NVIDIA RTX 5090 (Blackwell Consumer Flagship)
Product Overview
The NVIDIA RTX 5090, unveiled at CES 2025-01, is the consumer flagship bringing the Blackwell architecture to consumer GPUs for the first time. With 32GB GDDR7 memory, 21,760 CUDA cores, and a 575W TDP, it delivers 3,352 TOPS of AI compute (FP4) — 2.5× that of the RTX 4090.
Positioned for local LLM inference (70B+ models), Stable Diffusion XL training, and consumer AI developers.
Core Specifications
| Parameter | Value |
|---|
| Architecture | Blackwell (GB202) |
| Process Node | TSMC 4N (custom 5nm) |
| CUDA Cores | 21,760 |
| Tensor Cores | 680 (5th Gen) |
| RT Cores | 170 (4th Gen) |
| Base Clock | 2.01 GHz |
| Boost Clock | 2.41 GHz |
| Memory | 32 GB GDDR7 |
| Memory Bandwidth | 1,792 GB/s (28 Gbps × 512-bit) |
| FP32 Compute | 104.8 TFLOPS |
| FP16 Tensor | 419 TFLOPS (sparse) |
| FP8 Tensor | 838 TFLOPS (sparse) |
| FP4 Tensor | 3,352 TOPS (sparse) |
| INT8 Tensor | 1,676 TOPS |
| TDP | 575 W |
| Power Connector | 1× 16-pin (12V-2x6) |
| MSRP | $1,999 |
| Launch Date | 2025-01-30 |
RTX 5090 vs RTX 4090 Comparison
| Metric | RTX 5090 | RTX 4090 | Improvement |
|---|
| Architecture | Blackwell | Ada Lovelace | New gen |
| CUDA Cores | 21,760 | 16,384 | 1.33× |
| Memory | 32GB GDDR7 | 24GB GDDR6X | 1.33× |
| Memory Bandwidth | 1,792 GB/s | 1,008 GB/s | 1.78× |
| FP16 Tensor | 419 TFLOPS | 165 TFLOPS | 2.5× |
| FP4 Tensor | 3,352 TOPS | N/A | New |
| TDP | 575W | 450W | 1.28× |
| Price | $1,999 | $1,599 | 1.25× |
Blackwell New Features
FP4 Precision Support
- Native FP4 Tensor Cores (first time on consumer GPUs).
- Reduces inference memory footprint by 50% (vs FP8).
- 70B LLM can run FP4 quantized within 32GB memory (~40GB model compressed).
DLSS 4 Multi Frame Generation
- Multi Frame Generation: Generates 3 frames from 1 (vs DLSS 3's 1 frame from 1).
- Gaming-only, but showcases Blackwell's compute power.
GDDR7 Memory
- 28 Gbps speed (vs GDDR6X 21 Gbps).
- 1,792 GB/s bandwidth = 2× RTX 4090.
- Alleviates the memory-bound bottleneck in LLM inference.
| Model | Quantization | RTX 5090 (32GB) | RTX 4090 (24GB) | Improvement |
|---|
| Llama 3 8B | FP16 | ~95 tok/s | ~70 tok/s | 1.36× |
| Llama 3 70B | FP4 | ~28 tok/s | OOM | Breakthrough |
| Llama 3 70B | INT4 | ~22 tok/s | ~15 tok/s | 1.47× |
| Mixtral 8x7B | INT4 | ~45 tok/s | ~32 tok/s | 1.41× |
| Qwen 2.5 72B | FP4 | ~26 tok/s | OOM | Breakthrough |
70B model FP4 quantized (~40GB) fully fits in VRAM — 32GB memory is the key enabler.
Use Cases
- ✅ Local 70B LLM inference (FP4 quantized, 32GB VRAM)
- ✅ Stable Diffusion XL / Flux training and inference
- ✅ Video production (DaVinci Resolve AI acceleration)
- ✅ 8K gaming + frame generation
- ❌ Data center (use H100/B200 instead)
- ❌ Multi-node training (lacks NVLink)