AMD MI250
vs
Nvidia L40
Two data center GPUs, one CDNA2 and one Ada Lovelace: 128 GB HBM2e against 48 GB GDDR6, 3.7x the memory bandwidth on the MI250, and Infinity Fabric 400 GB/s on the MI250 where the L40 is PCIe only.
At a glance
Memory
- MI250
- 128 GB
- L40
- 48 GB
2.7x more on MI250
Memory bandwidth
- MI250
- 3,200 GB/s
- L40
- 864 GB/s
3.7x higher on MI250
Cheapest on-demand
- MI250
- --
- L40
- $0.78 /GPU/hr
MI250 is not sold on demand
Price comparison
Median price per GPU per hour by billing type, with the cheapest listing under it.
| Billing | MI250 | L40 | Difference |
|---|---|---|---|
| On-demand | No on-demand listings | $1.05 /GPU/hr from $0.78 | |
| Reserved | -- /GPU/hr from $1.30 (12 mo) | -- /GPU/hr from $0.70 | |
| Spot | No spot listings | $0.79 /GPU/hr from $0.77 |
Heads up: A provider's own page may quote a different figure, for example on tax, region or a promotion. A monthly-only plan shows a derived hourly rate. Medians are across providers, one vote each, so a large catalog does not outweigh a small one. No median for reserved on the MI250 and reserved on the L40: we only show one when at least 3 providers list the GPU on that billing type. Verify before provisioning. More on how we price.
Price history
MI250 vs L40 price history
Weekly median on-demand price per GPU per hour
Specs comparison
| Spec | MI250 | L40 | Difference |
|---|---|---|---|
| Memory | |||
| Capacity | 128 GB HBM2e | 48 GB GDDR6 | 2.7x more on MI250 |
| Bandwidth | 3,200 GB/s | 864 GB/s | 3.7x higher on MI250 |
| Bus width | 8,192-bit | 384-bit | 21x wider on MI250 |
| Compute · dense | |||
| FP8 | Not supported | 362 TFLOPS | L40 only |
| FP16 / BF16 | 362.1 TFLOPS | 181.1 TFLOPS | 2x higher on MI250 |
| INT8 | 362.1 TOPS | 362 TOPS | Within 5% |
| FP32 | 90.5 TFLOPS | 90.5 TFLOPS | Identical |
| FP64 | 45.3 TFLOPS | Not supported | MI250 only |
| Platform | |||
| Interconnect | Infinity Fabric 400 GB/s | PCIe only | Infinity Fabric on MI250 only |
| Board power | 500 W | 300 W | 1.7x higher on MI250 |
| Released | Q4 2021 | Q4 2022 | 11 months newer on L40 |
| Precision | |||
| Supported formats | FP16BF16FP32FP64INT4INT8 | FP8FP16BF16TF32FP32INT4INT8 | |
Compute: Vendor peak figures, dense and per GPU. Not measured throughput.
Source: MI250 datasheet and L40 datasheet.
Where to rent them
Providers listing these GPUs, those with both first, so you can switch without moving clouds.
| Provider | MI250 /GPU/hr | L40 /GPU/hr |
|---|---|---|
Cirrascale
|
Reserved (12mo) $1.30 | -- |
GPU.ai
|
-- | On-Demand $0.78 |
Thunder Compute
|
-- | On-Demand $0.79 |
Massed Compute
|
-- | On-Demand $0.86 |
Sesterce
|
-- | On-Demand $0.97 |
Runcrate
|
-- | On-Demand $0.97 |
Oblivus
|
-- | On-Demand $1.05 |
Spheron
|
-- | On-Demand $1.06 |
EmpirioLabs AI
|
-- | On-Demand $1.50 |
1 provider list the MI250 and 12 the L40. Every listing is on the MI250 page and the L40 page.
Similar GPUs
Other comparisons in the same class.
Our data for AMD MI250 was last updated on Sept. 7, 2026, and for Nvidia L40 on Sept. 7, 2026.