Nvidia Details Vera CPU at Hot Chips: 88-Core Chip Targets Agentic AI, Outpaces AMD EPYC in Early Benchmarks

Nvidia used its Hot Chips 2026 presentation to flesh out the architecture of Vera, its first CPU built around a completely custom core design (dubbed Olympus). Unlike the previous Grace CPU, which used a stock Arm core, Vera represents Nvidia's deepest foray into processor design and is explicitly aimed at the rapidly evolving 'agentic AI' market — autonomous software agents that perform complex, multi-step tasks.

Vera ships as a single monolithic 88-core SKU, a contrast with the chiplet-based designs typical of x86 competitors. It employs what Nvidia calls 'spatial multithreading,' a novel approach to parallel processing that the company says is optimized for the unpredictable, branching workloads characteristic of AI agents. The chip is paired with a new memory subsystem using LPDDR5X (marketed as SOCAMM2) that delivers 1.2 TB/s of bandwidth.

Nvidia provided concrete performance comparisons against AMD's 96-core EPYC 9655P (Turin). In a headless browser test designed to simulate how an AI agent would fetch and process web information, Vera was 24% faster. In Linux kernel compilation — a proxy for the software-building tasks agents often perform — Vera achieved a 22% improvement when compiling natively for AArch64 and 14% faster when cross-compiling for x86.

However, Nvidia acknowledged that benchmarking agentic AI remains a challenge, calling it 'the most complex computing workload in history.' Performance can vary dramatically depending on how an agent's workflow is optimized. For instance, agents can strip out GUI rendering, font loading, and media decoding to run a browsing workflow up to 4.5x faster than a human-driven process.

Notably, AMD executive Joe Macri was quoted as being 'very happy to see Nvidia's Vera performance results,' adding, 'I actually thought we were beating them by smaller numbers.' The comment suggests that while Nvidia holds a lead in these specific benchmarks, AMD's Venice architecture (due to compete with Vera) may be closer than Nvidia's marketing suggests.

Nvidia has previously disclosed that Vera's Olympus core features a 10-wide decode, a large branch prediction unit with neural prediction, and a massive L1 cache — design choices aimed at minimizing latency for the erratic instruction streams typical of AI agents. The company also confirmed that Vera will sample to customers in the first half of 2026, with volume shipments expected later that year.

Why This Matters

Nvidia's entry into the CPU market with a custom core breaks the x86 duopoly of AMD and Intel in data center processors, giving cloud providers and AI companies another architecture to consider.
The focus on agentic AI workloads signals that Nvidia sees autonomous agents — not just LLM inference — as the next major compute demand driver, potentially reshaping server procurement strategies.
If Vera's spatial multithreading and memory architecture deliver real-world gains, it could pressure AMD and Intel to accelerate their own custom-core or multi-threading innovations.

Background

Nvidia acquired Arm in 2022 (though the deal was ultimately blocked), but the company had already begun developing its own Arm-compatible cores. The Grace CPU in 2023 used off-the-shelf Arm Neoverse cores, giving Nvidia a foothold in CPU design. Vera represents a major leap: Olympus is a ground-up Nvidia design, optimized not for general-purpose computing but for the specific demands of AI agent pipelines — which mix traditional CPU tasks (code compilation, web scraping, data processing) with GPU-accelerated neural network calls.

Hot Chips has historically been the venue where chipmakers reveal architecture details. Nvidia's presentation this year is notable for offering direct competitive benchmarks, a departure from the more theoretical disclosures typical of the event.

Key Perspectives

Nvidia: Argues that the CPU is becoming the bottleneck in agentic AI workflows, and that existing architectures (including its own Grace) are not optimized for the mixed, latency-sensitive workloads agents generate. Spatial multithreading and the massive bandwidth of SOCAMM2 are designed to keep cores fed with data. AMD: Appears confident that its upcoming Venice architecture (based on the new Zen 6 core) will close the gap. Macri's public 'happiness' with the results hints that AMD may have internal data showing narrower differences — or that Venice has unannounced features that make the comparison moot. Critics/Skeptics: Some analysts question whether headless browser benchmarks and kernel compilation accurately predict agentic AI performance, which involves complex chains of GPU calls, network I/O, and decision-making. Nvidia's own admission that agentic workloads are 'the most complex in history' underscores the difficulty. Moreover, Vera is a monolithic 88-core die; if yields are poor or power consumption is high, cost and adoption could suffer.

What to Watch

Real-world agentic AI benchmarks (e.g., SWE-bench, GAIA) that measure end-to-end agent performance, not just proxy tasks.
AMD's Venice launch and independent benchmarks — expected later in 2026, they will provide the first direct head-to-head comparison.
Arm ecosystem maturity: Vera may need software optimization from OS vendors, compiler teams, and AI frameworks to realize its full potential; watch for announcements from Red Hat, Canonical, and LLVM.
Nvidia's own GPU integration: How tightly Vera will pair with upcoming Rubin GPUs over NVLink could be a decisive advantage over AMD's chiplet approach.

ZOTPAPER