XMC-4906

NVIDIA® Ampere® GPU-based XMC with Tensor Cores for AI Deployment, & Up to 8.25 TFLOPS of Peak Processing

Designed and manufactured by WOLF Advanced Technology, the XMC-4906’s NVIDIA Ampere architecture GPU XMC card has 2560 CUDA cores, 80 Tensor core dedicated to AI accelerated compute, and 20 Ray Tracing (RT) cores for superior rendering speeds. Designed to work in conjunction with Tensor RT, CUDA, and cuDDN, the Ampere’s Tensor cores add support for new data types as well as a sparsity feature. This new generation of Tensor and RT cores delivers twice the throughput of the previous generation. The NVIDIA Ampere architecture introduces significant improvements, which can increase throughput in the CUDA cores two to four times depending on the workload. Adding the move to the 8 nm manufacturing process, the Ampere architecture provides 144 GFLOPS/W compared to 51 GFLOPS/W of the P2000 used in the previous generation XMC-4902.

Moving data quickly and efficiently is the key to achieving the best performance out of a GPU. The XMC-4906 uses 128-bit GDDR6 memory for a maximum memory bandwidth of 192 GB/s. The GPU supports up to Gen-4 PCIe; however, the XMC connector and carrier width determine the maximum throughput. The PCIe is a x8 to the connector.