This gigantic processor offers twice the performance of its predecessor, the WSE-2, it has 4 billion transistors manufactured in a 5nm processincludes 44 GB of onboard SRAM, offers 21 petabyte/s bandwidth, and is specifically designed for training AI models.
62 NVIDIA H100 GPUs in a single processor
Indeed, during the official presentation of this new processor, the company wanted to compare it with the most powerful NVIDIA GPU currently available on the market, the H100. The performance of this processor is equivalent to 62 NVIDIA H100 GPUs.
A very important section to which Cerebras has paid special attention is the energy consumption
In terms of AI model training capabilities, the WSE-3 is capable of working with up to 24 trillion parameters and supports a maximum of 1.2 Petabytes of memory in order to store large models in a logical unit without using partitions, which helps make the process much faster.
The scalability of this new processor allows configuration of up to 2048 CS-3 systems, enabling the creation of gigantic supercomputers capable of performing AI tasks in much less time than servers equipped with GPUs currently sold. Nvidia
With the next generation already in the oven, the H-200 This should represent a significant change compared to the first, although at the moment all the details of the new generation are still unknown.
In addition to all the performance improvements brought by this new processor, where innovation has also been made, there is the amount of code required to train LLM, reducing up to 97% compared to the NVIDIA H-100 GPU
This new processor, focused on large companies with cloud developments, is also aimed at medical research through a collaborative project between Cerebras and the Mayo Clinic and the Argonne National Laboratory in the United States.
The thing doesn’t stop there
This is not the only project Cerebras is working on. Through the collaboration with G42, they are working on building the Condor Galaxy 3, a much more powerful computer than the WSE-3, since it will feature up to 64 CS-3 systems, which means 57,600,000 cores. When the Condor Galaxy 3 becomes available, the combined computing power of these three supercomputers will be 16 ExaFLOP as they claim, the CTO of the G42.
This will be their third collaboration between the two companies, following the Condor Galaxy 1 and Condor Galaxy 2 models currently located in California and which offer a combined performance of 8 ExaFLOPs, leaving the new WSE-E far behind.