Essential Knowledge Systems, LLC Information

The capabilities of recent AI models are truly astounding and transformative. Everyone who wants a piece of this AI pie is in for another equally astounding fact: the cost of AI. Even if you want to just put together a single server with consumer grade GPUs, you are looking at over 100K in US Dollars. Plus, you need to call your utility to upgrade your power supply lines to 240 volts or higher. The cost of cloud GPUs are even worse. If you want a decent (not top of the line), dedicated GPU server, you are looking at 50K to 100K in US Dollars per month. Yet you feel you can’t lag behind the others in your market segment and you are not alone. The projected cumulative effect of all the demands for AI, at the national scale, in the foreseeable future is staggering and unprecedented. Over the past few years companies and researchers have brought to the market many innovations in hardware and AI algorithm designs to bring down the power requirements of AI. Innovations such as data and model sharding, tensor parallelism, lower precision arithmetic, etc., enable the training of much larger models than otherwise possible on a given set of GPUs. Model-specific techniques such as flash attention bring additional boost in performance. However, all these gains are constrained by an empirical scaling law which says larger models learn better than smaller models. That is where EKS technologies become relevant to bring down your total cost of compute. We offer flexible and favorable PoC options, including money-back guarantees. Contact us right away!