Product Listing
In late February 2026, JH Semiconductor completed intake inspection and system entry for NVIDIA L40S data center GPU cards. The L40S is built on the Ada Lovelace architecture, carries 48GB of GDDR6 memory, supports FP8-precision inference acceleration, and has a 350W TDP. It is now officially listed, with quotes available for both single-card and volume purchases.
Positioned for Mixed Workloads
The L40S is distinctive in serving two workload classes at once. It supports low-precision inference acceleration such as FP8, enabling it to run large language model inference; at the same time it retains a full graphics pipeline, making it well suited to 3D visualization, rendering and video processing. As a standard PCIe add-in card it can be deployed in mainstream general-purpose server platforms without a dedicated system architecture, which is why enterprise customers often choose it as a pragmatic option — alongside flagship training GPUs — for mixed inference and graphics workloads.
Complementary Support
Deploying GPU compute in practice usually requires coordinated selection of the server system, memory and high-performance storage. JH Semiconductor operates server, memory and enterprise SSD product lines in parallel, and can offer complementary selection guidance based on the characteristics of a customer's inference and rendering workloads. The products have been warehoused and are open for customer inquiries.