Until this week, Cerebras remained a stealth mode startup backed by premier venture capitalists. This week Cerebras revealed its first product, a new computer chip built on Cerebras Wafer Scale Engine (WSE) optimized entirely for deep learning. Cerebras mentioned that it is the world’s largest (46,225 square millimeters) chip containing more than 1.2 Trillion transistors and computation.
Cerebras’ chip blows Nvidia’s comparable chip out of the water. Check out the mind-blowing numbers below.
• 46,225 mm2 chip: 56x larger than the biggest GPU ever made
• 400,000 cores: 78x more cores
• 18 GB on-chip SRAM: 3000x more on-chip memory
• 100 Pb/s interconnect: 33,000x more bandwidth
All the AI-optimized cores on this chip are connected entirely on silicon by the Swarm fabric in a 2D mesh enabling huge bandwidth and low latency at a fraction of the power when compared to traditional techniques used to cluster graphics processing units. In addition to the chip hardware, Cerebras software stack:
• Allows users to program the WSE using familiar, high-level ML frameworks
• Automatically compiles ML framework-defined models into WSE executables
• Maximizes utilization of all 400,000 cores and takes advantage of the WSE’s unique scale
You can learn more about Cerebras
here.
source
