Cerebras presents 7nm Wafer Scale chip with 850,000 cores
Cerebras has presented the second version of its Wafer Scale Engine. It is a chip that occupies an entire 7nm wafer from TSMC. The chip has 2.6 trillion transistors and 850,000 simple cores for AI calculations.
As with the first Wafer Scale Engine, which was made at 16nm, each core is equipped with 18KB of sram. This means that the new chip contains a total of 40 GB of SRAM. The memory bandwidth and interconnect bandwidth is also much higher due to the increased number of cores.
Because the yields of TSMC’s 7nm process are so good, Cerebras assumes a yield of 100 percent, writes AnandTech. The chip is designed in such a way that the functioning of any non-functional cores can be absorbed by the surrounding cores.
Cerebras already announced last year that it would come up with a successor to its Wafer Scale Engine, but not all details were known at that time. The first version was presented in 2019 and was made at 16nm. 400,000 cores fit on the same chip surface.
According to the makers, the first version has been used at various supercomputer labs and there is interest in the new model from the same angle. Cerebras also makes a complete system that contains the chip. For the new chip, this is the CS-2 computer, with twelve 100Gbit/s Ethernet connections. That system will be available in the third quarter of this year and will cost “several million dollars.” The first version was quoted as a price of 2 to 3 million dollars.
chip | Cerebras Wafer Scale Engine (2nd Generation) | Cerebras Wafer Scale Engine (1st Generation) |
Process | 7nm, TSMC | 16nm, TSMC |
Format | Whole wafer (300mm) 21.5×21.5cm (462.3cm²) 46.225 mm2 |
Whole wafer (300mm) 21.5×21.5cm (462.3cm²) 46.225 mm2 |
Transistors | 2.6 trillion | 1.2 trillion |
Number of cores | 850,000 | 400,000 |
sram | 48KB per core, 40GB total | 48KB per core, 18GB total |
Memory bandwidth | 20PB/s | 9PB/s |
Interconnect bandwidth | 220Pb/s | 100Pb/s |
Consumption (chip/system) | 15kW / 20kW | 15kW / 20kW |