I lead a research group that studies accelerator design, multi-core processors, and specialized co-processors and the design of tools that make them easier to program and design. My research style focuses on the construction of real prototypes, in FinFET nodes.I direct the Bespoke Silicon Group (http://bsg.ai), which authored some of the earliest research on the Dark Silicon problem, including the first research paper to have the phrase "dark silicon" in the title. I also authored the first research paper on bitcoin mining chips. In 2002, my team and I designed the first scalable general-purpose multicore processor, MIT Raw, which had 16 general purpose cores on a tiled on-chip network, an approach that has since been adopted by Intel in the Skylake SP series. More recently we developed an open source 2048-core RISC-V processor.My current projects include:- Chiplet Cloud, a way of exponentially reducing the cost of serving Large Language Models in the data center- HammerBlade manycore, a next generation parallel architecture that is more flexible and programmable than a GPU. We have several prototype tapeouts in 12nm FinFET technology, including a 99 mm^2 chip that breaks all world records for RISC-V silicon performance. Funded by DARPA for > $13M.- BlackParrot, an open source, SystemVerilog based RISC-V Linux capable multicore..- ASIC Clouds, a method of scaling out accelerators into the datacenter. Our work predicted the Google TPU before it was announced, and also predicted the deployment of video transcoding ASIC clouds such as the Google VPU.- BaseJump (http://bjump.org), open source hardware infrastructure for accelerating the hardware design process, as well as prototyping infrastructure.- Cortex Suite and the San Diego Vision Benchmark Suite, benchmark suites for computer vision and machine learning. I have an active consulting practice, focused on high performance digital system design.

Worked in Google's video transcoding accelerator team.

I directed the UCSD Center for Dark Silicon (http://darksilicon.org), which authored some of the earliest research on the Dark Silicon problem, including the first research paper to have the phrase "dark silicon" in the title. I also authored the first research paper on bitcoin mining chips. In 2002, my team and I designed the first scalable general-purpose multicore processor, MIT Raw, which had 16 general purpose cores on a tiled on-chip network.I led a research group that studies the design of multi-core processors, accelerators, and specialized co-processors and the design of tools that make them easier to program and design. My research style focuses on the construction of real prototypes, for instance, the 16nm Celerity RISC-V chip, which had 511 RISC-V cores and broke the world record for RISC-V performance by 100X.My projects included:- GreenDroid, a multicore processor design for cell phones that leverages automatically generated energy-saving coprocessors called Conservation Cores to fight Dark Silicon (greendroid.ucsd.edu)- Kremlin, a kind of "gprof for parallelization" i.e. a profiler, that given a serial program and representative inputs, tells you which regions to parallelize (kremlin.ucsd.edu)- The San Diego Vision Benchmark Suite, a benchmark suite that attempts to characterize the exciting area of computer vision. (parallel.ucsd.edu/vision)- The first academic work analyzing Bitcoin mining ASIC systems- ASIC Clouds, a method of scaling out accelerators into the datacenter

MIT Raw Project

Co-author of Virtual PC 1.0.

Microkernel Group.