NVIDIA · Blackwell

Rent RTX 5090

NVIDIA GeForce RTX 5090, 32 GB.LLM inferenceRenderinggeneral ml

Full price from

—

in stock, per GPU-hour

Spot from

—

in stock, per GPU-hour

Offers in stock

0 of 1

full price or spot

Providers in stock

0 of 1

full price or spot

Offers (1)

NosanaSold outSelf-service

1× RTX 509032 GB VRAM

Nosana network, hosts worldwide

Details

Stock and price checked 8 min ago.

You start it yourself once there is stock.

nvidia-5090

Full price

€0.45

per hour, whole instance, 1 GPU

View offer

Price and availability history

Price per GPU-hour

Excluding VAT.

Providers with stock

Each provider counts once. Full price and spot are separate.

Share of offers in stock

Full price. Offers we can't say are really sold there are left out until they've been in stock once.

Units free

Only providers that report a number. Each line is one provider and one server size, because their counts cover different things and can't be added up. Providers that only say yes or no are under Stock by provider.

Stock by provider

In stock, yes or no, for every provider selling this GPU. A filled band means in stock: green at full price, orange as spot.

NosanaSold out

Reports a number

Using several together

Can't be linked. This GPU has no NVLink, so several in one server only talk over PCIe. Each card keeps its own 32 GB. That suits running a separate job on each card. Splitting one model across them works, but slowly.

Specifications

Consumer card: memory bandwidth and FLOPS figures come from NVIDIA's RTX Blackwell whitepaper (Table 3), not a product datasheet. FP4 (FP32 accumulate), FP8 (FP16 accumulate), FP16 (FP16 accumulate), BF16 (FP32 accumulate) and INT8 are with sparsity. fp64Tflops is derived: NVIDIA states GB202 FP64 is 1/64th of FP32 (104.8 / 64). No NVLink/SLI.

Memory

VRAM: 32 GB
Memory type: GDDR7
Memory bandwidth How fast the GPU reaches its memory. Often limits LLM speed: 1,792 GB/s
Bus width: 512 bit

Compute

FP64 Speed at 64-bit precision, used for scientific simulation: 1.64 TFLOPS
FP32 Speed at 32-bit precision: 104.8 TFLOPS
FP16 Speed at 16-bit precision: 838 TFLOPS
BF16 16-bit AI format. Standard for training and full-quality inference: 419 TFLOPS
FP8 8-bit AI format. About half the memory of BF16, slightly lower quality: 1,676 TFLOPS
FP4 4-bit AI format. About a quarter of the memory of BF16: 3,352 TFLOPS
INT8 8-bit integer speed, used by quantized models: 1,676 TOPS
FP8 support: Yes
FP4 support: Yes
Transformer Engine Hardware that speeds up FP8 and FP4: Yes
Compute cores: 21,760

Multi-GPU and connectivity

NVSwitch Lets every GPU in a server talk to every other at full speed: No
PCIe generation: 5

Sharing and security

MIG partitioning Splits one GPU into several isolated smaller GPUs: No
Max MIG instances: 0
Confidential computing: No

Hardware

Architecture: Blackwell
Launch year: 2025
Process node: TSMC 4N
Transistors: 92.2 billion
Die size: 750 mm²
Power (TDP) Maximum power draw: 575 W
Cooling: active

Video

NVENC engines: 3
NVDEC engines: 2
AV1 encode: Yes
AV1 decode: Yes

Sources (40)