The video discusses the current situation in the AI development landscape, where top tech companies like Google, Tesla, and Microsoft are buying GPUs at high prices, leading to a bottleneck in AI development. It introduces a startup called Cerebras, which offers AI chips with better performance than Nvidia GPUs.
Cerebras has recently announced a supercomputer called the Contour Galaxy one, capable of four exaflops of AI compute. This supercomputer is built with 64 wafer scale engine chips, each of which is approximately 56 times larger than an Nvidia 8 100 GPU. The largest chip, called the wafer scale engine 2, can maintain the entire network parameters like weights and activations on a single chip, speeding up the training process.
Cerebras plans to build nine supercomputers in total, with the first one, Condor Galaxy one, already up and running. These supercomputers are designed to be interconnected and function like one big cloud. They are expected to significantly reduce power consumption compared to similar computers, and are equipped with Cerebras' custom compiler which can run PyTorch code as it is on their hardware without any changes.
The video also highlights the partnership between Cerebras and G42, a company operating the largest cloud in the Middle East. G42 plans to buy over 500 wafer scale Engine 2 chips for their AI training needs.
The video concludes by emphasizing the potential of Cerebras in the AI market, citing several publications that report better performance of Cerebras chips compared to Nvidia GPUs.
1. Top tech companies like Google, Tesla, and Microsoft are buying GPUs, but supply is limited, meaning not everyone can get them. This is seen as a bottleneck in the current pace of AI development.
2. There is a company, Siri breast, that offers AI chips with better performance than Nvidia GPUs and is available right now.
3. Siri breast has recently announced a new Contour Galaxy one supercomputer that is capable of four exaflops of AI compute. It is one of the largest and fastest AI supercomputers in the world.
4. The Contour Galaxy one is built of 64 wafer scale engine chips, which are approximately 56 times larger than an Nvidia 8 100 GPU.
5. The Contour Galaxy one consumes about 1.75 megawatts of power, which is less than half of the power draw of the h100.
6. Cerebrus plans to build nine supercomputers, including the Condor Galaxy one, and they will be interconnected, functioning like one big cloud.
7. The first supercomputer is called Condor Galaxy one and is based in Santa Clara, California. It is a four exaflop supercomputer.
8. The Condor Galaxy one and the other eight supercomputers will be linked together into a constellation, creating a 36 exaflop AI compute constellation, the largest in the world.
9. The Condor Galaxy one uses direct fiber for data transfer and runs strictly data parallel due to its wafer scale engine two.
10. Cerebrus has developed a custom compiler that can run PyTorch code written for a GPU on their hardware without any changes.
11. A single wafer scale 2 engine chip costs about 1.2 to 1.7 million dollars, so a Condor Galaxy supercomputer will cost roughly 100 million dollars.
12. The main customer for this supercomputer is g42, which operates the largest cloud in the Middle East. They plan to buy over 500 wafer scale Engine 2 chips.
13. g42 is using machines to train Arabic language chat. They have a huge database of medical data and plan to train medical AI based on it.
14. Cerebrus is already working on the next generation of their chip, wafer scale Engine 3, which will be taped out by TSMC in a five-nanometer process node.
15. Cerebrus is considered a solid alternative to Nvidia GPUs. Published research shows that Cerebral's wafer scale Engine 2 outperforms Nvidia GPUs.