NVIDIA Reveals Vera CPU Architecture Details at Hot Chips 2026, Taking Aim at Server Market
Nashnova编辑部
Nvidia unveiled architecture details of its in-house CPU, Vera, at Hot Chips 2026 — 88 cores, low-power memory, purpose-built for AI agents. Reuters reports the chip is already in mass production, with a $20 billion revenue target for fiscal 2027, directly challenging Intel and AMD on their home turf.
What exactly is Vera?
Vera is Nvidia's first CPU built on its own Olympus core — the prior Grace chip used an off-the-shelf Arm design. This time Nvidia drew the blueprint from scratch.
This means → Nvidia is no longer just "licensing an Arm core and slapping a label on it." It is following the Apple M-series playbook — designing its own CPU from the ground up.
A fixed 88-core configuration, prioritizing single-core performance and execution consistency over raw core count — the polar opposite of AMD's 256-core approach.
Why build a CPU specifically for AI agents?
GPUs handle model training and most inference. But AI agents generate heavy workloads *outside* the GPU: code execution, tool calls, data retrieval, sandboxing, multi-step orchestration — all of which land on the CPU.
In plain terms = the GPU is the brain; the CPU is the hands and feet. As agents act autonomously, the limbs' workload keeps growing — the CPU can no longer be a supporting actor.
Nvidia's internal benchmarks: headless-browser scaling 24% faster than AMD's 96-core EPYC 9655P; Linux kernel compilation 22% faster for native Arm targets, 14% faster for x86 cross-compilation. These results are not independently verified, and no unified benchmark standard for agent workloads exists yet.
Why do two architecture choices matter?
Spatial Multithreading — a technique that reduces interference between two threads sharing the same core. Where rivals chase higher core counts, Vera chases steadier per-core execution.
LPDDR5X memory replaces traditional server RDIMM: eight hot-swappable SOCAMM2 slots, up to 1.5 TB capacity and 1.2 TB/s aggregate bandwidth.
This means → the power gap is the key number: a fully loaded 1.5 TB Vera memory subsystem draws roughly 30–40 watts, versus over 100 watts for a high-capacity RDIMM setup — within the same rack power budget, Vera fits more compute.
Who is already buying in?
SpaceX's AI division announced on August 24 it will deploy standalone Vera CPUs for agent orchestration, code execution, data processing, and simulation workloads, while expanding Grok infrastructure around the broader Vera Rubin platform.
A major Chinese cloud provider plans to test over 300 servers, each equipped with two Vera CPUs, before deciding whether to scale up — still in evaluation, no purchase commitment yet.
This reflects a forking commercialization path: top U.S. customers are deploying; the Chinese market remains at "test first, decide later."
How is AMD responding?
AMD's sixth-generation EPYC 9006 series (codenamed Venice) takes the opposite route: Zen 6 cores on TSMC's 2 nm process, up to 256 cores / 512 threads per socket — nearly three times Vera's core count.
On memory, Venice supports up to 16-channel DDR5 and JEDEC-standard MRDIMM at up to 12,800 MT/s — pushing traditional server memory to the extreme.
In plain terms = same exam, different bets. Nvidia wagers on "fewer, sharper cores"; AMD wagers on "more, wider cores" — 88-core single-thread strength vs. 256-core massive parallelism. Which bet pays off depends on what real-world agent workloads actually look like.
Can the $20 billion target be met?
Reuters reported in June that Nvidia expects Vera revenue to reach $20 billion in fiscal 2027 (ending January 2027). The chip has entered mass production.
This means → for that number to hold, Vera must complete the full arc from mass production to large-scale delivery in under a year — a very tight window.
The real open question is not chip performance but software compatibility and workload migration: server customers are entrenched in the x86 ecosystem, and switching to Vera's Arm architecture means clearing real deployment hurdles.
市场有风险，内容仅供研究参考，不构成投资建议。