Nvidia’s vera CPUs supercharge AI chip design workflows

Drip

The gist

NVIDIA’s new Vera CPUs and AI-driven Agent Toolkit are turbocharging chip design, slashing verification times and giving NVIDIA and its partners an uncatchable edge.

What to know

  • The Vera CPU architecture—88 custom Olympus cores on a 3nm die—delivers up to 1.5x faster EDA workflows, while the Vera Rubin NVL72 platform slashes inference costs by 90%.
  • NVIDIA’s Agent Toolkit, used by giants like Samsung and TSMC, accelerates quantum chemistry and lithography simulations by up to 50x and 20x respectively.
  • Major EDA players like Cadence and Synopsys are building autonomous AI engineers atop NVIDIA’s platform, reporting design workflows up to 15x faster and RTL validation 50x speedier.

Vera CPUs: Purpose-Built for AI EDA

NVIDIA’s Vera CPUs leverage 88 custom cores, spatial multithreading, and record memory bandwidth to offload EDA workloads from GPUs, boosting system density and throughput for next-gen chip design.

NVIDIA's Vera CPU architecture, featuring 88 custom Olympus cores on a monolithic TSMC 3nm die, represents a deliberate departure from chiplet designs to minimize latency and optimize for agentic AI workloads. With advanced features like spatial multithreading—where each core supports two dedicated threads for concurrent high-throughput and lighter tasks—and a high-bandwidth LPDDR5X memory subsystem delivering up to 1.2 TB/s aggregate bandwidth, Vera is engineered to meet the demanding latency-sensitive and memory-intensive needs of EDA verification and simulation workflows. This design enables Vera to shoulder computational burdens traditionally handled by GPUs, supported by cutting-edge packaging, networking, and cooling solutions, thereby enhancing overall system density and throughput in chip design environments.

Integrated within NVIDIA’s Vera Rubin NVL72 platform—a rack-scale system combining 36 Vera CPUs with 72 Rubin GPUs and advanced interconnects like NVLink 6 and ConnectX-9 SuperNICs—Vera CPUs contribute to a tightly coordinated hardware-software ecosystem that dramatically improves EDA workflow efficiency. Early benchmarks from CoreWeave’s DeepSeek-R1 highlight that Vera Rubin achieves roughly 10 times more tokens per second per megawatt compared to previous generations, with inference costs per token reduced by up to 90%, underscoring the platform’s power-efficient performance gains in verification and simulation tasks.

NVIDIA’s internal testing with leading EDA tools such as Cadence Jasper and Synopsys VCS demonstrates that Vera CPUs deliver up to a 1.5x performance boost on compute-intensive verification and simulation workloads, enabling faster bug detection and more design iterations within the same development cycle. This acceleration not only shortens individual verification runs but also increases throughput, allowing engineers to explore a broader range of design alternatives and complete more comprehensive validation, thereby enhancing productivity in critical chip design stages.

By embedding Vera CPUs directly into its own chip design workflows, NVIDIA is creating a recursive performance feedback loop where accelerated EDA processes enable quicker design cycles, which in turn produce improved Vera architectures for future iterations. This strategic integration of custom CPU hardware with advanced GPU-accelerated sparse solver libraries and AI-driven toolkits exemplifies NVIDIA’s holistic approach to co-designing silicon, software, and systems engineering, potentially conferring a sustainable competitive advantage in next-generation semiconductor development.

Sources

Agent Toolkit Powers AI Engineers

PhysicsNeMo and CUDA-X libraries embedded in NVIDIA’s Agent Toolkit let companies like Samsung and TSMC compress chip simulation cycles from weeks to hours by tightly coupling physics-based AI with GPU acceleration.

NVIDIA’s Agent Toolkit has been significantly enhanced by integrating its PhysicsNeMo physics-AI libraries alongside a suite of GPU-accelerated sparse solver libraries, including the newly introduced cuISS for iterative solvers, cuDSS for direct sparse solutions, and cuEST for quantum chemistry simulations. This integration empowers autonomous AI engineers to seamlessly invoke advanced physics-based simulation tools within chip design workflows, enabling partners like Samsung, TSMC, and Synopsys to achieve dramatic speedups—up to 50x faster quantum-chemistry workloads and 20x faster computational lithography—thereby compressing traditionally weeks-long verification and simulation cycles into mere hours.

