RTX 5090
Our complete AI-workload review of the RTX 5090: VRAM analysis, Flux & ComfyUI throughput, local LLM performance, power efficiency, and workstation fit.
Overview
The RTX 5090 delivers 100/100 on our composite AI workload index — a balanced view of Flux generation throughput, SDXL inference, and quantized local LLM performance. With 32GB of VRAM and a 575W TDP, it's positioned for serious AI creators running large workflows locally.
Pros & Cons
Pros
- 32GB VRAM handles most Flux & SDXL workflows
- 4.8 it/s on Flux.1 dev FP16
- Excellent CUDA ecosystem support
- Strong resale value
Cons
- 575W TDP requires serious PSU
- Premium pricing
- Limited stock at MSRP
Performance benchmarks
| Flux.1 dev FP16 · 1024² · 25 steps | 4.8 it/s | |
| SDXL Base · 1024² · 20 steps | 18.2 it/s | |
| Llama 3 70B · Q4_K_M · 2k ctx | 42 tok/s | |
| Hunyuan Video · 720p · 5s | 1.44 it/s |
VRAM analysis
With 32GB of VRAM, the RTX 5090 can comfortably run: Flux.1 dev FP16 with all loaders resident, SDXL with multiple LoRAs, and Llama 3 70B at Q4 quantization.
Flux performance
At 4.8 it/s, a standard Flux.1 dev 25-step generation completes in ~5.2 seconds. For batch workflows, expect linear scaling up to VRAM limits.
ComfyUI performance
SDXL throughput of 18.2 it/s makes the RTX 5090 ideal for iterative ComfyUI workflows. Block-swap and tiled VAE are rarely needed.