Cerebras Systems has unveiled the largest AI chip based on the 7nm process node, the Wafer Scale Engine 2. Succeding the first generation WSE, the WSE2 is a singular monolithic chip that features a range of new features including the largest transistor count & most number of cores ever featured on a chip.
Cerebras Wafer Scale Engine Is A Massive AI Chip Featuring 2.6 Trillion Transistors & Nearly 1 Million Cores
The Cerberas Wafer Scale Engine 2 has two parts, the first is the chip itself and the second is the CS-2 system which it will actually be featured within. We will with the chip itself which has insane numbers to go by. For starters, the WSE-2 is the largest AI chip ever made. It is based on TSMC’s 7nm process node and has a 100% yield rate which is quite impressive. For one, if you are using the entire wafer as a die, you are either going to get 100% yield if the design can absorb defects or 0% if it cannot.
Coming to the technical details, the Cerebras WSE-2 measures at 46,225mm2 which is 56 times larger than the largest GPU ever made, the NVIDIA Ampere A100. The chip features almost one million cores (850,000 to be precise), a 123x increase over the Ampere A100, 40 GB of on-chip memory (SRAM) vs 40 MB on the Ampere A100, 20 Petabytes per second of memory bandwidth and 220 Petabits per second interconnect fabric (2D Mesh) bandwidth.
Cerebras Wafer Scale Engine 2 Specs
As for the cores on the Cerebras WSE-2, they are AI-optimized Sparse Linear Algebra Compute (SLAC) cores and are fine-tuned towards AI processing. Compared to the previous generation (Cerebras WSE-1), Cerebras’s WSE-2 offers a 2.16x increase in transistor density and features more than twice the number of cores.
Moving over to the Cerebras WSE-2 powered CS-2 system, we are looking at a system that is 26" tall and composed of 15 rack units (15U). The system features 12 100GbE lanes, a custom water-cooled solution through a closed-loop setup, and twelve standard PSUs in a 9+3 redundant configuration that delivers up to 23,000 Watts of power to the CS-2 system. Some words from Cerebras on their cooling solution:
Cerebras
“TSMC has long partnered with the industry innovators to manufacture advanced processors with leading performance. We are pleased with the result of our continuous collaboration with Cerebras Systems in manufacturing the Cerebras WSE-2 on our 7nm process, another extraordinary accomplishment and milestone for wafer scale development after the introduction of the Cerebras 16nm WSE less than two years ago.”
Sajiv Dalal, SVP of Business Management, TSMC North America
Currently, there’s no word on pricing for either the Cerberas CS-2 or the WSE-2 but the first generation Wafer Scale Engine had a price tag of around $2 million US.
The top right of the system is for the movement of water. Two hot-swappable pumps move water through a manifold across the back of the WSE-2, cooling the wafer and warming the water. Warm water is then pumped into a heat exchanger. This heat exchanger presents a large surface area for the cold air blown in by the four hot-swappable fans at the bottom of the CS-2. These fans move air from the cold aisle, cool the warm water via the heat exchanger and exhaust the warm air into the warm aisle.
