LoRA fine-tuning · 4 models · 11 GPUs measured first-party · Updated October 2026

Best GPU for LoRA fine-tuning

Which graphics card to use for LoRA fine-tuning, from first-party measurements of Qwen2.5 1.5B LoRA, SmolLM2 1.7B LoRA, TinyLlama 1.1B LoRA and more on 11 GPUs.

Fastest we measured

NVIDIA B300

22351.1 train tok/s on Qwen2.5 1.5B LoRA, the ceiling. Measured on our bench. 288GB of VRAM, $40,000 at launch.

Pros

• 22351.1 train tok/s on Qwen2.5 1.5B LoRA
• 288GB, clears the Qwen2.5 1.5B LoRA floor
• Rentable by the hour rather than bought

Cons

• 1400W board rating
• Datacenter or workstation hardware, not a retail purchase

Cheapest card that runs it

NVIDIA RTX PRO 6000 Blackwell Workstation Edition

15753.9 train tok/s on Qwen2.5 1.5B LoRA, lowest launch price that still fits. Measured on our bench. 96GB of VRAM, $8,565 at launch.

Pros

• 15753.9 train tok/s on Qwen2.5 1.5B LoRA
• 96GB, clears the Qwen2.5 1.5B LoRA floor
• Rentable by the hour rather than bought

Cons

• 600W board rating
• Datacenter or workstation hardware, not a retail purchase

LoRA fine-tuning teaches an existing language model new behaviour by training a small set of extra weights instead of the whole model. It is how most people customise a model on their own data, and unlike chatting with a model it keeps the GPU at full load for the whole run, often for hours.

We measured 4 models for LoRA fine-tuning on 11 GPUs. Speed is training tokens processed per second. Every number below is a first-party run on our own harness; cards absent from a model's chart have not been run on it yet.

What the numbers show.

Qwen2.5 1.5B LoRA: fastest on the NVIDIA B300 at 22351.1 train tok/s, 15.58x the slowest card we measured (NVIDIA T4); it used about 8.2GB of VRAM.

SmolLM2 1.7B LoRA: fastest on the NVIDIA B300 at 26114.3 train tok/s, 16.72x the slowest card we measured (NVIDIA T4); it used about 8.2GB of VRAM.

TinyLlama 1.1B LoRA: fastest on the NVIDIA B300 at 26938.2 train tok/s, 17.72x the slowest card we measured (NVIDIA T4); it used about 4GB of VRAM.

Qwen2.5 7B LoRA: fastest on the NVIDIA B300 at 14323.1 train tok/s, 13.46x the slowest card we measured (NVIDIA L4); it used about 20.6GB of VRAM.

Which model to pick. On the same card, the NVIDIA B300, TinyLlama 1.1B LoRA runs at 26938.2 train tok/s in about 4GB; SmolLM2 1.7B LoRA runs at 26114.3 train tok/s in about 8.2GB; Qwen2.5 1.5B LoRA runs at 22351.1 train tok/s in about 8.2GB; Qwen2.5 7B LoRA runs at 14323.1 train tok/s in about 20.6GB. TinyLlama 1.1B LoRA gets through the work 1.9x as fast as Qwen2.5 7B LoRA, so the model you choose moves the speed as much as the card does.

How it compares. B300: Qwen2.5 1.5B LoRA 22351.1 train tok/s, SmolLM2 1.7B LoRA 26114.3, TinyLlama 1.1B LoRA 26938.2, Qwen2.5 7B LoRA 14323.1. 2 of 3 beat Qwen2.5 1.5B LoRA here.

Cost on a rented GPU. 100M training tokens of Qwen2.5 1.5B LoRA: $1.81 on a L40S ($0.79/hr, 2.3 hours), $8.62 on a B300 ($6.94/hr, 75 min, 4.8x the cost).

Speed tiers for Qwen2.5 1.5B LoRA. 10000+ train tok/s: 6 (B300, H100 80GB HBM3, RTX PRO 6000 Blackwell Workstation Edition); 3000-10000 train tok/s: 4 (A100 80GB SXM4, A100 40GB SXM4, A10G); under 3000 train tok/s: 1 (T4). 10,000 train tok/s gets through 100M tokens in under three hours.

VRAM for Qwen2.5 1.5B LoRA. Measured peak 7.8GB, so 12GB is the smallest common card size; smallest card it ran on: T4 (16GB).

Power on Qwen2.5 1.5B LoRA. Most efficient: B300, 371W, 0.46 kWh per 100M training tokens. Hungriest: B200, 388W, 0.70 kWh. At $0.15/kWh: $0.069 per 100M training tokens.

Our verdict

For LoRA fine-tuning, the NVIDIA B300 is the fastest card we measured. We have not measured a consumer card on this job yet; the picks above are datacenter and workstation hardware. Check the fit table before buying: VRAM, not speed, is what rules a card out.

FAQ

What is the fastest GPU for LoRA fine-tuning?

In our runs, the NVIDIA B300 at 26938.2 train tok/s on TinyLlama 1.1B LoRA. We measured 4 models on 11 GPUs for this page.

How much VRAM do I need for LoRA fine-tuning?

The lightest model here, TinyLlama 1.1B LoRA, used about 4GB. The table above shows which models fit 8, 12, 16, 24 and 32GB cards, from measured peaks.

Are these numbers measured or estimated?

Measured. Every number on this page is a first-party run on our own harness, with power and VRAM sampled during the run. Cards we have not run yet are simply absent, not filled in.

What GPU do I need to run Qwen2.5 1.5B LoRA?

About 8GB. Smallest card that ran it: NVIDIA T4 (16GB).

How much does it cost to run Qwen2.5 1.5B LoRA in the cloud?

$1.81 per 100M training tokens on a NVIDIA L40S at $0.79/hr, cheapest of 10 rentable cards we measured.

Can I run Qwen2.5 1.5B LoRA on a 12GB, 16GB or 24GB card?

It used 7.8GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.

Is the H100 80GB HBM3 or the A100 80GB SXM4 faster for Qwen2.5 1.5B LoRA?

The H100 80GB HBM3: 16034.9 vs 8252.8 train tok/s, 94% faster on our bench.

How we test

Each model runs a fixed workload on every card: a warmup, then timed runs with power, temperature and VRAM sampled every half second through NVML. Models run at the precision and settings from their model card. Datacenter cards run on Modal; consumer cards on rented machines. Non-commercially licensed models are not part of this page.