Cisco UCSC-GPU-L4M6 | L4 24GB AI Inference GPU, 72W, Single-Slot PCIe 4.0

Q: Which Cisco UCS servers are compatible with the UCSC-GPU-L4M6?
A: The UCSC-GPU-L4M6 is compatible with Cisco UCS C220 M6 and C240 M6/M8 rack servers. It requires Cisco-specific SBIOS ID integration with CIMC and UCSM management, so the GPU must be procured as a Cisco part number rather than using generic NVIDIA L4 cards. Multiple L4 GPUs can be installed depending on available PCIe riser slots.

Q: What are the primary use cases for the NVIDIA L4 GPU?
A: The L4 excels at AI inference workloads including recommendation engines, natural language processing, computer vision, and generative AI at up to 2.5X the performance of the previous-generation T4. It also delivers hardware-accelerated video transcoding with AV1 codec support, virtual desktop infrastructure (VDI) for up to hundreds of concurrent users, and lightweight graphics rendering with ray tracing and DLSS 3 support.

Q: Does the UCSC-GPU-L4M6 require auxiliary power connectors?
A: No, the L4 draws its entire 70-72W power budget directly from the PCIe 4.0 x16 slot without requiring auxiliary 6-pin or 8-pin power cables. This simplifies installation and makes it ideal for dense server deployments and edge locations where power distribution and cooling capacity are limited.

Q: What AI precision formats does the L4 support?
A: The L4 supports multiple precision formats optimized for different workloads: FP32, TF32, FP16, and BFLOAT16 for training and mixed-precision inference; FP8 and INT8 with fourth-generation Tensor Cores for high-throughput inference. FP8 support delivers up to 485 teraFLOPS of AI performance, making it highly efficient for deploying large language models and transformer-based architectures.

Q: Can I install the L4 in a server that wasn't ordered as GPU-ready?
A: Yes, but retrofitting requires the UCSC-GPUKIT-240M8= GPU kit (for C240 M8 servers) which includes low-profile CPU heatsinks, GPU air duct, thermal paste, and air blockers. You only need one GPU kit per server regardless of how many L4 GPUs you add. Note that the L4 does not require the air duct that double-wide GPUs need, simplifying installation.