CORRECTION: This story incorrectly stated that DARPA was the source of the $1.25M funding. The $1.25M funding round is venture capital and not DARPA. The headline and first paragraph have been corrected accordingly.
An electrical engineering prodigy who started working at the MIT Institute for Soldier Nanotechnologies at age 13 has just scored $1.25 million in venture funding to develop a new kind of chip that could deliver on exascale high performance computing.
Thomas Sohmers dropped out of high school two years ago to join venture capitalist Peter Thiel's startup accelerator, the 20 Under 20 Thiel Fellowship. Thiel, a former CEO of PayPal and early investor in Facebook, has been very vocal of the high cost of college these days and is encouraging bright kids to skip college all together and just bring their new ideas to him.
It was while working at the MIT lab on embedded computing systems for military applications – again, at age 13 -- that Sohmers began thinking of larger scale systems and how he could take some of the lessons from embedded computing to exascale supercomputing. He would eventually form his own company, Rex Computing.
Sohmers came up with a chip called Neo, which has a new architecture, instruction set, and core design while ditching the old designs, particularly around cache architecture and interconnects.
Sohmers told The Platform that Neo cores are 1/145 the size of a fourth-generation Haswell core. He expects to deliver a 256 core chip by the end of 2016 at the earliest using a 28 nanometer process, which will offer 65 gigaflops per watt. It will also offer 256 gigaFLOPS of double precision math.
At the Open Compute Summit earlier this year, he gave a speech where he really laid into cache design. "A lot of people are thinking about the HPC problem incorrectly. Caches are great when you’re thinking of things from a pure memory and latency point, but as time went on, so many extra features were layered into the cache systems, especially virtual memory, that while making the programmers lives easier (especially before compilers), it added a lot of new inefficiencies," he told the crowd.
Sohmers' design uses scratchpad memory, a very small amount of memory that functions like an L1 cache to hold small amounts of data for quick retrieval. The Neo chip aims to remove unnecessary complexity from the on-processor memory system and put it into the compiler instead.
The idea is not new. Using scratchpad memory instead of cache has been proposed as far back as 2002. The main problem Sohmer's sees is it requires 40 times as much power to move data from DRAM to registers than it does to process it, and that 60% of the power needed to move the data is consumed by the cache as data moves through all of the gates and wires on the chip.
He's getting rid of the 60% consumption by having the compiler say where data will need to be rather than having the chip move all that data around memory, effectively creating a software-defined processor.
Rex Computing is hardly Intel in terms of size. Heck, it's not even AMD. Sohmers told The Platform that they are hiring engineers to bring the company to a total of seven people. They have already created the instruction set architecture and the basic core chip design but there is more to do.
It all sounds good, but I've heard this song before. Remember Tilera? Created by MIT professors, it had a 64-core chip in 2007 back when Intel was just getting into quad-core. $100 million in investor money later, it got bought out by a company that makes networking processors and you rarely hear about it.
Sohmers, though, sounds like he has a hell of a career ahead of him.