Major EDA and semiconductor companies such as Cadence, Siemens, Synopsys, Samsung, and TSMC are building their own autonomous AI engineers atop NVIDIA’s Agent Toolkit, embedding these physics-aware CUDA-X libraries at the core of their design automation stacks. Cadence, for example, reports up to 15x faster design workflows and a halving of time to market, while Synopsys’s AI-driven verification closure flow achieves a 50x acceleration in RTL validation with a 20% boost in coverage. This collaborative ecosystem illustrates how NVIDIA’s AI agents are not only accelerating chip design but also enhancing design quality and decision-making by enabling engineers to explore vastly more design alternatives through high-fidelity, accelerated iterative loops.

While PhysicsNeMo is open source under Apache 2.0 and CUDA-X libraries serve as free drop-in replacements for hand-tuned parallel code, their exclusive dependence on NVIDIA GPUs strategically ties this advanced AI-driven chip design acceleration tightly to NVIDIA’s hardware ecosystem. This hardware-software-AI synergy is further amplified by the deployment of NVIDIA’s Vera CPUs, which deliver up to 1.5x higher performance on critical verification and simulation tasks, creating a recursive performance feedback loop that continuously accelerates and refines next-generation processor development.

Sources

Recursive Loop Accelerates Innovation

NVIDIA’s own EDA workflows run on Vera CPUs, creating a self-reinforcing cycle where each hardware generation speeds up its own development and deepens NVIDIA’s competitive moat.

NVIDIA’s strategic embedding of its Vera CPUs within its own electronic design automation (EDA) workflows creates a powerful recursive performance feedback loop that accelerates chip design cycles and innovation. By running Vera-enhanced tools like Cadence Jasper and Synopsys VCS faster than previous iterations, NVIDIA shortens each design cycle, producing improved Vera CPUs that in turn further speed up verification and simulation workloads. This self-reinforcing ecosystem not only boosts design velocity but also deepens NVIDIA’s technical moat by tightly integrating hardware, software, and AI into a cohesive platform, thereby establishing a competitive edge that is difficult for rivals to replicate.

Internal testing validates the efficacy of this recursive loop, with NVIDIA reporting up to a 1.5x performance improvement on critical verification and simulation tasks using Vera CPUs. This tangible acceleration not only enhances NVIDIA’s internal chip development efficiency but also serves as the most credible product demonstration to ecosystem partners and customers, reinforcing confidence in Vera’s capabilities and NVIDIA’s broader AI-accelerated design toolkit. By showcasing Vera’s real-world impact within its own workflows, NVIDIA effectively signals a robust, battle-tested platform that underpins its next-generation CPU and GPU innovations.

Sources
Onpode

Ecosystem Integration Drives AI Leadership

By unifying CPUs, GPUs, networking, and AI software with partners like Siemens and OpenAI, NVIDIA is building a full-stack, secure platform that anchors its dominance across semiconductor, robotics, and enterprise AI markets.

NVIDIA is strategically deepening its ecosystem partnerships with industry leaders like Siemens to embed self-verifying agentic AI workflows within semiconductor and PCB design, significantly enhancing automation and validation processes. By integrating Siemens Fuse EDA with NVIDIA’s NeMo Gym, Nemotron models, and CUDA-X, this collaboration accelerates library characterization by over 10x, exemplifying NVIDIA’s commitment to a tightly integrated hardware-software-AI ecosystem that boosts design quality and speeds engineering outcomes, as highlighted by NVIDIA VP Timothy Costa.

Beyond hardware, NVIDIA is advancing a comprehensive AI infrastructure strategy that transcends GPUs to deliver a full-stack platform encompassing CPUs, networking, and DPUs, all synchronized to support complex engineering workflows. This approach leverages NVIDIA OpenShell for secure, sandboxed execution with strict IP controls, enabling autonomous AI agents to perform multi-step engineering tasks validated against physics-based verification engines, reflecting a shift toward integrated AI platforms tailored for semiconductor design and broader AI applications.

NVIDIA’s ecosystem expansion also targets a diversified data center customer base beyond traditional hyperscalers, now including AI labs, AI-native companies, and enterprises such as Anthropic, OpenAI, Eli Lilly, Samsung, Tesla, and TSMC, which collectively represent about half of its Data Center revenue. This broadening of partnerships underscores NVIDIA’s role as the sole provider of an end-to-end AI platform that integrates training (DGX), simulation (Omniverse + Cosmos), and deployment (Jetson Thor) capabilities, positioning it uniquely across industrial automation, robotics, and autonomous vehicle markets.

Sources

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.