Llama 3.3 70B on an RTX 4090
No fine-tuning method fits Llama 3.3 70B on one RTX 4090 (24 GB) at 2,048 tokens. 4-bit inference: no, the weights alone are 40 GB.
Verdict by method
| Fine-tuning | Memory at 2K tokens | One RTX 4090? | Longest sequence (batch 1) | Largest batch at 2K | Largest LoRA rank |
|---|---|---|---|---|---|
| QLoRA (4-bit base) | 48 GB | No | — | — | — |
| LoRA (16-bit base) | 150 GB | No | — | — | — |
| Full fine-tuning | 1,134 GB | No | — | — | n/a |
| Inference | Memory at 8K | One RTX 4090? | Longest context | Concurrent 4K sequences |
|---|---|---|---|---|
| 16-bit (BF16/FP16) | 144 GB | No | — | — |
| 8-bit | 73 GB | No | — | — |
| 4-bit | 42 GB | No | — | — |
- QLoRA: no, it needs about 48 GB at 2K tokens, more than the 24 GB available.
- LoRA: no, it needs about 150 GB at 2K tokens, more than the 24 GB available.
- Full fine-tuning: no, it needs about 1,134 GB at 2K tokens, more than the 24 GB available.
Where the memory goes
| Method | Weights | Training state | Activations | Total | On one RTX 4090 | 80 GB GPUs needed |
|---|---|---|---|---|---|---|
| QLoRA (4-bit base) | 40 GB | 3.3 GB | 5.6 GB | 48 GB | Doesn't fit | 1 |
| LoRA (16-bit base) | 141 GB | 3.3 GB | 5.6 GB | 150 GB | Doesn't fit | 3 |
| Full fine-tuning | 141 GB | 988 GB | 5.6 GB | 1,134 GB | Doesn't fit | 16 |
| Precision | Weights | KV cache at 8K | Total | At 128K context | On one RTX 4090 |
|---|---|---|---|---|---|
| 16-bit (BF16/FP16) | 141 GB | 2.7 GB | 144 GB | 184 GB | Doesn't fit |
| 8-bit | 71 GB | 2.7 GB | 73 GB | 114 GB | Doesn't fit |
| 4-bit | 40 GB | 2.7 GB | 42 GB | 83 GB | Doesn't fit |
The cheapest way to run it
Renting an RTX 4090 costs $0.89 an hour on Runpod Secure Cloud ($0.34 on Runpod Community Cloud). Since no fine-tuning method fits, the table shows the cheapest hardware that does.
| Job (default settings) | Memory | Cheapest rental that fits | Per hour |
|---|---|---|---|
| QLoRA fine-tuning | 48 GB | A100 80GB | $2.19 |
| LoRA fine-tuning | 150 GB | 4 × L40S | $4.36 |
| 4-bit inference | 42 GB | L40S | $1.09 |
| 16-bit inference | 144 GB | 2 × A100 80GB | $4.38 |
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.
Try other settings
| Job | Memory | RTX 4090 (24 GB) | Smallest single GPU |
|---|---|---|---|
| QLoRA (4-bit base)2K tokens × batch 1, rank 16 | 48 GB | 3 GPUs | A100 80GB (80 GB) |
| LoRA (16-bit base)2K tokens × batch 1, rank 16 | 150 GB | 7 GPUs | B200 (180 GB) |
| Full fine-tuning2K tokens × batch 1 | 1,134 GB | 53 GPUs | None: needs several |
| Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence | 144 GB | 7 GPUs | B200 (180 GB) |
| Inference, 8-bit8K-token context × 1 sequence | 73 GB | 4 GPUs | H200 (141 GB) |
| Inference, 4-bit8K-token context × 1 sequence | 42 GB | 2 GPUs | L40S (48 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 RTX 4090's dense BF16 peak of 165.2 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 Llama 3.3 70B on an RTX 4090?
- QLoRA: no, it needs about 48 GB at 2K tokens, more than the 24 GB available. LoRA: no, it needs about 150 GB at 2K tokens, more than the 24 GB available. Full fine-tuning: no, it needs about 1,134 GB at 2K tokens, more than the 24 GB available.
- Can I run Llama 3.3 70B on an RTX 4090?
- 16-bit (BF16/FP16) inference: no, the weights alone are 141 GB. 8-bit inference: no, the weights alone are 71 GB. 4-bit inference: no, the weights alone are 40 GB.
Related
- Llama 3.3 70B VRAM requirementsEvery GPU and method
- Llama 3.3 70B on H100 SXM80 GB
- Llama 3.3 70B on H200141 GB
- Fine-tuning Llama 3.3 70B: costPer million tokens
- Llama 3.1 8B on RTX 4090Meta
- Llama 3.1 70B on RTX 4090Meta
- Llama 3.2 1B on RTX 4090Meta
- Llama 3.2 3B on RTX 4090Meta
- Qwen3 4B on RTX 4090Alibaba Qwen
- Qwen3 8B on RTX 4090Alibaba Qwen
Sources
- Hugging Face. meta-llama/Llama-3.3-70B-Instruct config.json
- Meta. meta-llama/Llama-3.3-70B-Instruct model card
- Meta. The Llama 3 Herd of Models (arXiv:2407.21783)
- NVIDIA. RTX 4090 specifications
- Runpod. GPU pricing (updated 12 November 2026)
- Lambda. GPU cloud pricing
- Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
- Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
- Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
- 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 Llama 3.3 70B is worth the GPU time.
Start training free