News Overview
- NVIDIA has revealed its strategic roadmap for GPUs, CPUs, and system architectures through 2028, focusing on performance enhancements to support AI and high-performance computing demands.
- The roadmap includes the introduction of new technologies, such as the Vera Rubin NVL144 system, Rubin Ultra NVL576 system, and the Feynman GPU.
In-Depth Analysis
Vera Rubin NVL144 System (2026)
- Performance: The Vera Rubin NVL144 system is expected to deliver 3.6 exaflops of FP4 and 1.2 exaflops of FP8 performance.
- Architecture: It will feature the Vera CPU, which has 88 custom cores and 176 threads, paired with the Rubin GPU providing 50 petaflops of FP4 performance.
- Memory and Interconnect: The system will be equipped with 288GB of HBM4 memory and utilize the NVLink 6 interconnect, which offers a data transfer speed of 260TB/second—double that of NVLink 5.
Rubin Ultra NVL576 System (Late 2027)
- Enhanced Specifications: This system will double the number of GPU dies and HBM modules, with four GPU dies and 16 HBM modules.
- Performance: The Rubin Ultra NVL576 is projected to exceed 100 petaFLOPS of FP4 performance, supported by 1 terabyte of HBM4e memory.
- Power Requirements: The system is expected to require up to 600 kW per rack, indicating a need for advanced cooling and power management.
Feynman GPU (2028)
- Introduction: NVIDIA will release the Feynman GPU in 2028, featuring next-generation HBM memory. This will complement the Vera CPUs in system configurations, enhancing performance further.
Commentary
NVIDIA’s roadmap illustrates a clear path toward advancing AI and high-performance computing capabilities. While the focus on performance and innovation is commendable, the significant power demands, particularly for systems like the Rubin Ultra NVL576, may challenge current data center infrastructure and energy efficiency. Nevertheless, NVIDIA’s strategic vision positions it as a leader in AI-driven computing.