The shift to agentic AI creates a new CPU requirement for the AI factory: fast cores, massive memory bandwidth and the ability to sustain high performance when all cores are active.<\/p>\n
Initial benchmark results published by Phoronix<\/a> today show that the NVIDIA Vera CPU meets this need. For this first public look, the benchmark scope was centered on the agentic workloads Vera was designed for in the modern data center.<\/p>\n The Vera CPU delivers the throughput AI factories need while optimizing platform power. Eighty-eight NVIDIA custom Olympus cores, 1.2TB\/s of memory bandwidth and a high-speed, on-chip fabric results in a CPU platform that combines core performance and memory bandwidth in an efficient power envelope.<\/p>\n NVIDIA Olympus Delivers Aggressive Performance<\/p>\n At the heart of Vera are custom NVIDIA Olympus CPU cores. Fully compatible with the Armv9.2 instruction set architecture, Olympus is designed for the sequential CPU work underpinning agentic AI: branch-heavy runtimes, sandboxed code, data processing and orchestration.\u00a0<\/p>\n Vera\u2019s monolithic die, wide cores, advanced branch prediction and the second-generation NVIDIA Scalable Coherency Fabric help Vera keep data moving across all 88 cores.<\/p>\n Phoronix\u2019s testing of a single-socket Vera CPU \u2014 rated at 450-watt thermal design power with less than 30 watts of memory power \u2014 showed that it delivers outstanding performance within that power profile, along with generational gains across a broad array of workloads spanning code compilation, file compression, video transcoding, Python, Java and database management.\u00a0<\/p>\n These are the same kinds of CPU-heavy tasks that agents and AI factories run every day: compiling code, executing runtimes, compressing data, querying databases and coordinating large software stacks.<\/p>\n \u201cGoing into this, I didn\u2019t really know what to expect of NVIDIA\u2019s Vera with the new Olympus cores,\u201d wrote Michael Larabel, founder and principal author of Phoronix. \u201cBut in the end I was left realizing this is the most formidable competition to Intel and AMD x86_64 processors ever realized.\u201d\u00a0<\/p>\n \u2018Incredible Advantage\u2019 in Memory Performance<\/p>\n Agentic workloads are not limited by core count alone. They need high core utilization and sustained memory bandwidth, making memory performance per watt a critical part of overall CPU efficiency.\u00a0<\/p>\n Vera incorporates a second-generation LPDDR5X memory subsystem, enabling dramatically lower energy per bit compared to DDR5. This allows Vera to offer up to a massive 1.2 TB\/s of bandwidth \u2014 up to 2x the peak memory bandwidth compared with traditional CPUs in less than 30 watts of memory power, as opposed to more than 100 watts for traditional DDR5.\u00a0<\/p>\n In Phoronix STREAM TRIAD testing, Vera sustained 90% of its peak memory bandwidth \u2014 achieving the highest percentage of rated peak bandwidth of any CPU tested by Phoronix \u2014 and delivered over 4x the memory bandwidth per core compared with traditional x86 CPUs.<\/p>\n \u201cNVIDIA Vera with its LPDDR5X memory was showing its incredible advantage in memory performance over current Intel Xeon and AMD EPYC processors,\u201d Larabel wrote.<\/p>\n However, peak bandwidth is only part of the story. AI factory workloads run many sandboxes, tool calls and data services at the same time. In separate testing with Vera, Prime Intellect found<\/a> that Vera maintained high bandwidth and low, consistent memory latency as more workloads ran in parallel \u2014 the kind of predictable performance needed for agentic AI.<\/p>\n A Large Generational Leap \u2014 and Leadership in Phoronix Testing<\/p>\n Compared with the prior-generation NVIDIA Grace CPU, Vera delivered a 1.6x geometric mean increase in Phoronix\u2019s testing \u2014 an incredible generation-over-generation gain.\u00a0<\/p>\n \u201cThe difference from Grace to Vera was consistently exceeding my expectations for gen-on-gen performance we typically see for processors,\u201d Larabel wrote. \u201cNVIDIA\u2019s Vera CPU with its in-house-designed Olympus CPU cores ends up packing a heavy-hitting punch with competitiveness to Intel\/AMD x86_64 CPUs that I have never seen out of any other ARM or non-x86_64 processors.\u201d<\/p>\n Vera led the tested CPU field, delivering a 1.5x overall performance advantage compared with a latest-generation 128-core x86 processor. The gains showed up in practical developer workloads. Single-socket Vera compiled a default Linux kernel in just 20 seconds, the fastest result Phoronix measured in that test. Vera delivered 2x faster Linux kernel compilation on a per-core basis compared with a 128-core processor.<\/p>\n \u201cOn a [geometric] mean basis, the NVIDIA Vera delivered 10% better performance than the AMD EPYC 9575F 5.0 GHz high frequency processor,\u201d Larabel wrote.<\/p>\n Vera in Customer Testing, Coming Soon From Partners<\/p>\n At NVIDIA GTC, NVIDIA announced widespread ecosystem support for Vera, spanning AI natives, supercomputing centers, cloud service providers and infrastructure providers.\u00a0<\/p>\n