Cerebras launched the WSE-3 AI chip which they say is the fastest AI inference chip in the world. What's more interesting about it is that it is also the largest computer chip ever produced in terms of size. 215 x 215 mm. A processor chip is usually cut from a silicon wafer. WSE-3 uses the entire silicon wafer to produce a 5nm technology chip with 4 trillion transistors, 900,000 cores and a built-in memory of 44GB SRAM.
In comparison, the WSE-3 has 52 times more processor cores than the NVIDIA H100. This gives it a peak processing speed capability of 125 petaFLOPs. Cerebras added that, on average, Cerebras CS-3 computers using WSE-3 are capable of performing AI inference processing up to 20 times faster than competing systems.
For example it can process 1,800 tokens per second for the Llama 3.1 8B model and 450 tokens per second for the Llama 3.1 70B model. Other chips on the market are only capable of 300 tokens per second. Therefore Cerebras says the difference in processing speed is like changing from dial-up internet to broadband.
The ability to process tokens faster allows AI chatbots and even generative AI to generate responses and images faster than now. Anyone who has ever used an AI service will notice that the processing time for commands given to the AI at this point is quite slow.
Cerebras now offers their cloud AI services through three tiers which are free with API limits, Developer which offers access to Llama 3.1 8B and 70B and finally Enterprise with a model tuned for specific customer needs.