Introduction
NVIDIA has released a 45-page whitepaper detailing Vera, its first server CPU. Built around the Olympus core, Vera promises impressive performance with its 88-monolithic cores and advanced features. However, the narrative marketing around these innovations is somewhat lacking. Let's dive into the technical details to understand what truly makes Olympus stand out.
Technical Features of Vera
Vera is a server CPU based on the Olympus core, a 10-wide Arm v9.2 core. It features value prediction, a graph prefetcher, 2 MB of private L2 cache per core, and 164 MB of shared last-level cache. With eight LPDDR5X memory interfaces, it promises a bandwidth of 1.2 TB/s. An interconnect fabric of 3.4 TB/s and a distributed system-level cache support these powerful cores.
Innovations of Olympus
Value Prediction
Value prediction is one of Olympus's key innovations. It allows the prediction of an operation's result so that dependent instructions can continue execution without waiting. This technology has been explored by other players like Apple and AMD, but Olympus appears to offer a more robust implementation.
Graph Prefetcher
The graph prefetcher is not exclusive to NVIDIA. Intel, for instance, uses a similar technology called Data-Dependent Prefetcher, available since 2022. However, Olympus's implementation might offer specific improvements for graphical workloads and AI applications.
Disappointing Marketing
Despite these innovations, the whitepaper dedicates a significant portion to an implicit critique of the x86 architecture. The design choices are framed as a moral narrative rather than technical advancements. This marketing approach detracts from the processor's actual hardware performance.
Conclusion
NVIDIA has clearly developed a powerful product with Vera and Olympus. However, the marketing could have better highlighted the technical advancements rather than engaging in superficial comparisons. For tech decision-makers and entrepreneurs, Olympus's features are what truly matter.
Let's discuss your project in 15 minutes.