L40S Specifications
Strong FP8 inference and media engines, but GDDR6 bandwidth and no NVLink make it a poor multi-GPU training part.
Buy it for
Inference and rendering where FP8 and media engines matter more than interconnect. Strong throughput per dollar.
Look elsewhere if
You need multi-GPU training — there is no NVLink, so scaling runs over PCIe.
Will it fit your system?
Passively cooled — needs server airflow
This card has no fans. It relies on a chassis forcing air through its heatsink. In a normal desktop it will throttle within minutes and can overheat outright. It needs a rack server with high-static-pressure fans, or an aftermarket shroud and blower.
- Form factor
- PCIe dual-slot
- Cooling
- passive (no fans)
- Power connectors
- 1x 16-pin CEM5 (12VHPWR)
- Board power
- 350 W
- Suggested PSU
- 850 W or better
- Interface
- PCIe Gen 4
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 L40S does and does not support in hardware.
Blackwell-generation feature — B200 has it
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.
Precision must be managed manually
Multi-GPU traffic goes over PCIe — a real limit for distributed training
Hardware ray tracing for rendering workloads
Corrects single-bit errors — expected for long training runs
What fits in memory
Largest model that fits on one L40S, assuming ~20% of the 48 GB is reserved for KV cache and activations.
| Model | FP16 / BF16 | FP8 |
|---|---|---|
| Llama 8B | Fits | Fits |
| Llama 70B | 4× GPUs | 2× GPUs |
| Mixtral 8x22B | 8× GPUs | 4× GPUs |
| Llama 405B | 22× GPUs | 11× 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
Compute outruns bandwidth. Strong on compute-bound phases such as prefill and training, but batch-1 decode will leave FLOPS idle waiting on memory.
Cost per unit of compute
Breakeven counts hardware cost against rental only. Power, cooling, hosting and depreciation push the real crossover substantially later — see colocation pricing for what hosting actually costs.
Cost to run
Electricity for one card at 100% duty, including cooling and PSU overhead (4,292 kWh/year).
At the US average rate, three years of continuous operation costs $1,545 — roughly 17% of the card's $9,000 launch price, before any cooling or hosting fees.
Availability & market
Readily available and much easier to buy than Hopper. A common choice for self-hosted inference.
Memory
- Capacity
- 48 GB
- Type
- GDDR6 ECC
- Bandwidth
- 864 GB/s
- Bus width
- 384-bit
- GB per TFLOPMemory relative to compute — higher favours large models
- 0.133
Compute
- FP64
- —
- FP32
- 91.6 TFLOPS
- TF32 dense
- 183 TFLOPS
- FP16 dense
- 362 TFLOPS
- FP16 sparseRequires structured sparsity; rarely reached in practice
- 733 TFLOPS
- FP8 dense
- 733 TFLOPS
- INT8 dense
- 733 TOPS
Silicon
- Codename
- AD102
- Process node
- TSMC 4N
- Transistors
- 76.3B
- Die size
- 609 mm²
- CUDA cores
- 18,176
- Tensor cores
- 568
- SM count
- 142
- Boost clock
- 2,520 MHz
Platform & power
- TDP
- 350 W
- TFLOPS per wattDense FP16 ÷ TDP — drives your power bill at scale
- 1.03
- Form factor
- PCIe dual-slot
- Cooling
- passive
- Interconnect
- None
- Interconnect speed
- —
- PCIe generation
- Gen 4
- MIG
- No
- Launch
- 2023-08
- MSRP at launch
- $9,000
L40S — 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 L40S have?
- The L40S has 48 GB of GDDR6 ECC memory with 864 GB/s of bandwidth. In practice plan on roughly 38 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 L40S run?
- On a single L40S, about 19B parameters at FP16, or roughly 38B quantised to FP8. 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 L40S in a desktop PC?
- Only with added cooling. The L40S is a passively cooled card — it has no fans at all and depends on a rack chassis forcing air through its heatsink. Dropped into an ordinary desktop it will throttle within minutes and can overheat outright. It needs either a rack server with high-static-pressure fans or an aftermarket shroud and blower.
- How much power does the L40S use?
- The L40S is rated at 350 W board power. It is fed through 1x 16-pin CEM5 (12VHPWR). For a single-card system, a 850 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 L40S?
- At the US commercial average of $0.12/kWh, running one L40S continuously costs about $515 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 17% of what the card cost at launch.
- Does the L40S support FP8?
- Yes. The L40S has hardware FP8 tensor support, so FP8 weights and activations run at full tensor-core rate — 733 dense TFLOPS. This is the format most current inference servers default to for large models.
- Does the L40S support FP4?
- No. The L40S predates hardware FP4. You can run 4-bit quantised weights on it and the memory saving is real, but they are dequantised before compute, so there is no arithmetic speedup.
- Does the L40S support INT8?
- Yes. The L40S supports INT8 acceleration at 733 dense TOPS. INT8 remains the most widely supported quantised inference format across frameworks.
- Does the L40S support BF16?
- Yes. The L40S 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 L40S have NVLink?
- No. The L40S has no NVLink. Multi-GPU work goes over PCIe Gen 4, which is an order of magnitude slower and makes tensor parallelism across cards expensive.
- Does the L40S have RT cores?
- Yes. The L40S includes dedicated ray-tracing cores, so it is usable for rendering and visualisation as well as compute.
- Does the L40S have ECC memory?
- Yes. The L40S has ECC, so single-bit memory errors are corrected rather than silently corrupting a long run.
- Is the L40S worth buying?
- Buy it for: Inference and rendering where FP8 and media engines matter more than interconnect. Strong throughput per dollar. Look elsewhere if: You need multi-GPU training — there is no NVLink, so scaling runs over PCIe.