Informasi penting
NVIDIA has officially revised the technical specifications for its upcoming Vera Rubin product line, resulting in lower performance targets for both memory and interconnect fabric. The company reduced the HBM4 memory speed from 11.7 Gbps per pin to 9.6 Gbps per pin, which lowers the total memory bandwidth for the NVL72 rack-scale solution from 1,580 TB/s to 1,400 TB/s.
The company also decreased the NVLink bandwidth per GPU from 1.8 TB/s to 1.5 TB/s, effectively dropping bidirectional bandwidth from 3.6 TB/s to 3 TB/s. This adjustment stems from a reduction in the NVLink Nyquist frequency target from 56 GHz to 47 GHz. By lowering these electrical demands on cables and SerDes, NVIDIA aims to improve manufacturing yields and overall system stability at the cost of peak interconnect performance.
These revisions place the Vera Rubin platform behind AMD's Helios system in key performance metrics. According to the report, the Instinct MI455X provides 432 GB of HBM4 per GPU compared to Rubin's 288 GB. Furthermore, the Helios system offers superior aggregate memory bandwidth and a higher bidirectional interconnect capacity of 3.6 TB/s per GPU, compared to the updated 3 TB/s per GPU for NVIDIA's solution.
