Nvidia Says Vera Runs AI Agents Twice as Fast as AMD's 256-Core EPYC Venice, Measured per Megawatt

Nvidia published a set of CPU benchmarks built around AI agents this week, claiming its Vera processor runs agent workloads twice as fast as AMD's 256-core EPYC "Venice" — a comparison measured not in cores or clock speed but in completed tasks per megawatt. In the same material, Nvidia says Vera fits 1.6x more agents per watt than the most energy-efficient x86 part, and delivers 60 percent higher performance per watt than a frequency-optimized 96-core Venice configuration. Every figure in the comparison is vendor-run: Nvidia chose the workloads, the baselines and the 100-megawatt framing.

The argument is about where the bottleneck has moved. An agent does not simply generate an answer: it searches for information, queries databases, writes and tests code, and takes actions in external systems, bouncing between GPU reasoning and CPU execution at every step. As companies put more agents into production, the CPU becomes the queue every task has to pass through. Nvidia is now selling orchestration capacity — how much agent work a data center can finish inside a fixed power budget — rather than raw throughput.

The hardware behind the claim is Vera's 88 custom Armv9.2 "Olympus" cores on a single die with 176 threads, backed by up to 1.5TB of LPDDR5X memory at roughly 1.2TB/s of bandwidth. Nvidia argues that single-thread speed, not core count, decides agent latency: its published material says Olympus cores are up to 1.8x faster single-threaded than a baseline CPU, which it claims lets 85 percent of reinforcement learning evaluations finish inside the same training window, versus 45 percent on the baseline. Nvidia's own platform page lists a different figure — 50 percent higher single-core performance — almost certainly against a different baseline, a discrepancy worth remembering before quoting either number.

Nvidia also argues from fleet behavior. Telemetry from 163,594 agentic sessions shows that more than 97 percent have unique trajectory profiles, which the company says makes it impractical to right-size a fleet with multiple specialized CPU design points. It adds that disabling cores to raise per-core clocks can strand up to 1.5TB of memory capacity on a Vera-based system — a direct answer to the suggestion that buyers can simply trade cores for clock speed.

Early customers supply the deployment evidence, with the same caveat that none of it is independently audited. DeepInfra reports 2.2x faster agent orchestration than its best x86 processor across four tested architectures. Perplexity says agentic coding tasks ran 1.5x faster on Vera and plans to deploy it in an upcoming production system. ClickHouse reports near-linear scaling across 176 cores using its regular software binary, without Vera-specific optimization. Daytona says it brought 1,000 agent sandboxes online in 10.8 seconds using Vera, versus 88.4 seconds on x86, and sustained 12.7 completed jobs per second at 2,000 concurrent sandboxes.

AMD has answered with numbers of its own: in AMD's testing, a 256-core EPYC 9996 Venice platform delivers 1.2x the per-core performance and 2.24x the platform-level performance of an 88-core Vera system. Venice is not a paper product — AMD said on its third-quarter earnings call that revenue shipments began in the quarter, with the first OEM systems on track for the fourth. The gap between AMD's 1.2x per-core claim and its 2.24x platform claim is itself informative: most of the advantage comes from core count, memory configuration and sockets, which is a throughput argument, not a latency one. Aireport covered AMD's September pitch of Venice as an agentic-AI processor, including the same 2.24x figure.

What it means: both vendors can be telling the truth because they are measuring different things, and neither set of numbers has survived third-party replication. Buyers evaluating either chip should ask for the exact workload, the x86 configuration and the software versions behind each multiplier, then rerun the closest match on their own agent code. The real fight is over which metric procurement uses. If agents-per-megawatt displaces core counts as the deciding number, Nvidia has defined the contest; if throughput per socket holds, Venice's 256 cores keep the argument alive.