Cerebras Systems, a semiconductor company based in California, announced a major news. Its third generation of Wafer Scale AI acceleration chip "WSE-3" (Wafer Scale Engine 3) is officially available.

According to the report, this chip is known as "crazy specifications" in the industry. It not only maintains a stable advantage in power consumption and price, but also pushes the performance to a new height, doubling the performance under the premise of the same power consumption and price.

It was revealed that the WSE-3 process was upgraded to TSMC's 5nm process, and although the area of the chip has not yet been announced, it is widely expected to remain at a similar level.

However, it is worth mentioning that the number of transistors has increased to a staggering 4 trillion. At the same time, the number of AI cores has further increased to 900,000, the cache capacity has reached 44GB, and the external memory capacity can be 1.5TB, 12TB, and 1200TB to meet the needs of different scale applications.

Despite the modest increase in the number of cores and cache size, WSE-3 has seen a leap in performance, with a peak AI power of 125PFlops, or 12.5 peflops per second, which puts it on par with the top supercomputers.

According to Cerebras Systems, WSE-3 was able to train Llama with 70 billion parameters in just one day. This result will undoubtedly bring broader possibilities for the research and application of artificial intelligence, and provide powerful computing power support for scientific computing, data analysis, image processing and other fields.