Cerebras CS-4 rack systems take AI performance to the next level
Next-gen AI accelerators push performance to new heights.
1Cerebras' next-gen AI accelerators
Cerebras unveiled its next-gen Wafer Scale Engine (WSE) and Nexus rack systems, aiming to boost AI performance tenfold in throughput per watt compared to the previous generation. The new WSE-3T chip offers 21.6 petabytes per second of memory bandwidth and 250 petaFLOPS of AI compute power, making Cerebras 1,000x faster than its competitors.
2Turbo technology boosts performance
The WSE-3T leverages Turbo technology to maximize every hertz of the chip's potential, doubling the clock speed from 1.4 GHz to 2.8 GHz. With 44 GB of SRAM and 2.4 Tbps off-die connectivity, it provides a significant advantage in AI computations. These innovations position Cerebras as a leader in AI accelerators.
3Collaboration with AWS and AMD
Cerebras has partnered with Amazon Web Services (AWS) and AMD to accelerate AI inference tasks. This collaboration offloads compute-intensive prompt processing to AWS's Trainium XPUs and AMD's Instinct GPUs, allowing Cerebras chips to function primarily as decode accelerators. This strategy enhances Cerebras' efficiency in AI inference.
4Modular rack system design
Cerebras' CS-4 rack systems feature a modular design, separating compute, power delivery, and cabling for easier deployment, maintenance, and upgrades. Each CS-4 rack can be equipped with three 'backpack' form factor chips. This design optimizes the system's power consumption and scalability.
5Power consumption and cooling
The CS-4 rack systems are notable for their high power consumption. Each backpack consumes approximately 46 kW, with total system power ranging between 120 kW and 140 kW. These figures are significant even for liquid-cooled systems. However, Cerebras' efficient power delivery allows the chip to operate at higher frequencies.
6SRAM capacity and future expectations
Cerebras chose to double performance this generation rather than increase SRAM capacity, which hasn't seen a meaningful boost since the WSE-2. However, the next-gen WSE-4 is expected to double SRAM capacity while offering modest FP16 performance gains. This shift signals a new era for Cerebras' AI accelerators.
7Competitor comparison
Cerebras' WSE-3T chip offers higher memory bandwidth compared to AMD and Nvidia's latest GPUs. However, these performance figures may be theoretical. In AI inference, Cerebras chips may not fully utilize SRAM, potentially limiting their advantage over competitors. Still, Cerebras excels in decode operations.
8The future of AI accelerators
Cerebras' CS-4 rack systems provide key insights into the future of AI accelerators. Modular design, high performance, and efficient power delivery set new standards in AI computing. Collaborations with AWS and AMD further strengthen Cerebras' position in AI inference. These developments will shape the future of AI technologies.