BTC
ETH
USDT
XMR
LTC
GPU comparison · Blackwell vs Ada Lovelace
NVIDIA RTX 5090 vs RTX 4090.
The RTX 5090 (Blackwell) has 32 GB of GDDR7 at 1.8 TB/s against 24 GB of GDDR6X at 1 TB/s on the RTX 4090, plus FP4 tensor cores. For AI servers the memory is the main difference; the 5090 costs $60 more per card per month.
- 32 vs 24 GBGPU memory
- 1.79 vs 1.01 TB/smemory bandwidth
- $329 vs $269per GPU, per month
RTX 5090 vs RTX 4090: which one to rent.
Choose the RTX 5090 if
- You need more than 24 GB: larger image and video models, or LLMs up to about 45B in 4-bit.
- Your inference is memory-bound: 1.78× the bandwidth of the 4090.
- You want FP4 inference on a GeForce card.
Choose the RTX 4090 if
- Your models fit in 24 GB, like SDXL-class image models or LLMs up to about 30B in 4-bit.
- You want the lowest price per card: $269 a month.
- You render with Blender, Octane or Redshift and scale out with more cards rather than more memory.
RTX 5090 vs RTX 4090 specs, side by side.
Figures from NVIDIA’s published specifications, per GPU. Tensor figures are peak dense throughput.
| Per GPU | NVIDIA RTX 5090 | NVIDIA RTX 4090 | RTX 4090 / RTX 5090 |
|---|---|---|---|
| Architecture | Blackwell | Ada Lovelace, TSMC 4N | — |
| Form factor | PCIe Gen5 graphics card | PCIe Gen4 graphics card | — |
| GPU memory | 32 GB GDDR7, 512-bit | 24 GB GDDR6X, 384-bit | 0.75× |
| Memory bandwidth | 1,792 GB/s | 1,008 GB/s | 0.56× |
| GPU-to-GPU link | None (PCIe Gen5) | None (PCIe Gen4) | — |
| FP4 tensor (dense) | 1.7 PFLOPS | Not supported | — |
| FP8 tensor (dense) | 419 TFLOPS | 330 TFLOPS | 0.79× |
| FP16 / BF16 tensor (dense) | 209.5 TFLOPS | 165 TFLOPS | 0.79× |
| FP64 | — | — | — |
| Max power | 575 W | 450 W | — |
| GPUs per server | 1, 2, 4 | 1, 2, 4 | — |
| Per GPU in our servers | 16 vCPU · 64 GB RAM · 1 TB | 12 vCPU · 64 GB RAM · 1 TB | — |
Sources: NVIDIA GeForce RTX 5090 · RTX Blackwell whitepaper · NVIDIA GeForce RTX 4090 · Ada whitepaper. Full sheets: RTX 5090 · RTX 4090
RTX 5090 vs RTX 4090 rental price.
Our monthly prices, set 30% below the market median of public on-demand prices and paid in crypto. Hourly figures are the monthly price divided by 730 hours.
| CryptGPU | NVIDIA RTX 5090 | NVIDIA RTX 4090 | RTX 4090 / RTX 5090 |
|---|---|---|---|
| Price per GPU, per month | $329 | $269 | 0.82× |
| Equivalent per GPU-hour | $0.45 | $0.37 | — |
| Per GB of GPU memory, per month | $10.28 | $11.21 | 1.09× |
| Per PFLOPS of dense FP8, per month | $785 | $815 | 1.04× |
| Largest server | 4× · $1,316/mo | 4× · $1,076/mo | — |
| Market median, per GPU | $481 | $387 | — |
| Below the median | −32% | −30% | — |
Same price per GPU at every server size, no hourly metering. How we compare prices · Configure an RTX 5090 server · Configure a RTX 4090 server
Which models fit on each.
Smallest server that holds each model, by precision. Rule of thumb with headroom for the KV cache, 92% of GPU memory usable.
| Model (total parameters) | RTX 5090 · FP8 | RTX 4090 · FP8 | RTX 5090 · 4-bit | RTX 4090 · 4-bit |
|---|---|---|---|---|
| Qwen3.5-9B | 1× | 1× | 1× | 1× |
| Gemma 4 31B | 2× | 2× | 1× | 1× |
| Llama 3.3 70B | 4× | 4× | 2× | 4× |
| gpt-oss-120b | — | — | 4× | 4× |
| DeepSeek-V4-Flash | — | — | — | — |
| Qwen3.5-397B-A17B | — | — | — | — |
| DeepSeek-R1 (671B) | — | — | — | — |
| Kimi K2.6 (1T) | — | — | — | — |
FP8 ≈ 1.2 bytes and 4-bit ≈ 0.65 bytes per parameter. — = larger than the biggest server of that GPU. Size another model · How much VRAM does an LLM need?
RTX 5090 vs RTX 4090 FAQ.
More on each GPU: NVIDIA RTX 5090 · NVIDIA RTX 4090.
Is the RTX 5090 better than the RTX 4090 for AI?
Yes when memory or bandwidth limits you: 32 GB against 24 GB and 1.8 TB/s against 1 TB/s. Dense FP8 and BF16 throughput is about 27% higher. For models that fit in 24 GB, the 4090 costs less.
Can I run a 70B model on RTX 5090s?
In 4-bit a 70B model needs about 46 GB by our rule of thumb: two RTX 5090 hold it (about 59 GB usable), two RTX 4090 do not (about 44 GB usable), four RTX 4090 do. GeForce servers come with 1, 2 or 4 cards.
How many cards per server?
Both come in servers of 1, 2 or 4 cards, connected over PCIe (no NVLink on GeForce).
Rent the RTX 5090 or the RTX 4090.
Dedicated servers with 1 to 4 GPUs, one monthly price, paid in crypto. Online in under 10 minutes, no KYC.