Comparison · HBM against GDDR, at similar money

# A100 PCIe or L40S?

Both are dedicated, both are monthly, both open without an identity check. What separates them is memory, bandwidth and $377 a month — and which of those matters depends entirely on what you intend to run.

- $1,091/mo A100 PCIe, one card
- $714/mo L40S, one card
- 80 vs 48 GB memory per card
- 2.2× bandwidth gap

## The short answer

Which one, and when

- **Pick the A100 PCIe:** when your model needs more than 48 GB on a single card. That is a threshold, not a preference — below it, the extra $377/month buys you nothing.
- **Pick the L40S:** when 48 GB holds your model. It saves $377/month, and memory you do not use is memory you paid for.
- **Neither, if:** your model does not fit on a single card of either. Sharding costs synchronisation at every layer — see the [interconnect guide](https://gpuserver.io/guides/nvlink-vs-pcie) before buying a multi-GPU node.

## Side by side

**A100 PCIe compared with L40S**

|  | [A100 PCIe](https://gpuserver.io/gpu/a100-80gb) | [L40S](https://gpuserver.io/gpu/l40s) | Difference |
|---|---|---|---|
| Architecture | Ampere | Ada Lovelace | Ampere is a different generation |
| Memory per card | 80 GB HBM2e | 48 GB GDDR6 ECC | 32 GB in favour of the A100 PCIe |
| Memory bandwidth Predicts generation speed | 1,935 GB/s | 864 GB/s | 2.24× in favour of the A100 PCIe |
| Form factor | PCIe add-in card | PCIe add-in card | Same |
| Node sizes | 1×, 2×, 4×, 8× | 1×, 2×, 4×, 8× | Up to 640 GB in one node |
| From, per month | $1,091 | $714 | $377/month apart |

## Price at every node size

**Monthly price of both cards at each available node size**

| Node | A100 PCIe | L40S | Total VRAM |  |
|---|---|---|---|---|
| 1 × GPU | $1,091/mo | $714/mo | 80 GB / 48 GB | [A100 PCIe](https://gpuserver.io/configure?c=a100-80-x1) [L40S](https://gpuserver.io/configure?c=l40s-x1) |
| 2 × GPU | $2,159/mo | $1,427/mo | 160 GB / 96 GB | [A100 PCIe](https://gpuserver.io/configure?c=a100-80-x2) [L40S](https://gpuserver.io/configure?c=l40s-x2) |
| 4 × GPU | $4,157/mo | $2,754/mo | 320 GB / 192 GB | [A100 PCIe](https://gpuserver.io/configure?c=a100-80-x4) [L40S](https://gpuserver.io/configure?c=l40s-x4) |
| 8 × GPU | $7,906/mo | $5,251/mo | 640 GB / 384 GB | [A100 PCIe](https://gpuserver.io/configure?c=a100-80-x8) [L40S](https://gpuserver.io/configure?c=l40s-x8) |

## What each one holds on a single card

At 8k context, best precision that fits. This is the difference that decides a purchase — not the specification sheet.

### Only the A100 PCIe

Nothing. Every model the A100 PCIe holds on one card, the L40S holds too — the difference between them is speed, not capability.

### Only the L40S

Nothing. The A100 PCIe holds everything the L40S does.

**The same model on both: Llama 3.3 70B.** On the A100 PCIe, roughly 25 tokens/second at 4-bit (AWQ, GPTQ); on the L40S, roughly 11. Single-stream and deliberately conservative — with continuous batching the aggregate is several times higher on both. Treat the *ratio* as the useful number, not the absolute.

## Cost per GB and per TB/s

Two ratios that cut through the specification sheet. The first tells you what memory costs; the second what speed costs.

**Cost per gigabyte of VRAM and per terabyte per second of bandwidth**

| Card | $/GB of VRAM, per month | $/TB per second, per month | Better at |
|---|---|---|---|
| [A100 PCIe](https://gpuserver.io/gpu/a100-80gb) | $13.64 | $564 | Both ratios |
| [L40S](https://gpuserver.io/gpu/l40s) | $14.88 | $714 | Neither |

These ratios rank cards; they do not choose one. A card that is cheaper per gigabyte is worthless if it has too few gigabytes to hold your model at all — capacity is a threshold, not a slope. Use them to break a tie between two cards that both fit.

## A100 PCIe against L40S, in short

Is the A100 PCIe worth the extra money over the L40S?

It costs $377 more a month — 53% — for 32 GB more memory and 124% more memory bandwidth. Both hold the same set of models on a single card, so the difference is speed rather than capability.

Which one is faster for LLM inference?

Generation speed is bounded by memory bandwidth, because each token requires re-reading the active weights. The A100 PCIe has 1,935 GB/s against 864 GB/s, so the A100 PCIe leads by roughly 124% on a model both of them hold.

Do both come as multi-GPU nodes?

A100 PCIe: 1×, 2×, 4×, 8×. L40S: 1×, 2×, 4×, 8×. Both are PCIe cards, joined peer-to-peer over the bus rather than by NVLink.

Can I rent either without an identity check?

Yes, both, on the same terms. No document, no selfie, no phone number and no company registration, at any level of spend. Payment is in cryptocurrency, each invoice gets its own address, and root access arrives under 5 minutes after the first confirmation.

## Other comparisons

- [A100 PCIe vs A100 PCIe  The same GPU with twice the memory Compare](https://gpuserver.io/compare/a100-80gb-vs-a100-40gb)
- [H100 PCIe vs A100 PCIe  The two 80 GB PCIe cards Compare](https://gpuserver.io/compare/h100-pcie-vs-a100-80gb)
- [L40S vs RTX A6000  Two 48 GB cards, one built for data centres Compare](https://gpuserver.io/compare/l40s-vs-rtx-a6000)

## Both are racked and waiting, in all 6 data centres.

Whichever you pick, it is monthly, paid in crypto, opened without an identity check and delivered in under 5 minutes.

[Open the configurator](https://gpuserver.io/configure) [Read the guides](https://gpuserver.io/guides)

---

Source: https://gpuserver.io/compare/a100-80gb-vs-l40s/. This file is generated from the same data as the website; if a figure here differs from a page, the page is authoritative and this file is stale — the canonical source is https://gpuserver.io/.