RTX 5090 Specifications
Consumer Blackwell. 1.8 TB/s GDDR7 and FP4 support, but 32 GB and no NVLink keep it single-GPU work.
Buy it for
The fastest single desktop card for local inference: 32 GB of GDDR7 at around 1.8 TB/s, which is HBM-class bandwidth in a consumer part, plus FP4 support for quantised serving.
Look elsewhere if
You need ECC, multi-GPU scaling, or a datacentre licence. There is no NVLink, and NVIDIA's driver terms restrict GeForce in datacentre deployments.
Will it fit your system?
- Form factor
- PCIe dual-slot
- Cooling
- active
- Power connectors
- 1 × 16-pin (12V-2×6)
- Board power
- 575 W
- Suggested PSU
- 1200 W or better
- Interface
- PCIe Gen 5
PSU figure assumes one card plus a mainstream CPU and drives, with headroom for transient spikes — modern cards can draw well above their rated TDP for milliseconds, which is what trips an undersized supply.
Supported formats & features
What the RTX 5090 does and does not support in hardware.
Hardware FP4 — roughly doubles inference throughput over FP8
E4M3 and E5M2 — the standard format for modern inference
Wider exponent than FP16 — fewer overflow issues when training
Drop-in FP32 speedup with no code changes
Quantised inference supported
Not yet confirmed against vendor documentation.
Per-layer automatic FP8 precision management
Multi-GPU traffic goes over PCIe — a real limit for distributed training
Hardware ray tracing for rendering workloads
No ECC, so silent corruption is possible on multi-day runs
What fits in memory
Largest model that fits on one RTX 5090, assuming ~20% of the 32 GB is reserved for KV cache and activations.
| Model | FP16 / BF16 | FP8 | INT4 / FP4 |
|---|---|---|---|
| Llama 8B | Fits | Fits | Fits |
| Llama 70B | 6× GPUs | 3× GPUs | 2× GPUs |
| Mixtral 8x22B | 12× GPUs | 6× GPUs | 3× GPUs |
| Llama 405B | 32× GPUs | 16× GPUs | 8× GPUs |
Weights only. Long contexts, large batch sizes and optimiser state during training all need substantially more — full fine-tuning typically needs 3–4× the weight memory.
Workload character
Plenty of bandwidth per unit of compute. Memory-bound work — single-stream token decode above all — runs close to its ceiling, and compute is usually what runs out first.
Cost to run
Electricity for one card at 100% duty, including cooling and PSU overhead (7,052 kWh/year).
At the US average rate, three years of continuous operation costs $2,539 — roughly 127% of the card's $1,999 launch price, before any cooling or hosting fees.
Where it sits in the line
Availability & market
Supply is tight and street prices sit above MSRP. Use a native 12V-2×6 cable rather than an adapter — melted connectors on this generation trace back to adapters and partial seating.
Memory
- Capacity
- 32 GB
- Type
- GDDR7
- Bandwidth
- 1,792 GB/s
- Bus width
- 512-bit
- GB per TFLOPMemory relative to compute — higher favours large models
- 0.153
Compute
- FP64
- —
- FP32
- 104.8 TFLOPS
- TF32 dense
- 209.5 TFLOPS
- FP16 dense
- 209.5 TFLOPS
- FP16 sparseRequires structured sparsity; rarely reached in practice
- 419 TFLOPS
- FP8 dense
- 838 TFLOPS
- INT8 dense
- 838 TOPS
Silicon
- Codename
- GB202
- Process node
- TSMC 4NP
- Transistors
- 92.2B
- Die size
- 750 mm²
- CUDA cores
- 21,760
- Tensor cores
- 680
- SM count
- 170
- Boost clock
- 2,410 MHz
Platform & power
- TDP
- 575 W
- TFLOPS per wattDense FP16 ÷ TDP — drives your power bill at scale
- 0.36
- Form factor
- PCIe dual-slot
- Cooling
- active
- Interconnect
- None
- Interconnect speed
- —
- PCIe generation
- Gen 5
- MIG
- No
- Launch
- 2025-01
- MSRP at launch
- $1,999
RTX 5090 — common questions
Answered from the specifications above. Questions this card's data cannot answer with confidence are left out rather than guessed.
- How much VRAM does the RTX 5090 have?
- The RTX 5090 has 32 GB of GDDR7 memory with 1,792 GB/s of bandwidth. In practice plan on roughly 25 GB for weights: KV cache, activations and fragmentation take the rest, and long contexts or large batches take considerably more.
- What size model can the RTX 5090 run?
- On a single RTX 5090, about 12B parameters at FP16, or roughly 51B quantised to FP4. Anything larger has to be split across several GPUs, which is slow on this card because it has no NVLink. These are weight-only figures — a long context window can consume tens of gigabytes on its own.
- Can you put the RTX 5090 in a desktop PC?
- Yes. The RTX 5090 is a self-cooled PCIe dual-slot card and works in a normal desktop, provided the case has clearance and the power supply is rated around 1200 W or better.
- How much power does the RTX 5090 use?
- The RTX 5090 is rated at 575 W board power. It is fed through 1 × 16-pin (12V-2×6). For a single-card system, a 1200 W supply or better leaves headroom for the CPU, drives and the transient spikes that trip undersized units.
- How much does it cost to run the RTX 5090?
- At the US commercial average of $0.12/kWh, running one RTX 5090 continuously costs about $846 a year in electricity. That figure includes cooling and power-supply overhead at a factor of 1.4 over board power, and over three years comes to roughly 127% of what the card cost at launch.
- Does the RTX 5090 support FP8?
- Yes. The RTX 5090 has hardware FP8 tensor support, so FP8 weights and activations run at full tensor-core rate — 838 dense TFLOPS. This is the format most current inference servers default to for large models.
- Does the RTX 5090 support FP4?
- Yes. The RTX 5090 supports 4-bit floating point in hardware, which roughly doubles throughput again over FP8 and halves the memory a model occupies.
- Does the RTX 5090 support INT8?
- Yes. The RTX 5090 supports INT8 acceleration at 838 dense TOPS. INT8 remains the most widely supported quantised inference format across frameworks.
- Does the RTX 5090 support BF16?
- Yes. The RTX 5090 supports bfloat16. Its wider exponent range makes training far less prone to the overflow that plagues FP16, which is why most current training recipes assume it.
- Does the RTX 5090 have NVLink?
- No. The RTX 5090 has no NVLink. Multi-GPU work goes over PCIe Gen 5, which is an order of magnitude slower and makes tensor parallelism across cards expensive.
- Does the RTX 5090 have RT cores?
- Yes. The RTX 5090 includes dedicated ray-tracing cores, so it is usable for rendering and visualisation as well as compute.
- Does the RTX 5090 have ECC memory?
- No. The RTX 5090 has no ECC. For short interactive work this rarely matters; for multi-week training runs it is a real risk, and it is one reason datacentre parts cost more than consumer ones.
- Is the RTX 5090 worth buying?
- Buy it for: The fastest single desktop card for local inference: 32 GB of GDDR7 at around 1.8 TB/s, which is HBM-class bandwidth in a consumer part, plus FP4 support for quantised serving. Look elsewhere if: You need ECC, multi-GPU scaling, or a datacentre licence. There is no NVLink, and NVIDIA's driver terms restrict GeForce in datacentre deployments.