NVIDIA (NVDA) Designs Vera CPU as a Single Die With 3.4TB/s Core Bandwidth

Single Die, Agent-First Design

NVIDIA (NVDA) has built its Vera CPU on a single silicon die rather than a set of connected chiplets, a design the company says is aimed at AI agents that execute code and call tools between GPU jobs. The Vera team set out the reasoning in a Wccftech interview published on October 18. Their argument runs against the grain of the cloud era, when server chips added cores to lift throughput and accepted lower per-core speed as the price. Agents break that trade-off. Much of an agent's work is sequential: running scripts, processing data, calling external tools before and after the GPU handles model computation. If those CPU stages run slowly, the most expensive asset in the system waits for its next job.

The chip itself is monolithic. All 88 Olympus cores sit on one die and talk through a shared cache and memory system, so core-to-core traffic never crosses an I/O boundary. NVIDIA's technical blog puts the bidirectional bandwidth at up to 3.4TB/s, which the company describes as roughly three times what leading x86 server chips offer. The performance claims deserve a careful read, because the two documents measure different things. The team said Vera is 1.5 times faster than its predecessor Grace and up to twice as fast as the best x86 processors on agent-style tasks such as Python scripting and code compilation. The technical blog, measuring a full socket load that includes compilers, scripting tools, runtimes and agent tool calls, gives up to 1.5 times a competing x86 platform on sandbox performance. NVIDIA has not published both test sets side by side, so the figures cannot be ranked on one basis. Vera also runs spatial multithreading, a patented scheme that splits resources between a thread doing the work and a thread handling input and output. Phoronix testing recorded a gain of about 20% from the feature, although the blog does not spell out the conditions behind that measurement.

Standalone Racks and a Tilting Ratio

The commercial plan behind the design has shifted as much as the silicon. Grace was built as a host node inside a GPU rack, feeding data to the accelerators. Customer deployments changed that picture: NVIDIA found that heavy agent tool calling and reinforcement learning workloads wanted whole racks of CPUs with no GPUs attached at all. Vera will therefore be sold on its own, aimed at racks built for agent inference and reinforcement learning, in addition to its place in the Vera-Rubin superchip rack. The team had no official answer for where the CPU-to-GPU ratio in AI systems lands, but said it believes the balance is moving toward the CPU. It offered one datapoint: some agent sandbox providers have already deployed as many as 400,000 sandboxes for a single customer. Sandboxes are the isolated environments where agents run generated code, and that work is pure CPU. The architectural reversal does not stop at NVIDIA either. Taiwan's server and semiconductor supply chain, which built capacity around the chiplet era's multi-die boards, sits directly in the path of a design trend returning to large single dies.

NVDA Presses Its Strongest Ceiling

The product story is long-cycle, but the tape is short. COINOTAG data shows NVIDIA stock trading at $217.86, up 2.11% on the day, sitting directly under the strongest resistance in our composite scoring at $217.95, which rates 100/100. The nearest floor, $209.53, scored 95/100 and coincides with a Fibonacci retracement level, so a dip toward it would be technically routine rather than a judgment on Vera. Momentum has not confirmed the move yet: the daily RSI reads 40.5 and the MACD signal stays bearish. A close above $217.95 would be the first sign the market is paying for the CPU rack thesis; failure there leaves the stock inside its daily downtrend.