Nvidia's Vera CPU: 88 Cores, 176 Threads Optimized for Agentic AI Workloads
Justgototheeffinmoon · reddit · 2026-07-22
Architectural details of Nvidia's upcoming server CPU, Vera, have surfaced. The chip features 88 custom Olympus cores and 176 threads (via spatial multithreading). It supports up to 1.5TB of LPDDR5X memory with 1.2TB/s bandwidth on a monolithic die, marking a significant step up from the previous Grace architecture.
Architecture
- The Olympus core is deliberately wide: 18 execution pipes, fetching up to 16 instructions per cycle, and decoding up to 10 fused instructions per cycle.
- It includes a neural branch predictor resolving two branches per cycle, alongside value prediction and a graph prefetcher tuned for pointer-heavy access patterns, explicitly targeting graph traversals common in agentic AI workloads.
Performance
- In SPEC CPU 2026 benchmarks, Vera posts a modest 3% multi-threaded integer lead over AMD's EPYC 9755 (925 vs 898).
- Nvidia claims more decisive single-thread performance and 29.3x scaling on a 32-core graph traversal benchmark versus 10x for AMD, though these results are vendor-supplied.
Related event: NVIDIA Reveals More Vera CPU Details Ahead of AMD Event(8 posts)→
More from Infra
- Anthropic's ~11 GW contracted compute estimated at ~42% NVIDIA GPUs, ~4.5 GW Google TPU — ns123abc · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27