The shift to agentic AI creates a brand new CPU requirement for the AI manufacturing facility: quick cores, large reminiscence bandwidth and the power to maintain excessive efficiency when all cores are energetic.
Preliminary benchmark outcomes revealed by Phoronix as we speak present that the NVIDIA Vera CPU meets this want. For this primary public look, the benchmark scope was centered on the agentic workloads Vera was designed for within the trendy information heart.
The Vera CPU delivers the throughput AI factories want whereas optimizing platform energy. Eighty-eight NVIDIA customized Olympus cores, 1.2TB/s of reminiscence bandwidth and a high-speed, on-chip material ends in a CPU platform that mixes core efficiency and reminiscence bandwidth in an environment friendly energy envelope.
NVIDIA Olympus Delivers Aggressive Efficiency
On the coronary heart of Vera are customized NVIDIA Olympus CPU cores. Absolutely appropriate with the Armv9.2 instruction set structure, Olympus is designed for the sequential CPU work underpinning agentic AI: branch-heavy runtimes, sandboxed code, information processing and orchestration.
Vera’s monolithic die, huge cores, superior department prediction and the second-generation NVIDIA Scalable Coherency Material assist Vera preserve information shifting throughout all 88 cores.
Phoronix’s testing of a single-socket Vera CPU — rated at 450-watt thermal design energy with lower than 30 watts of reminiscence energy — confirmed that it delivers excellent efficiency inside that energy profile, together with generational features throughout a broad array of workloads spanning code compilation, file compression, video transcoding, Python, Java and database administration.
These are the identical sorts of CPU-heavy duties that brokers and AI factories run each day: compiling code, executing runtimes, compressing information, querying databases and coordinating massive software program stacks.
“Going into this, I didn’t actually know what to anticipate of NVIDIA’s Vera with the brand new Olympus cores,” wrote Michael Larabel, founder and principal creator of Phoronix. “However in the long run I used to be left realizing that is probably the most formidable competitors to Intel and AMD x86_64 processors ever realized.”
‘Unbelievable Benefit’ in Reminiscence Efficiency
Agentic workloads should not restricted by core rely alone. They want excessive core utilization and sustained reminiscence bandwidth, making reminiscence efficiency per watt a vital a part of general CPU effectivity.
Vera incorporates a second-generation LPDDR5X reminiscence subsystem, enabling dramatically decrease power per bit in comparison with DDR5. This permits Vera to supply up to an enormous 1.2 TB/s of bandwidth — as much as 2x the height reminiscence bandwidth in contrast with conventional CPUs in lower than 30 watts of reminiscence energy, versus greater than 100 watts for conventional DDR5.
In Phoronix STREAM TRIAD testing, Vera sustained 90% of its peak reminiscence bandwidth — reaching the very best proportion of rated peak bandwidth of any CPU examined by Phoronix — and delivered over 4x the reminiscence bandwidth per core in contrast with conventional x86 CPUs.
“NVIDIA Vera with its LPDDR5X reminiscence was displaying its unbelievable benefit in reminiscence efficiency over present Intel Xeon and AMD EPYC processors,” Larabel wrote.
Nevertheless, peak bandwidth is simply a part of the story. AI manufacturing facility workloads run many sandboxes, device calls and information companies on the similar time. In separate testing with Vera, Prime Mind discovered that Vera maintained excessive bandwidth and low, constant reminiscence latency as extra workloads ran in parallel — the type of predictable efficiency wanted for agentic AI.
A Giant Generational Leap — and Management in Phoronix Testing
In contrast with the prior-generation NVIDIA Grace CPU, Vera delivered a 1.6x geometric imply enhance in Phoronix’s testing — an unbelievable generation-over-generation achieve.
“The distinction from Grace to Vera was persistently exceeding my expectations for gen-on-gen efficiency we sometimes see for processors,” Larabel wrote. “NVIDIA’s Vera CPU with its in-house-designed Olympus CPU cores finally ends up packing a heavy-hitting punch with competitiveness to Intel/AMD x86_64 CPUs that I’ve by no means seen out of some other ARM or non-x86_64 processors.”
Vera led the examined CPU subject, delivering a 1.5x general efficiency benefit in contrast with a latest-generation 128-core x86 processor. The features confirmed up in sensible developer workloads. Single-socket Vera compiled a default Linux kernel in simply 20 seconds, the quickest consequence Phoronix measured in that check. Vera delivered 2x quicker Linux kernel compilation on a per-core foundation in contrast with a 128-core processor.
“On a [geometric] imply foundation, the NVIDIA Vera delivered 10% higher efficiency than the AMD EPYC 9575F 5.0 GHz excessive frequency processor,” Larabel wrote.
Vera in Buyer Testing, Coming Quickly From Companions
At NVIDIA GTC, NVIDIA introduced widespread ecosystem assist for Vera, spanning AI natives, supercomputing facilities, cloud service suppliers and infrastructure suppliers.
NVIDIA has additionally delivered the primary Vera CPUs to main AI firms and cloud suppliers, marking an essential milestone as Vera strikes towards companion availability within the second half of the yr.
Vera can be accessible from companions in dual- and single-socket techniques, with air-cooled and liquid-cooled choices to assist AI manufacturing facility deployments, from commonplace enterprise information facilities to high-density agentic AI infrastructure.
Be taught extra about NVIDIA Vera.
