Power draw matters on a dedicated server more than on a desktop because you are paying for it continuously. On our UK hosting that cost is bundled into your monthly price, but it still shapes what we can provision together and which workloads pay back quickest. Here is every GPU in the lineup ranked by TDP.
Contents
TDP Ordered
| GPU | TDP | VRAM |
|---|---|---|
| RTX 5090 | 575 W | 32 GB |
| RTX 5080 | 360 W | 16 GB |
| RTX 3090 | 350 W | 24 GB |
| RTX 6000 Pro | ~300 W | 96 GB |
| R9700 | ~260 W | 32 GB |
| AMD RX 9070 XT | ~250 W | 16 GB |
| Intel Arc Pro B70 | ~220 W | 32 GB |
| RTX 4060 Ti | 165 W | 16 GB |
| RTX 5060 | 150 W | 8 GB |
| RTX 4060 | 115 W | 8 GB |
| Ryzen AI Max+ 395 (SoC) | ~120 W | 96 GB unified |
| RTX 3050 | 115 W | 6 GB |
Real vs Datasheet
TDP is a thermal design number, not a measured draw. At idle even the biggest cards pull 20-40 W. Under sustained AI load most cards run 85-100% of their rated TDP. LLM decode tends to be lighter than prefill or training – decoding Llama 3 8B on a 5090 pulls maybe 300 W, not 575 W. Batch prefill or training pushes close to the ceiling.
Tokens Per Watt
The interesting ranking is efficiency. For Llama 3 8B INT8 decode, rough tokens per watt:
| GPU | Rough tokens/sec/W |
|---|---|
| RTX 6000 Pro | ~0.30 |
| RTX 5090 | ~0.27 |
| RTX 5080 | ~0.26 |
| RTX 4060 Ti | ~0.22 |
| RTX 3090 | ~0.18 |
Newer silicon is more efficient even if the TDP numbers look higher. A 5090 draws more watts but puts them to better use. See our tokens per watt deep dive.
Fixed Monthly Pricing Includes Power
No metered electricity bills, no cloud pricing volatility – one number per month on our UK hosting.
Browse GPU ServersWhat It Means
First, the 5090 is the hungriest card in the lineup. Two of them in one box need a serious PSU and we price that accordingly. Second, the 6000 Pro is surprisingly efficient given its capacity – a huge card that runs cooler than a 5090. Third, the entry cards (3050, 4060, 5060) can fit in compact servers with modest power infrastructure, which is why they dominate the “your first AI server” bucket.
For purchasing strategy, see VRAM per pound.
