GPU · 1307.4 TFLOPS dense BF16
MI300X fine-tuning cost estimates
All sixteen scenarios on the MI300X: GPU-hours, wall-clock on 8 GPUs and the cost range across 4 providers, at 35% MFU. Prices read 12 November 2026. Estimates, not quotes.
The figure
Where 1307.4 TFLOPS comes from
“1307.4 TFLOPS peak theoretical Bfloat16 format precision (BF16)”
Scenarios
Cost by scenario
| Scenario | GPU-hours | 8 GPUs | Lowest | Median | Highest |
|---|---|---|---|---|---|
| 7B LoRA fine-tune on 100 million tokens | 1.7 | 13 min | $3.14 | $5.13 | $10 |
| 7B LoRA fine-tune on 1 billion tokens | 17 | 2.1 h | $31 | $51 | $102 |
| 7B full fine-tune on 100 million tokens | 2.55 | 19 min | $4.72 | $7.70 | $15 |
| 7B full fine-tune on 1 billion tokens | 25 | 3.2 h | $47 | $77 | $153 |
| 13B LoRA fine-tune on 100 million tokens | 3.16 | 24 min | $5.84 | $9.53 | $19 |
| 13B LoRA fine-tune on 1 billion tokens | 32 | 3.9 h | $58 | $95 | $189 |
| 13B full fine-tune on 100 million tokens | 4.73 | 36 min | $8.76 | $14 | $28 |
| 13B full fine-tune on 1 billion tokens | 47 | 5.9 h | $88 | $143 | $284 |
| 34B LoRA fine-tune on 100 million tokens | 8.26 | 1.0 h | $15 | $25 | $50 |
| 34B LoRA fine-tune on 1 billion tokens | 83 | 10 h | $153 | $249 | $495 |
| 34B full fine-tune on 100 million tokens | 12 | 1.5 h | $23 | $37 | $74 |
| 34B full fine-tune on 1 billion tokens | 124 | 15 h | $229 | $374 | $743 |
| 70B LoRA fine-tune on 100 million tokens | 17 | 2.1 h | $31 | $51 | $102 |
| 70B LoRA fine-tune on 1 billion tokens | 170 | 21 h | $314 | $513 | $1,020 |
| 70B full fine-tune on 100 million tokens | 25 | 3.2 h | $47 | $77 | $153 |
| 70B full fine-tune on 1 billion tokens | 255 | 32 h | $472 | $770 | $1,530 |
Prices used
MI300X price by provider
| Provider | $/GPU-hr | Configuration | Source |
|---|---|---|---|
| Vultr | $1.85 | 8× | price page |
| DigitalOcean (Paperspace) | $2.59 | 1× | price page |
| Crusoe | $3.45 | 1× | price page |
| Oracle Cloud | $6.00 | 1× | price page |
Each provider's lowest listed on-demand configuration for this GPU. Marketplace figures are left out.
Questions
MI300X: common questions
What is the lowest cost scenario on the MI300X?
7B LoRA fine-tune on 100 million tokens: about $3.14 on Vultr, 1.7 GPU-hours.
What throughput is assumed?
1307.4 TFLOPS dense BF16 as printed by the maker, times 35% MFU, so about 458 TFLOPS of useful work.
How many providers list it?
4 with a current provider-set on-demand price. Runcrate has no live listing for it.
Related