On this page
The NVIDIA L40S has twice the memory of the L4, 48 gigabytes against 24, and moves data nearly three times as fast. On 11 November 2026, one L4 rented for $0.894 an hour in Montreal. The bigger card starts at $1.861 an hour, and only in American regions. AWS, Azure and Google list no L40S in Canada; if your model fits in 24 GB, pick the L4.
If your model fits on the smaller card and the data must stay in Canada, pick the L4.
If you need 48 gigabytes on one card and American hosting is acceptable, pick the L40S.
If you need more memory inside the country, rent four or eight L4s in one machine and accept the cost of splitting the model.
- GPU memory is the memory on the graphics card itself, where the model's weights and the text it is working on have to fit.
- L4: 24 gigabytes of memory, read at 300 gigabytes a second, 72 watts, per NVIDIA's product page.
- L40S: 48 gigabytes, read at 864 gigabytes a second, 350 watts, per NVIDIA's product page.
- Where to rent: the smaller card rents in Montreal and Toronto; the bigger one does not, as of 11 November 2026.
What's the difference between the L4 and the L40S?
The L40S is the bigger, faster, hungrier card: twice the memory, 2.9 times the memory bandwidth, and close to five times the power draw. Both run 8-bit math natively. Neither has the fast bridge some bigger cards use to talk to each other.
| NVIDIA L4 | NVIDIA L40S | |
|---|---|---|
| GPU memory | 24 GB | 48 GB GDDR6 |
| Memory bandwidth | 300 GB/s | 864 GB/s |
| Max power | 72 W | 350 W |
| FP8 compute (dense / with sparsity) | about 242 / 485 TFLOPS | 733 / 1,466 TFLOPS |
| Size | Single slot, low profile | Dual slot |
| Link between cards | PCIe Gen4 x16, 64 GB/s | PCIe, no NVLink |
All figures are from NVIDIA's product pages, read on 11 November 2026. For the smaller card, NVIDIA quotes the 8-bit figure with sparsity and says the dense number is half of it.
Memory bandwidth matters more than the compute figure for chat-style work. To write each new word, the card reads the model's weights out of memory again. A card that moves 864 gigabytes a second can do that faster than one that moves 300. We have not measured either card's speed on a fixed task, so this post gives no tokens-per-second figure.
The power gap is the smaller card's whole case. At 72 watts it fits in ordinary servers without extra cooling, which is also why cloud providers stock it widely.
Which models fit on an L4 and which need an L40S?
A model at 8-bit takes about one byte per parameter, and at 4-bit about half a byte. So the smaller card holds a 4-bit model of about 30 billion parameters with little room left, and the bigger card holds the same model at 8-bit.
The table uses the weight file sizes published on each model's Hugging Face page, read on 11 November 2026. Where no official file exists at that precision, the row says so and gives the parameter count times bytes instead.
| Model (parameters) | Weights at 8-bit | Weights at 4-bit | One L4 (24 GB) | One L40S (48 GB) |
|---|---|---|---|---|
| yev1-4b (4.2B) | about 4.2 GB (arithmetic) | about 2.1 GB (arithmetic) | Yes, either way | Yes, either way |
| Qwen3-8B (8.2B) | 9.4 GB (official FP8 file) | about 4.1 GB (arithmetic) | Yes, at 8-bit | Yes, even at 16-bit (16.4 GB) |
| pplx-decider v1.2 (26.1B) | about 26 GB (arithmetic) | about 13 GB (arithmetic) | Only at 4-bit | Yes, at 8-bit |
| Qwen3-32B (32.8B) | 34.3 GB (official FP8 file) | 19.3 GB (official AWQ file) | Only at 4-bit, tight | Yes, at 8-bit |
| Llama 3.3 70B (70.6B) | about 70.6 GB (arithmetic) | about 35 GB (arithmetic) | No | Only at 4-bit |
Weights are not the whole bill. The card also holds the text in progress (the context cache), which grows with longer documents and with every request served at once, and the serving software reserves some memory of its own. Qwen3-32B at 4-bit leaves under 5 gigabytes on the smaller card for all of that, which suits short prompts and one user at a time. The same model at 8-bit on the bigger card leaves nearly 14.
For the full method, with much larger open models, see how much GPU memory you need to run open models.
Small models are where the cheap card shines. Our own decision model, Yev, has 4.2 billion parameters, and we trained it on one of the bigger cards in 4.71 hours. Serving a model that size needs a fraction of either card. The open-source decision models we tested run from 2 to 27 billion parameters, and most fit on the smaller card at 8-bit.
How much do the L4 and L40S cost per hour?
One L4 costs $0.894 an hour in Montreal and $0.805 in Virginia. The bigger card costs $1.861 an hour in Virginia and has no Canadian price. These are on-demand Linux prices from AWS's price list, read on 11 November 2026.
| AWS instance | GPUs | GPU memory | Montreal (ca-central-1) | Virginia (us-east-1) |
|---|---|---|---|---|
| g6.xlarge | 1x L4 | 24 GB | $0.894/h | $0.805/h |
| g6.12xlarge | 4x L4 | 96 GB | $5.109/h | $4.602/h |
| g6.48xlarge | 8x L4 | 192 GB | $14.823/h | $13.350/h |
| g6e.xlarge | 1x L40S | 48 GB | Not offered | $1.861/h |
| g6e.12xlarge | 4x L40S | 192 GB | Not offered | $10.493/h |
Per gigabyte of GPU memory, the two cards cost about the same in Virginia: $0.034 an hour for the smaller card and $0.039 for the bigger one. The bigger card's extra price buys bandwidth, plus room for a 48-gigabyte model on one card.
