Introducing the third generation of its Wafer Scale Engine, Cerebras has unveiled a chip manufactured on a 5nm node from TSMC. The new Wafer Scale Engine 3 boasts an impressive 4 trillion transistors and 900000 cores specially optimized for AI computations.
According to Cerebras' press release, the Wafer Scale Engine 3 is designed with 900000 cores tailored for AI operations. With the capability to deliver up to 125 petaflops of computing power, this latest chip surpasses its predecessor, the WSE-2, which could deliver up to 62.5 petaflops and was built on a 7nm process.
Notably, the new WSE-3 chip boasts a surface area of 46,225mm², as reported by the company on social media. The chip retains the same physical size as its predecessors and houses 44GB SRAM, translating to 48kB per core. With a memory bandwidth of 21PB/s and interconnect bandwidth of 214Pbit/s, the WSE-3 promises exceptional performance across the board.
Despite doubling in power compared to its predecessor, Cerebras asserts that the energy consumption of the Wafer Scale Engine 3 remains consistent. Reports indicate that the previous iteration, the WSE-2, consumed approximately 15kW. The WSE-3 is targeted towards businesses and organizations requiring hyperscale computing capabilities, such as cloud providers. These chips are integrated into Cerebras' CS-3 'supercomputers', enabling the linking of 2048 systems to create over 1.8 billion cores and achieve a compute power of up to 256 exaflops. The third-generation Wafer Scale Engine succeeds the WSE-2" from 2021, which was composed of 2.6 trillion transistors and 850000 cores optimized for AI tasks, alongside 40GB of SRAM.
English (US)