Nvidia's Vera CPU and the Olympus Cores Powering It
The AI giant challenges Intel and AMD's dominance with a groundbreaking new processor.
1Nvidia's Entry into the CPU Market
Nvidia has made a bold move to challenge Intel and AMD's dominance in the CPU market with its new Vera processor. The AI giant aims to offer standalone CPUs to major cloud providers. Companies like Alibaba, ByteDance, and Meta have already announced plans to adopt these chips. Vera succeeds Nvidia's Grace CPU, featuring 88 custom Armv9.2 cores and representing a significant leap in the company's processor technology.
2Vera's Technical Specifications
The Vera CPU boasts 88 custom Armv9.2 cores and 176 threads, supporting up to 1.5TB of LPDDR5X memory. Designed to operate independently from Nvidia's GPUs, the processor's architecture is optimized for AI workloads. Vera specifically targets AI agents and GPU management tasks, aiming to eliminate pipeline and execution bottlenecks that commonly affect performance in these areas.
3Monolithic Compute Architecture
Vera features a monolithic compute die design, with all 88 cores integrated into a single silicon piece manufactured using TSMC's 3nm process. Nvidia claims this architecture provides superior core-to-core bandwidth and reduced latency compared to competing designs. Surrounding the compute die are various chiplets handling memory controllers and I/O functions, creating a balanced and efficient package.
4Vera CPU Superchip Configuration
Nvidia introduced the Vera CPU Superchip for dual-socket configurations, connecting two Vera processors via NVLink-C2C with 1.8TB/s bidirectional bandwidth. This setup delivers 176 cores and 352 threads, supported by 16 LPDDR5X memory modules. The Superchip offers approximately double the memory bandwidth of AMD's 2024 Turin Epycs processors, demonstrating Nvidia's commitment to high-performance computing solutions.
5Design for AI Agent Workloads
Nvidia revealed a reference design capable of housing 256 Vera processors in a single liquid-cooled rack, totaling 22,528 cores and 384TB of memory. Optimized specifically for AI agents, this system differs from traditional large language models by targeting workloads that don't run on GPUs. Vera's architecture represents a specialized approach to handling these emerging computational requirements.
6Origins of the Olympus Cores
Previously relying on Arm's off-the-shelf cores like Neoverse V2 and Cortex X925 for its Grace CPUs, Nvidia has now developed its own ARMv9.2-compatible Olympus cores for Vera. However, the degree of customization remains debatable, as architectural diagrams suggest significant reliance on existing Arm IP. This transition marks an important step in Nvidia's processor development strategy.
7Olympus Core Architecture Details
The Olympus cores feature a 10-wide decoder, eight integer ALUs, six vector/floating-point pipelines (SVE 128), four load units, and two store units. While offering broader front and back ends compared to AMD's Zen 5 or Intel's Redwood Cove cores, they don't represent a radical departure from standard Arm core designs. The architecture focuses on balanced performance across various workloads rather than breaking specific performance records.
8Custom Neural Branch Predictor
Nvidia claims its custom neural branch predictor can explore two code paths simultaneously each cycle, a significant advancement in branch prediction technology. This innovation reduces misprediction rates, potentially boosting overall performance. Branch prediction remains crucial in modern CPUs, and Nvidia's approach specifically targets the complex, branch-heavy code typical in AI-generated scripts and other advanced workloads.
9Optimizations for AI Workloads
Nvidia has optimized Vera particularly for branch-intensive workloads like Python scripts generated by AI coding assistants. The company asserts its neural branch predictor achieves higher accuracy rates than traditional history-based methods. Additional mid-core innovations aim to reduce pipeline stalls, with the overall design focusing on maximizing instructions per clock (IPC) for AI-specific processing requirements.
10Memory and Value Prediction Innovations
Nvidia introduced memory renaming and value prediction schemes to enhance performance. Memory renaming allows dependent instructions to execute before load operations complete when data relationships can be determined. The value prediction scheme identifies stable dependency chains to predict future values speculatively. These features particularly benefit pointer-heavy software, graph traversal algorithms, and complex object-oriented workloads common in AI applications.
11Industry Impact and Implications
Nvidia's Vera CPU could mark a turning point in the AI processor market, challenging Intel and AMD's traditional CPU dominance. This move may establish new standards for cloud computing and AI infrastructure. With its optimization for AI agents, Vera could strengthen Nvidia's leadership in AI hardware. The processor's success will demonstrate Nvidia's capabilities beyond GPUs, potentially reshaping competitive dynamics in the semiconductor industry.
12The Future of CPU Design
Vera represents one of the first CPUs specifically designed for the AI era, moving beyond traditional processor paradigms. This workload-specific approach may become increasingly common as AI agents and complex object-oriented workloads grow. Nvidia's innovations could set new industry standards, positioning Vera not just as another processor but as a harbinger of a new computing paradigm tailored for emerging AI-driven applications.
Original article: www.theregister.com - Articles — Nvidia'nın Vera CPU'su ve Olympus Çekirdekleri