…
Cloud rental price, compute per dollar and specs, side by side.
Open in the interactive comparer →| Spec / price | RTX 3090 | RTX 4090 |
|---|---|---|
| Memory | 24 GB GDDR6X | 24 GB GDDR6X |
| Memory bandwidth | 936 GB/s | 1,008 GB/sbetter |
| FP16 / BF16 (dense) | 0.071 PF | 0.165 PFbetter |
| FP8 (dense) | — | 0.33 PF |
| FP4 (dense) | — | — |
| INT8 (dense) | 0.284 PF | 0.661 PFbetter |
| TF32 (dense) | 0.0356 PF | 0.0826 PFbetter |
| FP32 (dense) | 0.0356 PF | 0.0826 PFbetter |
| FP64 (dense) | 0.0006 PF | 0.0013 PFbetter |
| Power | 350 W | 450 W |
| Architecture | Ampere (2020) | Ada Lovelace (2022) |
| Lowest on-demand | $0.13better | $0.34 |
| Median on-demand | $0.17better | $0.40 |
| Lowest spot | — | — |
| FP16/BF16 $ per PFLOP·h | $1.76better | $2.06 |
| Providers | 2 | 5 |
The RTX 4090 has 2.3× more dense FP16/BF16 compute than the RTX 3090. The RTX 4090 has 1.1× more memory bandwidth than the RTX 3090. Real-world speedups depend on the workload: training and prefill scale with compute, while LLM decoding scales mostly with memory bandwidth.
Right now the RTX 3090 is cheaper per unit of compute: $1.76 vs $2.06 per PFLOP-hour at the lowest on-demand prices.
Running one GPU around the clock at the lowest on-demand price costs about $91 per month for the RTX 3090 and $248 for the RTX 4090.
Both have 24 GB of VRAM; the RTX 4090 has more bandwidth.