What's the price difference between NVIDIA H200 NVL and NVIDIA L40S?

On Runcrate today, NVIDIA H200 NVL starts at $4.63/hr and NVIDIA L40S starts at $0.54/hr per GPU on a 2-GPU machine; NVIDIA L40S is cheaper by 88%.

Can I deploy NVIDIA H200 NVL or NVIDIA L40S on Runcrate today?

NVIDIA H200 NVL is on demand in 2 regions from $4.63/hr, billed by the second on most GPUs. NVIDIA L40S is on demand in 5 regions from $0.54/hr per GPU on a 2-GPU machine, billed by the second on most GPUs.

How much VRAM does each GPU have?

NVIDIA H200 NVL has 141GB HBM3e. NVIDIA L40S has 48GB GDDR6.

H200 NVL or L40S: which holds a 70B model in 8-bit weights?

It needs 70 GB of 8-bit weights. On the H200 NVL: 1 GPU, 71 GB spare. On the L40S: 2 GPUs (96 GB). Pick the H200 NVL: fits on one GPU.

GPU comparison

H200 NVL vs L40S.

The H200 NVL is ahead on memory (141 GB vs 48 GB) and speed (2.3x the FP16 throughput and 5.6x the memory bandwidth). The L40S makes sense when your model fits in 48 GB and its hourly price is lower.

Launch L40S Reserve a cluster

Rent either on Runcrate

H200 NVLfrom $4.63/hr
$0.033 per GB/hr$0.55 per 100 TFLOPS/hr
Launch H200 NVL H200 NVL pricing
L40SCheaperfrom $0.54/hr per GPU
on a 2-GPU machine
$0.011 per GB/hr$0.15 per 100 TFLOPS/hr
Launch L40S L40S pricing

Live hourly price, then per GB of memory and per 100 FP16 TFLOPS.

Which is better for

Which to pick for each job. Memory fit and throughput, from the datasheets.

JobH200 NVLL40SPick

Hold a 70B model in 8-bit weights
70 GB of 8-bit weights
H200 NVL: 1 GPU, 71 GB spare
L40S: 2 GPUs (96 GB)
H200 NVL. Fits on one GPU.
Serve an 8B model in 16-bit weights
16 GB of 16-bit weights
H200 NVL: 1 GPU, 125 GB spare
L40S: 1 GPU, 32 GB spare
H200 NVL. Same GPU count, 5.6x the memory bandwidth.
Fine-tune a 70B model with QLoRA
About 46 GB with 4-bit weights
H200 NVL: 1 GPU, 95 GB spare
L40S: 1 GPU, 2 GB spare
H200 NVL. Same GPU count, 2.3x the FP16 throughput.
Train on a full node
FP16 tensor throughput per node
H200 NVL: 8x, 6,680 TFLOPS, NVLink in groups of 4
L40S: 4x, 1,448 TFLOPS
H200 NVL. 8 GPUs per node, not 4.
Throughput per watt
FP16 TFLOPS per 100 W of TDP
H200 NVL: 139 TFLOPS per 100 W
L40S: 103 TFLOPS per 100 W
H200 NVL. 35% more FP16 per watt.

Compare another pair

Hopper

NVIDIA H200 NVL

from$4.63/hr

Available now · 2 regions

Memory: 141 GB HBM3e
FP16 tensor: 835 TFLOPS
Bandwidth: 4.8 TB/s

Launch H200 NVL Reserve a cluster

Full pricing details

Ada Lovelace

NVIDIA L40S

from$0.54/hr per GPUCheaper

on a 2-GPU machine

Available now · 5 regions

Memory: 48 GB GDDR6
FP16 tensor: 362 TFLOPS
Bandwidth: 864 GB/s

Launch L40S Reserve a cluster

Full pricing details

At a glance

Spec by spec. The longer bar is the better number.

Memory

NVIDIA H200 NVL141 GB

NVIDIA L40S48 GB

Memory bandwidth

NVIDIA H200 NVL4.8 TB/s

NVIDIA L40S864 GB/s

FP16 tensor

NVIDIA H200 NVL835 TFLOPS

NVIDIA L40S362 TFLOPS

Hourly price, from (live)

NVIDIA H200 NVL$4.63/hr

NVIDIA L40S$0.54/hr

Full specs

The complete spec sheet.

Spec	NVIDIA H200 NVL	NVIDIA L40S
Architecture	Hopper	Ada Lovelace
Memory	141 GB	48 GB
Memory type	HBM3e	GDDR6
Memory bandwidth	4.8 TB/s	864 GB/s
FP32	60 TFLOPS	91.6 TFLOPS
FP16 tensor	835 TFLOPS	362 TFLOPS
INT8	1670 TOPS	733 TOPS
TDP	600 W	350 W
Form factor	PCIe	PCIe
NVLink	In groups of 4	No
GPUs per node	8	4
Hourly price (live)	from $4.63/hr	from $0.54/hr per GPU on a 2-GPU machine

Questions

H200 NVL or L40S. What buyers ask first.

Other comparisons