NVIDIA Unveils Rubin HPC Single Card with Over 50,000 CUDA Cores, Setting a Record
2 day ago / Read about 0 minute
Author:小编   

On November 16 local time, NVIDIA officially unveiled the Rubin HPC GPU under its Vera Rubin platform at the SC26 conference. Targeting scientific, engineering, and analytical workloads, its core mission is to accelerate AI agent-driven simulation and data processing. A single GPU integrates over 50,000 CUDA cores, setting a record for NVIDIA's GPU product line, with a core count nearly twice that of the existing Rubin GPU and nearly 2.5 times that of the consumer-grade RTX 5090. Its native FP64 double-precision performance returns to 33 TFLOPS, and with doubled CUDA cores and 3 times the vector cores of Blackwell, the effective throughput improvement potential for actual simulation workloads is even greater. The product adopts a balanced design, featuring sparse data acceleration capabilities and seventh-generation NVLink interconnect technology to prevent cores from idling. Rubin HPC is expected to launch in 2028, and NVIDIA has not yet disclosed details such as its HBM specifications and power consumption.