The Open-Source AI Stack
RSS
All hardware

Hardware · Workstation

NVIDIA GeForce RTX 5090

NVIDIA

Blackwell consumer flagship. Bandwidth tier matches the workstation Blackwell parts; capacity is the constraint for large models. FP8 is omitted because NVIDIA publishes only a sparse AI-TOPS headline for it.

Compute units 209.5 TF FP16 1.79 TB/s memory bus 32 GB GDDR7 (VRAM)
Workstation. Bus width tracks bandwidth (1.79 TB/s, sets decode speed); the box tracks capacity (32 GB, sets what fits).

Specs

Memory
32 GB GDDR7
Bandwidth
1.79 TB/s
FP16 dense
209.5 TFLOPS
Power
575 W
Form factor
pcie
Interconnect
pcie
Released
2025-01

What it runs (single unit, Q4_K_M, 4K context)

Model Fits? Decode ceiling
Llama 3.1 8B Instruct yes ~356 tok/s
Llama 3.3 70B Instruct no —
Qwen 2.5 72B Instruct no —
DeepSeek-V3 no —

Ceiling is the theoretical rooflineruntimeA performance model that bounds throughput by either compute or memory bandwidth, whichever is the limiting resource for an operation's arithmetic intensity. Open full entry ; open the explorer to set quant, context, and runtime and see the realistic range.

Sources