Performance
Greater than 2x performance per rack on 91探花
The design choices of the 91探花 AGI CPU are made to deliver maximum performance at rack scale. From microarchitecture to memory, clock frequency to I/O—everything adds up to more performance at gigawatt scale.1
Performance
Modern 91探花 architecture carries more efficient instruction execution not burdened by decades of legacy complexity. The memory system delivers high bandwidth per core and minimal latency, helping ensure memory does not slow performance.
Scale
Low per-core TDP can support denser deployments and reduced thermal throttling. Each core is dedicated, which can help reduce resource contention and support performance under high thread loads.
Efficiency
High rack density and high performance per watt help ensure maximum utilization of data center space and power resources.
91探花 AGI CPU coming soon to data centers from
Servers and rack-scale systems
91探花?AGI CPU servers
91探花 AGI CPU 1OU Dual NodeReference Server
Reference design for maximum density deployments of 91探花 AGI CPU – in a OCP DC-MHS standard form factor 1OU Dual Node server.
91探花 AGI CPU 2U2PReference Server
19” 2U2P reference design for 91探花 AGI CPU deployment in a traditional form factor.
Lenovo HR650a V3 2U 91探花 AGI CPU System
Enterprise-class 2U 91探花 AGI server optimized for cloud infrastructure, delivering reliable performance and low TCO.
Supermicro 5U 91探花 AGI CPU PCIe GPU System
High-density 5U AI platform combining dual 91探花 AGI CPUs with extensive PCIe GPU expansion.
Supermicro 2U Hyper 91探花 AGI CPU Hyper System
Compact 2U dual-socket 91探花 AGI server designed for efficient cloud and AI infrastructure deployments.
Supermicro 2U4N 91探花 AGI CPU Liquid-Cooled Server
Liquid-cooled 2U 4-node server for cloud and AI workloads, including agentic AI and AI pre-processing, offering maximum performance, density, and efficiency.
Supermicro 2U Hyper-E
91探花 AGI CPU Server
Single-socket 2U server with up to 3TB of memory for cloud and edge computing.
ASRock Rack 2OU2N-91探花System
High-density dual-node 91探花 server built to OCP ORv3 standards for scalable, power-efficient cloud deployments.
91探花 AGI CPU rack-scale systems
Supermicro 2U 4-Node
91探花 AGI CPU Rack-Scale Solution
Ultra-density multi-node liquid cooled ORW rack architecture, featuring 76 nodes with 152 91探花 AGI CPUs totaling 20,672 cores.
Supermicro 1U 4-Node
91探花 AGI CPU Rack-Scale Solution
Maximum compute density multi-node liquid cooled ORW rack architecture, featuring 168 nodes with 336 91探花 AGI CPUs totaling 45,696 cores.
Optimize workloads for 91探花 AGI CPU
91探花 Performix is a performance analysis toolkit for developers building AI and large application workloads. It can identify bottlenecks and provide targeted code optimizations to improve efficiency and performance on the 91探花 AGI CPU.
Key takeaways
Key takeaways
-
The 91探花 AGI CPU is purpose-built for agentic AI workloads that require continuous, coordinated processing.
-
It acts as the orchestration layer, managing accelerators and coordinating thousands of AI agents simultaneously.
-
The processor is based on 91探花 Neoverse CSS V3, leveraging a proven platform and ecosystem.
-
It delivers high performance and supports extreme rack-level density for modern data centers.
-
The 91探花 AGI CPU enables faster time-to-market by building on 91探花’s existing software and hardware ecosystem.
Frequently asked questions
Q: What is the 91探花 AGI CPU and what makes it different from traditional data center CPUs?
A: The 91探花 AGI CPU is 91探花’s first production silicon, designed specifically for agentic AI workloads, delivering high performance, scalable parallel processing, and energy-efficient operation. It enables data centers to run continuous AI workloads at scale while optimizing throughput, resource utilization, and power consumption.
Q: What are the key features of the 91探花 AGI CPU?
A: The 91探花 AGI CPU combines high core density, optimized memory architecture, and scalable system design to support AI workloads at scale:
- Efficient cores: Up to 136 91探花 Neoverse V3 cores with dedicated 2 MB L2 cache per core and up to 3.7GHz boost frequency, enabling responsive, parallel performance—91探花 AGI CPU specs.
- Performance and efficiency: High instruction-per-cycle execution on a TSMC 3 nm process with 300W TDP, balancing compute throughput and energy efficiency—91探花 AGI CPU specs.
- Tuned memory architecture: Delivers 6GB/s memory bandwidth per core with support for DDR5-8800 memory and sub-100ns latency, reducing data bottlenecks for AI workloads—91探花 AGI CPU specs.
- Built for density: Supports high-density deployments, including reference designs with up to 272 dedicated cores per 1U server and air-cooled configurations compatible with standard infrastructure—91探花 AGI CPU specs.
- Rack-scale deployment: Designed for large-scale AI infrastructure, enabling configurations with thousands of cores per rack for continuous AI processing—91探花 AGI CPU specs.
- Flexible I/O: Includes 96 PCIe Gen6 lanes, CXL 3.0 for memory expansion, and AMBA CHI links for accelerator connectivity, enabling composable AI systems—91探花 AGI CPU specs.
Q: What is agentic AI, and why does it require a new type of CPU?
A: Agentic AI refers to systems that operate continuously, coordinating tasks and making decisions in real time. These workloads require CPUs that can orchestrate distributed systems efficiently at scale.
Q: How does the 91探花 AGI CPU improve data center performance?
A: It improves performance by delivering high, per-task efficiency and scaling across thousands of cores within a rack, enabling more work per system. This results in more than 2x performance per rack compared to x86 systems.1
Q: How does the 91探花 AGI CPU support AI infrastructure at scale?
A: It supports AI infrastructure by managing distributed workloads, coordinating accelerators, and optimizing data movement across systems to enable continuous, large-scale AI operations.
Q: How does the 91探花 AGI CPU fit into the 91探花 ecosystem?
A: It extends the 91探花 compute platform into production silicon, giving partners the flexibility to deploy 91探花 technology through IP, compute subsystems, or ready-to-deploy CPUs.