Chip designer Graphcore unveils new AI processor on 7nm process
British chip designer Graphcore has unveiled a new processor. The Colossus MK2 is intended for running AI workloads. It has 1472 cores and 8832 parallel threads. The company will start deliveries on a large scale in the fourth quarter of 2020.
The Colossus MK2, also referred to by the company as GC200, is a so-called ipu that is used for performing AI calculations. The chip is equipped with 59.4 billion transistors and is made on a 7nm process from TSMC. The GC200 thus has 1472 cores and 8832 parallel threads. The ipu also has 900MB sram in the processor itself. The whole is housed in a building of 823mm².
The Graphcore GC200 chip (left) and the M2000 system
Graphcore also supports extra ‘exchange memory’ for the memory. This allows each ipu to be provided with a maximum of 448GB of memory with a maximum theoretical bandwidth of 180TB / s. Graphcore further claims that the GC200 performs about eight times better than its predecessor, the Colossus MK1. The latter was announced in 2018 and was made on a 16nm process.
Graphcore reports that the new ipu comes in an M2000 system. This M2000 contains four Colossus MK2 chips and has a 1U housing. That system must be good for a petaflops of computing power. Users can link up to 64,000 of these ipus together for a total computing power of up to sixteen exaflops. The company is also supplying a rack-scale ‘IPU-POD64’, which consists of 16 M2000 machines for a total of 64 Colossus MK2 chips. These Pod systems are placed in a standard 19 “server rack.
For communication between the different iPUs, each M2000 system contains a GC4000 gateway chip. The company calls this ‘ipu-fabric’, which functions as an interconnect between different systems, with a bandwidth of 2.8Tbit / s per M2000 system. According to Graphcore, this can be scaled up linearly to ‘several petabits per second’ when using many ipus.
The system uses standard copper or optical osfp connectors for ipu fabric. M2000 systems can be individually connected with each other. Larger configurations use tunneling over Ethernet. The use of standard qsfp interconnects and 100Gbit / s Ethernet switches is also possible.
Companies can pre-order the M2000 systems and IPU-POD64, Graphcore reports. Large-scale deliveries will start in the fourth quarter of 2020. The company does not disclose prices.
The design of ipu fabric