In August 2023, Nvidia recently released a new NVIDIA L40S GPU and an OVX server system equipped with the GPU. This computing acceleration graphics card can be used for graphics rendering, artificial intelligence large-scale model training and reasoning, 3D design, and video processing. , AI computing performance is higher than A100 GPU. Since this accelerator card does not support NVLink, and does not choose to use HBM video memory, but uses GDDR6 ECC video memory, the L40S is more suitable for edge computing than A100 and H100.
It is reported that this L40S GPU is paired with the newly released OVX server, which can be applied to AI large model training and reasoning, 3D design, visualization, video processing, industrial digitalization and other purposes. Compared with the A100 GPU, this set of "complementary" L40S The system can "enhance generative AI, graphics and video processing capabilities, and meet growing computing power needs . "
The L40S GPU is an upgraded version of Nvidia's previous L40 GPU, equipped with 48GB GDDR6 ECC memory, the GPU is based on the Ada Lovelace architecture, equipped with the fourth-generation Tensor Core and FP8 conversion engine, and the computing speed is 5 times that of the previous generation. But still does not support NVLink .
Nvidia claims that the L40S's generative AI inference performance is 1.2 times higher than that of the A100, and its training performance is 1.7 times higher. protrude.
Nvidia said that the Nvidia L40S GPU will be available this fall, and each Nvidia OVX server system can carry up to eight L40S accelerator cards, but Nvidia did not disclose the pricing of the GPU.
Nvidia also said that manufacturers such as ASUS, Dell, Gigabyte, HPE, Lenovo, QCT, and Supermicro will "soon launch" OVX server systems equipped with L40S GPUs.
Powerful General-Purpose GPU
NVIDIA L40S GPU gives you breakthrough multi-workload performance experience. The well-designed L40S GPU combines powerful AI computing performance with outstanding graphics and media acceleration functions, and is designed to support next-generation data center workloads. From generative AI and large language model (LLM) inference and training, to 3D graphics, rendering and video, it can handle it with ease.
NVIDIA L40S GPU Specifications
GPU architecture | NVIDIA Ada Lovelace Architecture |
GPU memory | 48GB GDDR6 with error checking |
memory bandwidth | 864GB/s |
interconnection interface | PCIe 4.0 x16: bi-directional 64 GB/s |
CUDA® Core based on NVIDIA Ada Lovelace Architecture | 18,176 |
NVIDIA third generation RT Core | 142 |
NVIDIA fourth generation Tensor Core | 568 |
RT Core performance (unit: TFLOPS) | 212 |
FP32 (unit: TFLOPS) | 91.6 |
TF32 Tensor Core (unit: TFLOPS) | 183 I 366* |
Bfloat16 floating-point performance Tensor Core (unit: TFLOPS) | 362.05 I 733* |
FP16 Tensor Core | 362.05 I 733* |
FP8 Tensor Core | 733 I 1,466* |
INT8 Tensor TOPS Peak INT4 Tensor TOPS Peak |
733 I 1,466* 733 I 1,466* |
form factor | 4.4"(H) x 10.5"(L), Dual Slots |
display port | 4 x DisplayPort 1.4a |
Maximum power consumption | 350 watts |
power interface | 16 pins |
cooling | passive |
Virtual GPU (vGPU) software support | yes |
Supported vGPU profiles | See the Virtual GPU Licensing Guide |
NVENC I NVDEC | 3x I 3x (including AV1 codec) |
Secure Boot via Root of Trust | yes |
NEBS support | level 3 |
Multiple Instance GPU (MIG) support | no |
NVIDIA® NVLink® support | no |
Has been focusing on scientific computing servers for many years, shortlisted for political mining platforms, H100, A100, H800, A800, RTX6000 Ada , a single dual-socket 192 core server is available for sale.