Skip to content

VRAM reference

Gemma 3 27B on an L40S

Yes, you can fine-tune Gemma 3 27B on one L40S (48 GB) with QLoRA. 4-bit inference: yes, 15 GB of weights, room for 128K tokens of context or 35 concurrent 4K-token sequences.

By Will Schulz · Checked 15 November 2026 · Prices as of November 2026

Verdict by method

Fine-tuningMemory at 2K tokensOne L40S?Longest sequence (batch 1)Largest batch at 2KLargest LoRA rank
QLoRA (4-bit base)24 GBYes8K tokens4128
LoRA (16-bit base)63 GBNo———
Full fine-tuning445 GBNo——n/a
InferenceMemory at 8KOne L40S?Longest contextConcurrent 4K sequences
16-bit (BF16/FP16)56 GBNo——
8-bit29 GBYes128K tokens20
4-bit17 GBYes128K tokens35
  • QLoRA: yes, about 24 GB at 2K tokens, batch 1, rank 16; up to 8K-token sequences at batch 1 or batch 4 at 2K, LoRA rank up to 128.
  • LoRA: no, it needs about 63 GB at 2K tokens, more than the 48 GB available.
  • Full fine-tuning: no, it needs about 445 GB at 2K tokens, more than the 48 GB available.

Where the memory goes

At 2,048 tokens per sequence, batch 1, LoRA rank 16, gradient checkpointing on
MethodWeightsTraining stateActivationsTotalOn one L40S80 GB GPUs needed
QLoRA (4-bit base)16 GB1.8 GB6.2 GB24 GBFits1
LoRA (16-bit base)55 GB1.8 GB6.2 GB63 GBDoesn't fit1
Full fine-tuning55 GB384 GB6.2 GB445 GBDoesn't fit7
One sequence, fp16 KV cache
PrecisionWeightsKV cache at 8KTotalAt 128K contextOn one L40S
16-bit (BF16/FP16)55 GB1.1 GB56 GB66 GBDoesn't fit
8-bit27 GB1.1 GB29 GB39 GBFits
4-bit15 GB1.1 GB17 GB27 GBFits

The cheapest way to run it

Renting an L40S costs $1.09 an hour on Runpod Secure Cloud ($0.79 on Runpod Community Cloud). QLoRA on it processes about 1,029 tokens a second by our estimate, so 10 million training tokens take 2.7 hours and cost about $2.94.

Job (default settings)MemoryCheapest rental that fitsPer hour
QLoRA fine-tuning24 GBL40S (this GPU)$1.09
LoRA fine-tuning63 GBA100 80GB$2.19
4-bit inference17 GBRTX 3090$0.50
16-bit inference56 GBA100 80GB$2.19

On-demand rates seen on 15 November 2026: Runpod Secure Cloud, Lambda for the A100 40GB. Consumer cards count as single GPUs only; data-centre GPUs up to eight per node. Kojen's GPU training can launch the L40S on Runpod, Vast.ai or Lambda, or pick the cheapest GPU in stock for you.

Try other settings

JobMemoryL40S (48 GB)Smallest single GPU
QLoRA (4-bit base)2K tokens × batch 1, rank 1624 GBFitsRTX 5090 (32 GB)
LoRA (16-bit base)2K tokens × batch 1, rank 1663 GB2 GPUsA100 80GB (80 GB)
Full fine-tuning2K tokens × batch 1445 GB11 GPUsNone: needs several
Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence56 GB2 GPUsA100 80GB (80 GB)
Inference, 8-bit8K-token context × 1 sequence29 GBFitsRTX 5090 (32 GB)
Inference, 4-bit8K-token context × 1 sequence17 GBFitsRTX 3090 (24 GB)

A job fits when the estimate is within 90% of the card's memory. Training assumes gradient checkpointing and FlashAttention; multi-GPU counts assume the memory splits evenly (FSDP, ZeRO-3 or tensor parallelism).

How we estimate it

Every figure comes from the model's published architecture (layers, hidden size, attention and KV heads, vocabulary) rather than a rule of thumb, using the same functions as the calculator above. Sizes are in GB (10⁹ bytes).

  • Full fine-tuning: 16 bytes per parameter for mixed-precision Adam (bf16 weights and gradients, fp32 master weights and two moments), as in the ZeRO paper.
  • LoRA: frozen bf16 weights (2 bytes per parameter) plus 16 bytes per adapter parameter, rank 16 on every linear layer. For mixture-of-experts models, adapters go on attention only and the experts stay frozen.
  • QLoRA: NF4 weights with double quantisation (4.127 bits per parameter); embeddings and the output head stay in bf16, as bitsandbytes leaves them.
  • Activations: gradient checkpointing and FlashAttention. Each layer's bf16 input is kept, one layer is recomputed at a time, and the output logits take 8 bytes per vocabulary entry per token (fp32 values and gradients). Chunked cross-entropy, as in Unsloth or Liger kernels, cuts that last term.
  • Inference: 2, 1 or about 0.56 bytes per weight (4-bit with group-wise scales), plus an fp16 KV cache of 2 × layers × KV heads × head dim × 2 bytes per token; sliding-window layers only cache their window.
  • Fits means the estimate is within 90% of the card's listed memory; the rest covers the CUDA context, kernels and fragmentation. Multi-GPU counts assume memory splits evenly.

Speed estimates use L40S's dense BF16 peak of 362 TFLOPS, about 4 FLOPs per active parameter per token for LoRA and 40% utilisation; see fine-tuning and training costs for the method.

Frequently asked questions

Can I fine-tune Gemma 3 27B on an L40S?
QLoRA: yes, about 24 GB at 2K tokens, batch 1, rank 16; up to 8K-token sequences at batch 1 or batch 4 at 2K, LoRA rank up to 128. LoRA: no, it needs about 63 GB at 2K tokens, more than the 48 GB available. Full fine-tuning: no, it needs about 445 GB at 2K tokens, more than the 48 GB available.
Can I run Gemma 3 27B on an L40S?
16-bit (BF16/FP16) inference: no, the weights alone are 55 GB. 8-bit inference: yes, 27 GB of weights, room for 128K tokens of context or 20 concurrent 4K-token sequences. 4-bit inference: yes, 15 GB of weights, room for 128K tokens of context or 35 concurrent 4K-token sequences.
What does fine-tuning Gemma 3 27B on an L40S cost?
With QLoRA at our planning rate of 40% utilisation, about 2.7 hours for 10 million training tokens, around $2.94 at $1.09 an hour (Runpod Secure Cloud, November 2026).

Sources

  1. [1]Hugging Face. google/gemma-3-27b-it config.json
  2. [2]Google DeepMind. google/gemma-3-27b-it model card
  3. [3]Google DeepMind. Gemma 3 Technical Report (arXiv:2503.19786)
  4. [4]NVIDIA. L40S specifications
  5. [5]Runpod. GPU pricing (updated 12 November 2026)
  6. [6]Lambda. GPU cloud pricing
  7. [7]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  8. [8]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  9. [9]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  10. [10]Korthikanti et al., 2022. Reducing Activation Recomputation in Large Transformer Models (arXiv:2205.05198)

Measure before you rent the GPU.

Prompt tuning and model search in Kojen run on hosted models against your own test cases, so you know whether fine-tuning Gemma 3 27B is worth the GPU time.

Start training free