GPU comparison
H100 vs L40S: Specs, monthly price and which to rent
The H100 is a training GPU: 80 GB of HBM3 at 3.35 TB/s, 1.98 PF of dense FP8 and NVLink between the GPUs of an 8-GPU server. The L40S is a PCIe inference and graphics card with 48 GB of GDDR6 at 864 GB/s and 0.73 PF of FP8, for less than half the monthly price. Pick the H100 for training, fine-tuning and large models; the L40S for 7B to 13B inference, image generation and rendering on a budget.
- Updated Nov 4, 2026
- Specs: vendor datasheets, dense Tensor figures
- Prices: live CoreGPU catalog
| Spec | NVIDIA H100 SXM | NVIDIA L40S |
|---|---|---|
| Architecture | Hopper | Ada Lovelace |
| GPU memory | 80 GB HBM3 | 48 GB GDDR6 |
| Memory bandwidth | 3.35 TB/s | 864 GB/s |
| FP8 dense | 1.98 PF | 0.73 PF |
| FP4 dense | n/a | n/a |
| Interconnect | NVLink 4 · 900 GB/s | PCIe 4.0 x16 |
| Power (TDP) | 700 W | 350 W |
| Price per GPU / month | $1,999 | $849 |
| Price per GB of memory | $24.99 | $17.69 |
| Price per FP8 PFLOPS | $1,010 | $1,163 |
| Availability | Only 3 left | 38 available |
Green marks the better value on each line. Prices are per GPU per month at CoreGPU, at least 25% below the market median.
- Training and fine-tuning across 8 NVLink GPUs
- Memory-bound inference of 30B to 70B models
- Multi-GPU jobs that need fast GPU-to-GPU links
- 7B to 13B model inference at the lowest price
- Image and video generation
- 3D rendering and VFX with RT cores
H100 vs L40S: monthly price by server size
| Server | H100 | L40S | Difference |
|---|---|---|---|
| 1 GPU | $1,999/mo | $849/mo | $1,150 less on the L40S |
| 2 GPUs | $3,998/mo | $1,698/mo | $2,300 less on the L40S |
| 4 GPUs | $7,996/mo | $3,396/mo | $4,600 less on the L40S |
| 8 GPUs | $15,992/mo | $6,792/mo | $9,200 less on the L40S |
Flat monthly price per GPU, the same at every server size. USD, taxes excluded.
Largest model that fits in GPU memory
| Weights | 1× H100 | 1× L40S | 8× H100 | 8× L40S |
|---|---|---|---|---|
| BF16 | 33B | 20B | 266B | 160B |
| FP8 | 66B | 40B | 533B | 320B |
| FP4 | 133B | 80B | 1.1T | 640B |
Parameters whose weights fit with 20% headroom for activations and runtime. Long contexts and large batches need more room for the KV cache: see how much GPU memory you need.
More GPU comparisons
H100 vs L40S: common questions
Already a customer? Open a ticket from the console: engineers reply within one business day.
How support worksIs the H100 faster than the L40S?
On dense FP8 Tensor throughput the H100 is faster: 1.98 PF against 0.73 PF per GPU. Memory bandwidth is 3.35 TB/s on the H100 and 864 GB/s on the L40S, which matters most for LLM decoding.
How much does the H100 cost per month compared with the L40S?
At CoreGPU the H100 is $1,999 and the L40S is $849 per GPU per month, so an 8-GPU server costs $15,992 or $6,792 a month. The L40S is 58% cheaper per GPU. Both are at least 25% below the market median, billed monthly from a prepaid balance.
Which is better for LLM inference, the H100 or the L40S?
For large models and long contexts the H100 has the edge with 80 GB of memory per GPU, so a model needs fewer GPUs. Choose the H100 for training and fine-tuning across 8 NVLink GPUs; the L40S for 7B to 13B model inference at the lowest price.
Can I rent the H100 and the L40S by the month?
Yes. Both are dedicated GPUs you rent by the month, from 1 to 8 per server, with root SSH access. Right now the H100 is only 3 left and the L40S is 38 available. You pay in BTC, ETH, XMR, USDT (TRC-20) or SOL.