The NVIDIA RTX 5090 is NVIDIA's flagship Blackwell consumer GPU, 32 GB of GDDR7 memory at 1.79 TB/s of bandwidth, 5th-generation Tensor Cores with native FP4 support, and 21,760 CUDA cores.
The RTX 5090 brings the GB202 Blackwell die, 92 billion transistors on TSMC's 4NP node, to a PCIe form factor, with 5th-gen Tensor Cores, native FP4, and GDDR7 memory.
Native FP4 precision delivers 3,352 AI TOPS: 2× the AI throughput of RTX 4090.
78% more bandwidth than RTX 4090. 32 GB fits 13B models at FP16 and 32B at Q4.
Drops into any PCIe Gen5 system, no SXM motherboard required.
4th-gen RT Cores and hardware AV1 encode make the RTX 5090 the best consumer GPU for AI video.
1.79 TB/s bandwidth makes the 5090 the fastest consumer GPU for token generation on 7B–13B models.
32 GB gives headroom for LoRA and QLoRA fine-tuning of 13B models without sharding.
DLSS 4, hardware AV1 encode, and 32 GB VRAM make this the best consumer GPU for FLUX and SDXL.
Full RTX 5090 reserved exclusively for you. Zero noisy-neighbour risk, 99.99% SLA.
Join waitlist →Reserved RTX 5090 at a flat monthly rate. Pricing confirmed at launch.
Launching soonScale across a full RTX 5090 cluster with InfiniBand interconnect.
Talk to clusters team →The RTX 5090 is NVIDIA's flagship Blackwell consumer GPU, 32 GB GDDR7 at 1.79 TB/s, 3,352 AI TOPS, 5th-gen Tensor Cores with FP4. The most powerful consumer GPU ever built.
Coming soon. Join the waitlist to be notified the moment capacity opens on packet.ai.
From $0.59/GPU-hour. Pricing to be confirmed at launch. See pricing →
13B at FP16 natively; 32B at Q4 on a single card. For 70B+, use H100 or H200.
H100 has more memory (80 GB vs 32 GB), ECC, NVLink, and 24/7 datacenter reliability. RTX 5090 wins on cost-per-token for sub-30B inference.
No. PCIe Gen5 only. For NVLink multi-GPU, use H100 SXM or B200.
The most powerful consumer GPU ever built. Join the waitlist for early access on packet.ai.
No commitment · we'll notify you by email