Artificial intelligence company Cerebras came out with the world’s largest processor this year. This time, the company raised the bar to 5nm level. The WSE-3 (Wafer Scale Engine-3) processor, which reached a record level in the number of cores, appeared with 900 thousand artificial intelligence cores and 4 trillion transistors. Cerebras receives the processor from TSMC at the wafer level. Thus, giant processors appear.
The world’s largest processor WSE-3 and its features
Cerebras is a startup that produces processors used in artificial intelligence studies. Previously, the company had the title of the world’s largest processor with its 16 nm WSE-1 and 7 nm WSE-2 processor. Now, the company that developed the same title is here with a much more powerful processor.
Processor to be used for artificial intelligence training, 46225 mm2 such a giant size. The wafer-sized processor thus provides much higher performance and power savings. Since the company produces these chips for supercomputers and artificial intelligence servers, it has more freedom in terms of processor sizes. In this sense, the startup uses the advantage of being able to produce wafer-sized processors.
WSE-3, which has the same power consumption as its predecessor, WSE-2, offers twice the performance. The processor has 900 thousand artificial intelligence cores, 44 GB on-chip SRAM and 125 FP16 PetaFLOPS peak performance.
The world’s largest processor, WSE-3, also powers Cerebras’ supercomputer CS-3. This supercomputer can be used to train artificial intelligence models with up to 24 trillion parameters. According to Cerebras, its CS-3 supercomputer is capable of training AI models 10 times larger than Gemini and ChatGPT.
This supercomputer can support 1.5 TB, 12 TB, or 1.2 PB of external memory. This allows it to store especially huge artificial intelligence models on its own, without partitioning them. In this sense, the CS-3 supercomputer has a significant performance advantage for training artificial intelligence.
Additionally, Cerebras is seeking to renew the record for the world’s largest artificial intelligence supercomputer with G42. Previously, Condor Galaxy-1 and Condor Galaxy-2 supercomputers held this record. The newly developed Condor Galaxy-3 supercomputer will bring together 64 CS-3 systems and have a structure of 57,600,000 cores. Currently, Condor Galaxy-1 and Galaxy-2 supercomputers have a total power of 8 exaFLOPs. The Condor Galaxy-3 supercomputer will have a power of 16 exaFLOPs alone.