For Just A Few Million Dollars And 20kW Of Power
If NVIDIA’s offerings just aren’t helping you build your own personal SkyNet quickly enough you do have another choice. Cerebras Systems, which was started by SeaMicro founder Andrew Feldman is offering a CS-1 system which will allow you to train a neural network in minutes as opposed to months, if you have the cash to buy one and keep it powered. That time saved could well make this a very lucrative investment.
If you are wondering just how a chip can be this much better than a GPU, the answer is because Cerebras does things very differently. For one, they have the largest wafers going, and they don’t cut them up like everyone else, instead that wafer is the CPU. The wafer still consists of numerous chips, interconnected in way to provide incredibly low latency communication between them as well as almost obscene memory bandwidth, not to mention 1.2 trillion transistors.
They also contrasted the performance of one of their CS-1 systems against number 82 of the current top 500 supercomputers, claiming their system was 200 times faster than the Joule Supercomputer. Even if you aren’t particularly interested in machine learning, it is worth clicking through the links at Slashdot to learn more about this unique architecture.