Montreal runs about 11% above Virginia for the same instance. AWS's price list showed no instances of either card in its Calgary region (ca-west-1). Our scale-to-zero cost comparison works through what a g6e.12xlarge costs over a month, including the one-year commitment discount.
Can you rent an L40S in Canada?
Not from the big three clouds. As of 11 November 2026, AWS offers the card (its g6e family) in American regions but not in Montreal. Azure's Canada Central lists no such sizes, and Google's GPU locations page lists the card in no region at all. The one Canada-hosted L40S we could confirm is at ISAIC, a University of Alberta initiative in Edmonton, whose pricing page lists a dedicated one-L40S virtual machine at $858 CAD a month, read 11 November 2026.
Here is what the three do offer in Canada:
| Provider | Canadian region | L4 | L40S |
|---|---|---|---|
| AWS | Montreal (ca-central-1) | Yes, g6 family | No |
| Azure | Canada Central | No | No |
| Google Cloud | Montreal and Toronto | Yes, G2 family | No |
On every one of them, GPU quota for a new account starts at zero. You request it before you can launch anything.
Three routes give you more memory with inference kept in the country:
- Several L4s in one machine. A g6.12xlarge puts four L4s and 96 gigabytes in one Montreal machine for $5.109 an hour.
- Eight L4s. A g6.48xlarge doubles that to 192 gigabytes for $14.823 an hour.
- A bigger card that is rentable in Canada. AWS Montreal and Azure Canada Central both offer A100 and H100 machines, though in our experience the H100 usually needs a one- to three-year commitment.
- Where to compare them. Our guide to GPUs you can rent in Canada lists them with prices, and L40S vs A100 compares the closest step up.
- Your own L40S. It is a standard dual-slot card, so it can sit in a server in a Canadian office or data centre. You then own the power, cooling and repairs.
We did not find a Canadian GPU provider whose own pages list this card, so we name none here.
Splitting a model across several L4s costs speed. The cards swap partial results after every layer of the model, and that traffic crosses the server's ordinary expansion slots at 64 gigabytes a second, far slower than each card's own memory. Four of them hold as much as two of the bigger cards, but a model spread over four cards loses time to that traffic. Measure your own workload before you commit to the larger box.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account or on hardware they own, sizes the GPUs to the model, and builds the workflow on top. See how we deploy, the Canadian residency guide, and where each hosting model puts the data. If you need a model running with inference kept in Canada, bring one job to a free 30-minute call.
What this comparison does not cover
We have not run the same task on both cards, so there are no speed or cost-per-job numbers here. The model sizes count weights only, and the context cache will take more on long documents or busy servers. Prices are on-demand and move often. Savings plans and spot capacity cost less. We could not query Calgary directly, because the region is opt-in and not enabled on our account.
FAQ
Is the L40S better than the L4?
Yes, on every spec: double the memory, almost triple the memory speed and much more 8-bit compute. It also draws close to five times the power. On AWS on 11 November 2026 it cost 2.3 times as much per hour. If your model fits on the smaller card, the L4 is often the better buy.
Can an L4 run a 70-billion-parameter model?
Not on one card. A 70-billion-parameter model such as Meta's Llama needs about 35 gigabytes for its weights, even squeezed to half a byte per parameter. Four L4s in an AWS g6.12xlarge give 96 gigabytes. That holds the 8-bit weights, about 70.6 gigabytes, with some room for context. The cards pass data to each other while they work, so expect slower answers than one larger card would give.
Is the L4 available in Canada?
Yes. On 11 November 2026, AWS listed one L4 in its Montreal region from $0.894 an hour. Google Cloud listed it in Montreal and Toronto too. Azure's Canada Central region had none. On all three, new accounts start with zero GPU quota, so request it early.
Why is the L4 so cheap?
NVIDIA built it as a small, low-power card: 72 watts, single slot, low profile, built for running models rather than training large ones. That makes it easy for cloud providers to stock in ordinary servers. Per gigabyte of GPU memory it costs about the same as the bigger card on AWS. The bigger card's premium buys almost three times the memory speed.
Sources
- NVIDIA, L4 Tensor Core GPU, specifications, read 11 November 2026
- NVIDIA, L40S GPU, specifications, read 11 November 2026
- AWS, Amazon EC2 G6 instances and G6e instances; prices from the EC2 on-demand price list, read 11 November 2026
- Google Cloud, GPU regions and zones, read 11 November 2026
- Microsoft, Azure GPU-accelerated VM sizes, with the Canada Central size list read 11 November 2026
- Hugging Face model pages for Qwen3-8B-FP8, Qwen3-32B-FP8, Qwen3-32B-AWQ, Llama 3.3 70B Instruct, pplx-decider v1.2 and yev1-4b, read 11 November 2026
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project