.
- 71.42 tok/s on Gemma 4 31B
- 32GB, clears the Gemma 4 31B floor
- 575W board rating
NVIDIA GeForce RTX 3090
39.2 tok/s on Gemma 4 31B, lowest launch price that still fits. Measured on our bench. 24GB of VRAM, $1,499 at launch.
- 39.2 tok/s on Gemma 4 31B
- 24GB, clears the Gemma 4 31B floor
- 350W board rating
NVIDIA GeForce RTX 4090
45.08 tok/s on Gemma 4 31B, most speed per dollar. Measured on our bench. 24GB of VRAM, $1,599 at launch. That is 28.19 tok/s per $1,000 of launch price.
- 45.08 tok/s on Gemma 4 31B
- 24GB, clears the Gemma 4 31B floor
- 450W board rating
What GPU Do You Need for Gemma 4 31B?, tok/s by GPU
Efficiency: tok/s per 100W drawn
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
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
Gemma 4 31B. Measured generation speed by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 76.59 | 2490.6 | 0.24 | 324.0 W |
| NVIDIA GeForce RTX 5090 | 71.42 | 4020.6 | 0.16 | 438.8 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 65.42 | 3717.9 | 0.22 | 298.5 W |
| NVIDIA GeForce RTX 4090 | 45.08 | 2915.6 | 0.17 | 265.9 W |
| NVIDIA A100 80GB SXM4 | 43.58 | 1197.4 | 0.2 | 223.0 W |
| NVIDIA GeForce RTX 3090 Ti | 43.49 | 1637.4 | 0.13 | 347.4 W |
| NVIDIA A100 40GB PCIe | 42.57 | 1201.7 | 0.23 | 182.9 W |
| NVIDIA GeForce RTX 3090 | 39.2 | 1266.1 | 0.13 | 294.9 W |
| NVIDIA L40S | 34.95 | 2581.5 | 0.14 | 247.4 W |
| NVIDIA A10G | 22.67 | 810 | 0.17 | 132.8 W |
| NVIDIA L4 | 12.84 | 704.4 | 0.19 | 67.1 W |
What the numbers show. Across 11 GPUs measured on our own bench, H100 80GB HBM3 is fastest at 76.6 tok/s. Per dollar of launch price, RTX 5090 gives the most (35.7 tok/s per $1,000). The measured peak was ~20GB, so it runs on 24GB cards and up. That floor is a property of the model, so it applies to every GPU, measured or not.
About Gemma 4 31B. Gemma 4 31B: from google, 31B parameters, on Hugging Face since March 2026, Apache 2.0 licence. 11,312,870 downloads in the last 30 days and 2 community quantizations.
How it compares. H100 80GB HBM3: Gemma 4 31B 76.59 tok/s, GLM-4.7-Flash 186.1 (31B), Nemotron 3.5 Lightning 30B A3B 323.6 (32B), Nemotron-3-Nano-30B-A3B 316.5 (32B), Qwen3 30B A3B 283.8 (31B). All 4 beat Gemma 4 31B here. Nemotron 3.5 Lightning 30B A3B, Nemotron-3-Nano-30B-A3B and Qwen3 30B A3B are mixture-of-experts, so per token they compute only a slice of their size.
Cost on a rented GPU. 1M generated tokens of Gemma 4 31B: $0.86 on a RTX 3090 ($0.12/hr, 7.1 hours), $7.75 on a H100 80GB HBM3 ($2.14/hr, 3.6 hours, 9.0x the cost).
Gemma 4 31B: 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 | 39.2 | $0.86 |
| NVIDIA GeForce RTX 5090 | $0.39/hr | 71.42 | $1.51 |
| NVIDIA GeForce RTX 3090 Ti | $0.27/hr | 43.49 | $1.72 |
| NVIDIA GeForce RTX 4090 | $0.34/hr | 45.08 | $2.07 |
| NVIDIA A100 40GB PCIe | $0.45/hr | 42.57 | $2.93 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 65.42 | $4.57 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 43.58 | $6.04 |
| NVIDIA L40S | $0.79/hr | 34.95 | $6.28 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 76.59 | $7.75 |
| NVIDIA L4 | $0.44/hr | 12.84 | $9.52 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Gemma 4 31B. 30+ tok/s: 9 (RTX 5090, RTX 4090, RTX 3090 Ti); 10-30 tok/s: 2 (A10G, L4). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before Gemma 4 31B writes anything it reads the input: 4020.6 tok/s on the RTX 5090 (1.0s for a 4,000-token prompt), 2915.6 on the RTX 4090 (1.4s), 704.4 on the L4 (5.7s). Long documents and big code files feel this number more than the generation speed.
VRAM for Gemma 4 31B. Measured peak 18.3GB, so 24GB is the smallest common card size; smallest card it ran on: RTX 4090 (24GB). With long context: Q4_K_M 22GB (tested), Q3_K_M 18GB, Q5_K_M 27GB, Q6_K 31GB, Q8_0 40GB.
Power on Gemma 4 31B. Most efficient: H100 80GB HBM3, 324W, 1.18 kWh per 1M generated tokens. Hungriest: RTX 5090, 439W, 1.71 kWh. At $0.15/kWh: $0.18 per 1M generated tokens.
Our verdict
Fastest on Gemma 4 31B: NVIDIA H100 80GB HBM3, 76.59 tok/s. Best desktop card: NVIDIA GeForce RTX 5090, 71.42 tok/s. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3090 ($1,499, 39.2 tok/s). Cheapest to rent per job: NVIDIA GeForce RTX 3090, $0.86 per 1M generated tokens.
FAQ
- What GPU do I need to run Gemma 4 31B?
- About 19GB. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3090 (24GB, 39.2 tok/s).
- How fast is Gemma 4 31B on the NVIDIA GeForce RTX 5090?
- 71.42 tok/s at 439W, 93% of the NVIDIA H100 80GB HBM3.
- How much does it cost to run Gemma 4 31B in the cloud?
- $0.86 per 1M generated tokens on a NVIDIA GeForce RTX 3090 at $0.12/hr, cheapest of 10 rentable cards we measured.
- Can I run Gemma 4 31B on a 12GB, 16GB or 24GB card?
- It used 18.3GB at the precision we tested. 12GB: no; 16GB: no; 24GB: yes.
- Is the RTX 4090 or the RTX 3090 faster for Gemma 4 31B?
- The RTX 4090: 45.08 vs 39.2 tok/s, 15% faster on our bench.
- Should I buy or rent a RTX 5090 for Gemma 4 31B?
- Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 1,321x 1M generated tokens of Gemma 4 31B. 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.