Llama-2-7B · 11 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for Llama-2-7B?
Llama-2-7B on 11 GPUs, measured first-party: NVIDIA B300 leads at 315.9 tok/s, T4 trails at 43 tok/s, and it peaked at 5GB of VRAM.
Benchmarked weights: TheBloke/Llama-2-7B-Chat-GGUF
NVIDIA B300
315.9 tok/s on Llama-2-7B, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.
- 315.9 tok/s on Llama-2-7B
- 288GB, clears the Llama-2-7B floor
- Rentable by the hour rather than bought
- 1400W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA RTX PRO 6000 Blackwell Workstation Edition
263.8 tok/s on Llama-2-7B, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 263.8 tok/s on Llama-2-7B
- 96GB, clears the Llama-2-7B floor
- Rentable by the hour rather than bought
- 600W board rating
- Datacenter or workstation hardware, not a retail purchase
What GPU Do You Need for Llama-2-7B?, 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.
Llama-2-7B. Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA B300 | 315.9 | 6180.1 | 0.83 | 378.9 W |
| NVIDIA B200 | 292 | 9698.8 | 0.75 | 390.8 W |
| NVIDIA H200 | 290.8 | 9389.5 | 1.33 | 218.0 W |
| NVIDIA H100 80GB HBM3 | 284.6 | 9774.8 | 1.27 | 224.5 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 263.8 | 13343.7 | 1.09 | 241.1 W |
| NVIDIA A100 80GB SXM4 | 180.1 | 4670.6 | 0.8 | 226.4 W |
| NVIDIA A100 40GB SXM4 | 172.5 | 4629.9 | 0.88 | 195.0 W |
| NVIDIA L40S | 150.3 | 10585.5 | 0.74 | 202.3 W |
| NVIDIA A10G | 96.69 | 3427.5 | 0.8 | 121.4 W |
| NVIDIA L4 | 56.06 | 3185.3 | 0.9 | 62.5 W |
| NVIDIA T4 | 42.98 | 1241.9 | 0.69 | 62.0 W |
What the numbers show. Across 11 GPUs measured on our own bench, B300 is fastest at 316 tok/s. The slowest, T4, manages 43.0, so the spread is 7.3x from top to bottom. H200 is the most efficient, 291 tok/s at 218W. Per dollar of launch price, A10G gives the most (34.5 tok/s per $1,000). The fastest card with 16GB or less is T4 at 43.0 tok/s.
About Llama-2-7B. Llama-2-7B: from meta-llama, 6.7B parameters, on Hugging Face since July 2023, llama2 licence (gated: accept the terms first). 1,198,490 downloads in the last 30 days and 1 community quantizations.
How it compares. H100 80GB HBM3: Llama-2-7B 284.6 tok/s, DeepSeek Coder 7B Instruct v1.5 293.9 (7B), Mistral-7B-Instruct-v0.2 276.7 (7B), Qwen2.5-7B 264.2 (8B), Qwen2.5-Coder 7B 263.9 (8B). 1 of 4 beat Llama-2-7B here.
Cost on a rented GPU. 1M generated tokens of Llama-2-7B: $0.76 on a A100 40GB SXM4 ($0.47/hr, 97 min), $6.10 on a B300 ($6.94/hr, 53 min, 8.0x the cost).
Llama-2-7B: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA A100 40GB SXM4 | $0.47/hr | 172.5 | $0.76 |
| NVIDIA T4 | $0.14/hr | 42.98 | $0.88 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 263.8 | $1.13 |
| NVIDIA L40S | $0.79/hr | 150.3 | $1.46 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 180.1 | $1.46 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 284.6 | $2.08 |
| NVIDIA L4 | $0.44/hr | 56.06 | $2.18 |
| NVIDIA H200 | $3.59/hr | 290.8 | $3.43 |
| NVIDIA B200 | $5.98/hr | 292 | $5.69 |
| NVIDIA B300 | $6.94/hr | 315.9 | $6.10 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for Llama-2-7B. 30+ tok/s: 11 (B300, B200, H200). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before Llama-2-7B writes anything it reads the input: 13343.7 tok/s on the RTX PRO 6000 Blackwell Workstation Edition (0.3s for a 4,000-token prompt), 1241.9 on the T4 (3.2s). Long documents and big code files feel this number more than the generation speed.
VRAM for Llama-2-7B. Measured peak 4.6GB, so 8GB is the smallest common card size; smallest card it ran on: T4 (16GB). With long context: Q4_K_M 6GB (tested), Q2_K 5GB, Q3_K_M 5GB, Q5_K_M 8GB, Q6_K 9GB.
Power on Llama-2-7B. Most efficient: H200, 218W, 0.21 kWh per 1M generated tokens. Hungriest: B200, 391W, 0.37 kWh. At $0.15/kWh: $0.031 per 1M generated tokens.
Our verdict
Fastest on Llama-2-7B: NVIDIA B300, 315.9 tok/s. Cheapest to rent per job: NVIDIA A100 40GB SXM4, $0.76 per 1M generated tokens.
FAQ
- What GPU do I need to run Llama-2-7B?
- About 5GB. Smallest card that ran it: NVIDIA T4 (16GB).
- How much does it cost to run Llama-2-7B in the cloud?
- $0.76 per 1M generated tokens on a NVIDIA A100 40GB SXM4 at $0.47/hr, cheapest of 10 rentable cards we measured.
- Can I run Llama-2-7B on a 12GB, 16GB or 24GB card?
- It used 4.6GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for Llama-2-7B?
- The H100 80GB HBM3: 284.6 vs 180.1 tok/s, 58% faster on our bench.
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.