L40 vs A40
Both run on the same nodes, the same templates and the same tiers here — so the choice comes down to memory, bandwidth and price. Every figure below is from the catalogue or the vendor datasheet.
- L40
- $0.09/h48 GB · Ada Lovelace · 32 available
- A40
- $0.06/h48 GB · Ampere · 32 available
- Price gap
- 33%On spot, per GPU-hour
- Over 100 h
- $3.00Difference for the same GPU-hours
What actually separates them.
Four differences that change a decision, computed from the numbers rather than asserted.
The A40 costs 33% less per hour
$0.06 against $0.09 on spot — $3.00 of difference over 100 GPU-hours.
Token generation is faster on the L40
864 GB/s against 696 GB/s. Inference reads the whole model from memory for every token, so bandwidth, not FLOPS, usually sets the ceiling.
More memory per dollar on the A40
800 GB per dollar per hour against 533 GB. The usual tie-breaker when both cards fit the job.
Every figure we hold on both cards.
A row is marked only when both values exist and one is genuinely better. L40 leads on 5 rows, A40 on 3.
| L40 | A40 | |
|---|---|---|
| GPU memoryDecides which models fit at all | 48 GB GDDR6 | 48 GB GDDR6 |
| Memory bandwidthSets token throughput on inference | 864 GB/s | 696 GB/s |
| FP16 tensor | 181 TFLOPS | 150 TFLOPS |
| FP32 | 91 TFLOPS | 37 TFLOPS |
| 8-bit float (FP8)Halves the memory a model needs, where the stack supports it | Supported | Not supported |
| Architecture | Ada Lovelace (2022) | Ampere (2020) |
| Board powerLower is cheaper to run at scale — for us, not for your bill | 300 W | 300 W |
| GPU-to-GPU linkOnly matters for multi-GPU training | PCIe only | PCIe only |
| Node sizes | 1×, 2×, 4×, 8× | 1×, 2×, 4×, 8× |
| Spot price | $0.09/h | $0.06/h |
| On-demand price | $0.19/h | $0.09/h |
| Memory per dollarGPU memory per $1 of spot time, per hour | 533 GB | 800 GB |
| 30-day reclaim rateShare of spot instances reclaimed | <5% | <5% |
| Available now | 32 GPUs | 32 GPUs |
Specifications come from each vendor's datasheet — linked on the L40 page and the A40 page. Prices are ours, per GPU-hour, billed per minute.
The same models, on one GPU of each.
Weights only, with 20% of the memory kept free for activations and the KV cache. A cross means one GPU is not enough — the model page has the multi-GPU arithmetic.
| Model | Precision | L40 48 GB | A40 48 GB |
|---|---|---|---|
| Llama 3.1 8B8B parameters | FP16 · 16 GB | Fits | Fits |
| INT4 · 4 GB | Fits | Fits | |
| Qwen2.5 32B32B parameters | FP16 · 64 GB | Needs 2× node | Needs 2× node |
| INT4 · 16 GB | Fits | Fits | |
| Llama 3.3 70B70B parameters | FP16 · 140 GB | Needs 4× node | Needs 4× node |
| INT4 · 35 GB | Fits | Fits | |
| Mixtral 8x22B141B parameters | FP16 · 282 GB | Needs 8× node | Needs 8× node |
| INT4 · 71 GB | Needs 2× node | Needs 2× node |
The same 100 GPU-hours on each.
Compute only, billed per minute. Storage is $0.08 per GB-month on either card.
| L40 | A40 | |
|---|---|---|
| One hour, spot | $0.09 | $0.06 |
| One day, spot | $2.16 | $1.44 |
| 100 GPU-hours, spot | $9.00 | $6.00 |
| 100 GPU-hours, on-demand | $19.00 | $9.00 |
| A month, reserved | $65.70 | $43.80 |
Take the L40 when…
- Fine-tuning with LoRA and QLoRA — Comfortable for parameter-efficient training on 13B–34B models without sharding.
- Inference servers — Holds 30B-class models quantised, or several smaller models side by side.
- Anything that can checkpoint — On the spot tier this model is 53% below its own on-demand price, with a 2-minute notice before a reclaim.
Take the A40 when…
- Fine-tuning with LoRA and QLoRA — Comfortable for parameter-efficient training on 13B–34B models without sharding.
- Inference servers — Holds 30B-class models quantised, or several smaller models side by side.
- Anything that can checkpoint — On the spot tier this model is 33% below its own on-demand price, with a 2-minute notice before a reclaim.
L40 or A40?
Which is cheaper, the L40 or the A40?
The A40, at $0.06 per GPU-hour on spot against $0.09 — 33% less. On-demand: $0.09 against $0.19.
Which one should I take for inference?
Take the one that holds the model with room for the KV cache, then the one with more memory bandwidth. Here that is the L40 on capacity (48 GB) and the L40 on bandwidth.
Can I run both on the same job?
Not inside one instance — a node holds GPUs of a single model. You can run two instances in parallel, on the same persistent volume in turn, or split the work: a big card for the model that needs capacity, a cheap one for the parts that do not.
Other comparisons: the full list · all 47 models
Try both for less than the price of deciding.
An hour on each costs $0.15 on spot. Nothing renews, and your disk moves between them.
Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.