The two architectures closest to consumers, x86 and ARM, have received multiple optimizations in recent years with the aim of reducing their energy consumption. The path was especially complex for Intel, which is still seeking to expand its presence in the mobile market, but the real benefits in energy savings are more evident in the server world. That's when Rex Computing appears, a company working on an architecture capable of consuming one-twentieth of the energy that Intel products do.
The history of Rex Computing begins with Thomas Sohmers, a 19-year-old who left his studies to found this startup with financial help from Peter Thiel, known for his angel investments and for being one of the founders of PayPal. The interest of Thiel and other elements like our friends at DARPA seems well justified, since the company's plan is to develop a new ultra-efficient architecture, with consumption twenty times lower than its Intel equivalent. Developing architectures from scratch is complicated due to compatibility issues. We all know very well that both ARM and x86 have room to be optimized, but the idea has always been to maintain a reasonable balance. However, high-performance environments operate under other rules, and what benefits a desktop system turns out to be completely unnecessary in a supercomputer.
The key behind the Neo chips
What is the key behind Sohmers' "Neo chips"? The explanation he gives is the following: about 4,200 picojoules are needed to move 64 bits from DRAM to registers, while a double-precision floating-point operation demands only 100 picojoules. In other words, moving data consumes 42 times more energy than working on it, and this is because all the switches, cables, and the cache hierarchy itself eat up the difference. The goal of the Neo chips is to eliminate all that "overhead" and reduce consumption, at the cost of greater complexity on the software side, which will have to absorb several management functions. This implies the existence of development kits, and more time invested in the code.
The numbers
What about the "hard" numbers of the Neo chip? 256 cores at 256 GFLOPs in double precision, and an efficiency of 64 to 128 GFLOPs per watt, that is, between 2 and 4W per chip, depending on the application. I imagine the improvement in efficiency will not extend to all cases, but even if it meets half of what it promises, the high-performance world will surely welcome the new Rex Computing architecture.
