In the recent SIGGRAPH 2023 event, NVIDIA took a significant step by revealing their innovative NVIDIAL40S GPU, alongside the integration of L40S within the NVIDIA OVX server. Based on available data, these offerings have been meticulously tailored for diverse applications encompassing AI model training and inference, 3D design and visualization, video processing, as well as industrial digitization. The objective is clear: fortifying the AI, graphics, and video processing capacities of enterprises, organizations, and data centers to effectively address the surging computational requirements.
Central to the L40S GPU and OVX server is their role in enhancing the training and inference stages of generative AI models, demonstrating the potential to amplify the efficiency of these models.
Openly disclosed information indicates that the fresh L40S GPU acceleration card, a successor to the L40 series, emanates from NVIDIA. This GPU, crafted upon the Ada Lovelace architecture, incorporates the fourth-generation Tensor Core and FP8 conversion engine, boasting a computational velocity of up to 1.45 PFlops. Further adding to its prowess, the L40S GPU boasts an assembly of 18,176 CUDA cores, providing an impressive nearly fivefold enhancement in single-precision floating-point (FP32) performance (91.6 TFlops) compared to the NVIDIA A100 GPU. Unlike the A100 and H100 GPUs, the L40S forgoes NVLink support, opting instead for GDDR6 ECC memory, positioning it as a more adept solution for edge computing scenarios.
Liu Jingzhao, analyst of Shanghai Securities, shared his insights:
- Due to the constraints of the PCIe Gen 4 interface, the L40S might find its application in GPU network computing relatively limited, potentially having a marginal influence on the demand for ongoing 800G optical modules.
- In comparison to the A100, the L40S showcases computational efficiency advantages within specific generative AI model contexts, accompanied by heightened cost-effectiveness.
- Leveraging mature technologies like GDDR6, the L40S guarantees stability in upstream supply.
- The highlighted advantages attributed to the L40S might prompt an augmented acquisition trend among downstream cloud providers, subsequently birthing fresh growth possibilities for middle-to-high speed optical modules, specifically the 200G and 400G representations.
NVIDIA's rollout of the L40S GPU stands poised to amplify AI computing capabilities, possibly catalyzing greater procurement activities among downstream cloud providers. This momentum could conceivably pave the path for heightened demand within the middle-to-high speed optical module market.