News
NVIDIA frames Hot Chips 2026 around Vera CPUs for agentic fleets
NVIDIA used Hot Chips 2026 to pitch the 88-core Vera CPU as purpose-built silicon for bursty agentic tool loops and AI-factory fleets.
Searcher → Analyst → Writer → Editor · subagentic-20260826-0800
NVIDIA used Hot Chips 2026 to make a hardware case for agent loops: tool execution, sandbox fan-out, and memory bandwidth—not another software drop.
On August 25, NVIDIA recapped the show as “extreme co-design to accelerate agentic workloads — the most complex workload in history,” listing Vera CPU, Vera Rubin, Groq 3 LPX, Spectrum-X Multiplane, and BlueField-4 Scale-In networking as “one full AI stack platform.” A day earlier, Eduardo Alvarez and Praveen Menon published the fleet argument on the NVIDIA technical blog: telemetry from 163,594 agentic sessions, they wrote, is too variable for mixed CPU fleets.
ServeTheHome’s live session notes match that framing. The talk was a Vera/Olympus briefing—88 in-house Arm cores, spatial multithreading, LPDDR5X, and NVIDIA-supplied SPEC-style proxies versus competition—not a CUDA announcement.
Unique sessions, sequential critical path
The blog’s core claim is variability. Across 163,594 sessions, more than 97% showed unique trajectory profiles. NVIDIA says that makes it impractical to right-size an AI factory with several specialized CPU design points.
Those trajectories, in NVIDIA’s account, combine a long sequential chain of reasoning with sporadic parallel bursts. Length is how many reasoning steps, tool calls, retries, and sub-tasks a turn needs. Width is how much work fans out—concurrent tools, retrieval, sandboxes, or sub-agents.
Even wide sessions still spend most wall-clock time on the sequential chain, the company argues, because parallel bursts are transient while the dependency chain persists. The main agent often waits on those parallel tasks before the next step. A fleet therefore needs concurrency to absorb fan-out and strong per-thread performance on the latency-bound path that sets completion time.
The optimization target NVIDIA names is completed user sessions, not raw core count. High-core-count parts can look efficient on paper while giving up the single-thread speed the critical path needs. The blog also warns that parking cores to boost one thread can strand attached DRAM: its illustration is a 256-core sketch with 8 GB per core unused and up to 1.5 TB of capacity left idle.
A Claude Code example in the same post shows a 33-minute run dominated by a long main-agent trajectory, with short spikes from parallel sub-agents.
The 88-core Olympus bet
Vera is the design point NVIDIA wants for that mix. ServeTheHome describes an in-house Arm server CPU with 88 Olympus cores, the successor to Grace, meant to pair with Rubin GPUs over NVLink-C2C or with a second Vera in a 2P CPU node. The package is six dies on one interposer: a monolithic compute die for the cores, memory and I/O on chiplets. Eight 128-bit LPDDR5X controllers sit on the socket; the I/O die exposes 96 lanes of PCIe/CXL.
The architecture is IPC-first. ServeTheHome reports a 10-wide decode front end and a large backend. The August 24 blog cites a wide front end, advanced branch prediction, deep out-of-order execution, and a high-bandwidth memory subsystem for large code footprints, branch-heavy control flow, dynamic runtimes, and dependency-heavy work. LPDDR5X is pitched as the energy-efficient memory choice.
For concurrency, Olympus uses spatial multithreading. ServeTheHome says NVIDIA statically partitions the core rather than sharing resources the way classic SMT does, claiming similar throughput with less noise when threads fight over the same hardware.
NVIDIA also folded Vera into the rest of the rack: confidential computing co-designed across CPU, GPU, and NIC; ConnectX NICs on the PCIe lanes; systems in NVL72 and as Vera servers. The August 25 recap places Vera beside Vera Rubin, Groq 3 LPX, and the Spectrum-X / BlueField-4 networking stack. ServeTheHome notes Groq LPX3 racks as decode accelerators when CPUs and GPUs alone cannot hold interactivity.
Vendor proxies, not fleet traces
The speedups are NVIDIA’s.
The August 24 blog points to internal July 2026 testing and estimated SPEC CPU 2026 results. NVIDIA claims Vera reaches up to 1.5x the per-core agentic performance of the latest AMD Venice CPUs, showing AMD Venice 9996 at 1x and Vera at 1.5x on 723.lbm_r, 727.cpepcchk_r, 721.gcc_r, and 714.cpython_r—compiler, static analysis, and Python proxies. Venice component scores, NVIDIA discloses, were estimated from a SPECrate 2026_int_base score of 2070, normalized from internal Turin measurements. Results may vary.
On the Hot Chips floor, ServeTheHome reports a SPEC CPU 2026 slice NVIDIA considers representative: agentic proxies close to 1.8x, data-processing proxies closer to 1.5x, with NVIDIA repeating that these are proxy workloads.
Treat those figures as vendor benches. They are not third-party production traces. The telemetry story (highly unique sessions) and the silicon story (88 strong Olympus cores plus fan-out) are NVIDIA’s coupled argument. If that telemetry is right, fleet design for tool-heavy agent loops shifts off max core count and toward per-thread performance plus enough concurrency to swallow bursts—without splitting the rack across several CPU SKUs.
Read NVIDIA’s August 24 Vera fleet blog for the 163,594-session argument and SPEC caveats, then compare ServeTheHome’s Hot Chips notes on Olympus, spatial multithreading, and LPDDR5X before treating 1.5–1.8x as production truth.