DeepSeek-R1-Distill-Llama-70B · 7 GPUs measured first-party · llama.cpp · Updated October 2026
What GPU Do You Need for DeepSeek-R1-Distill-Llama-70B?
DeepSeek-R1-Distill-Llama-70B on 7 GPUs, measured first-party: NVIDIA B300 leads at 48.11 tok/s, L40S trails at 16.5 tok/s, and it peaked at 43GB of VRAM.
Benchmarked weights: unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF
NVIDIA B300
48.11 tok/s on DeepSeek-R1-Distill-Llama-70B, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.
- 48.11 tok/s on DeepSeek-R1-Distill-Llama-70B
- 288GB, clears the DeepSeek-R1-Distill-Llama-70B 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
32.52 tok/s on DeepSeek-R1-Distill-Llama-70B, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.
- 32.52 tok/s on DeepSeek-R1-Distill-Llama-70B
- 96GB, clears the DeepSeek-R1-Distill-Llama-70B 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 DeepSeek-R1-Distill-Llama-70B?, 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.
DeepSeek-R1-Distill-Llama-70B. Measured tokens per second by GPU
| GPU | tok/s | Prompt t/s | tok/W | Avg power |
|---|---|---|---|---|
| NVIDIA B300 | 48.11 | 670.7 | 0.09 | 551.1 W |
| NVIDIA B200 | 44.49 | 1253.9 | 0.07 | 603.6 W |
| NVIDIA H200 | 42.76 | 1177.6 | 0.11 | 375.3 W |
| NVIDIA H100 80GB HBM3 | 41.17 | 1212.7 | 0.11 | 379.8 W |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 32.52 | 1772.5 | 0.09 | 358.2 W |
| NVIDIA A100 80GB SXM4 | 22.99 | 547.2 | 0.1 | 229.1 W |
| NVIDIA L40S | 16.49 | 1175.6 | 0.06 | 265.3 W |
What the numbers show. Across 7 GPUs measured on our own bench, B300 is fastest at 48.1 tok/s. The slowest, L40S, manages 16.5, so the spread is 2.9x from top to bottom. H200 is the most efficient, 42.8 tok/s at 375W. Per dollar of launch price, RTX PRO 6000 Blackwell Workstation Edition gives the most (3.8 tok/s per $1,000).
How it compares. H100 80GB HBM3: DeepSeek-R1-Distill-Llama-70B 41.17 tok/s, Meta-Llama-3.1-70B 41.19, Llama-3.3-70B-Instruct-abliterated 41.15, Hermes-4-70B 41.2, Llama-3.3-70B 41.0. 2 of 4 beat DeepSeek-R1-Distill-Llama-70B here.
Cost on a rented GPU. 1M generated tokens of DeepSeek-R1-Distill-Llama-70B: $9.19 on a RTX PRO 6000 Blackwell Workstation Edition ($1.08/hr, 8.5 hours), $40.07 on a B300 ($6.94/hr, 5.8 hours, 4.4x the cost).
DeepSeek-R1-Distill-Llama-70B: cost per 1M generated tokens on rented GPUs
| GPU | Cheapest rate | Speed (tok/s) | Cost per 1M generated tokens |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | $1.08/hr | 32.52 | $9.19 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 22.99 | $11.44 |
| NVIDIA L40S | $0.79/hr | 16.49 | $13.31 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 41.17 | $14.41 |
| NVIDIA H200 | $3.59/hr | 42.76 | $23.32 |
| NVIDIA B200 | $5.98/hr | 44.49 | $37.34 |
| NVIDIA B300 | $6.94/hr | 48.11 | $40.07 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for DeepSeek-R1-Distill-Llama-70B. 30+ tok/s: 5 (B300, B200, H200); 10-30 tok/s: 2 (A100 80GB SXM4, L40S). 30 tok/s is roughly where replies outpace reading.
Reading your prompt. Before DeepSeek-R1-Distill-Llama-70B writes anything it reads the input: 1772.5 tok/s on the RTX PRO 6000 Blackwell Workstation Edition (2.3s for a 4,000-token prompt), 547.2 on the A100 80GB SXM4 (7.3s). Long documents and big code files feel this number more than the generation speed.
VRAM for DeepSeek-R1-Distill-Llama-70B. Measured peak 40.5GB, so 48GB is the smallest common card size; smallest card it ran on: L40S (48GB). With long context: Q4_K_M 54GB (tested), Q2_K 34GB, Q3_K_M 44GB, Q5_K_M 64GB, Q8_0 96GB.
Power on DeepSeek-R1-Distill-Llama-70B. Most efficient: H200, 375W, 2.44 kWh per 1M generated tokens. Hungriest: B200, 604W, 3.77 kWh. At $0.15/kWh: $0.37 per 1M generated tokens.
Our verdict
Fastest on DeepSeek-R1-Distill-Llama-70B: NVIDIA B300, 48.11 tok/s. Cheapest to rent per job: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, $9.19 per 1M generated tokens.
FAQ
- What GPU do I need to run DeepSeek-R1-Distill-Llama-70B?
- About 43GB. Smallest card that ran it: NVIDIA L40S (48GB).
- How much does it cost to run DeepSeek-R1-Distill-Llama-70B in the cloud?
- $9.19 per 1M generated tokens on a NVIDIA RTX PRO 6000 Blackwell Workstation Edition at $1.08/hr, cheapest of 7 rentable cards we measured.
- Can I run DeepSeek-R1-Distill-Llama-70B on a 12GB, 16GB or 24GB card?
- It used 40.5GB at the precision we tested. 12GB: no; 16GB: no; 24GB: no.
- Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for DeepSeek-R1-Distill-Llama-70B?
- The H100 80GB HBM3: 41.17 vs 22.99 tok/s, 79% 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.