» Simple Machines Wiki Page

SimpleMachines Inc. was a US-based semiconductor company that developed AI accelerator chips. It was one among the 100 or so chip startups formed around 2017-2020 to go after AI problems with semiconductor innovation (see AI Accelerator Survey and Trends). It's key technology was to develop a processor architecture called Reuse Exposed Dataflow that combined data reuse and dataflow at the Instruction Set Architecture level. Its technology allowed software maturity and high performance. The company was one of the first to observe the importance of software maturity, which they define as the ability of the hardware to run the deeply entrenched software stack (sse Composable and Modular Code Generation in MLIR: A Structured and Retargetable Approach to Tensor Compiler Construction) common in AI.

SimpleMachines was founded in 2017 based on technology developed at UW-Madison in Karu Sankaralingam's Vertical Research group. The founding team included Karu Sankaralingam (Founder and CEO), Jeff Thomas, Vinay Gangadhar, Preyash Shah, and Vijay Thiruvengadam. The company raised around $25 million dollars of VC funding.

It taped out its first chip Mozart in March 2020, manufactured on TSMC's 16nm technology. The chip included one ASIC die and two HBM dies packaged in a CoWos process. The software friendliness of the chip architecture allowed running of complete AI applications from the MLPerf benchmark suite and other applications. At the time, one of the few chips outside of NVIDIA's GPUs and Google TPUs to have the ability to run entire applications (see Growing AI Diversity and Complexity Demands Flexible Data-Center Accelerators).

The technology was featured in several industry forums including HOTCHIPS 2021, ISCA 2022, Microprocessor Report, and the Linley Guide to Processors for Deep Learning 2021. As CEO, Karu spoke at Mentor's U2U Fireside chat about the technology in the scope of other semiconductor trends. The company was also covered in the several other media reports: Wisconsin State Journal, November 2020, Forbes describing the company's innovative approach to speedup up AI processing.

The company ceased operations in early 2021 because the technology did not develop enough business traction. The market leaders NVIDIA and Google TPU were able to develop different technologies that made SimpleMachines' superior performance on small batch size less attractive. The Mozart chip implemented several academic ideas including: dataflow, reused exposed dataflow, streams to access data, and the concept of podcast and prethrow that allowed combining of multiple tiles to work on one unit of data. These ideas are explained in detail in the ISCA 2022 publication. More information on the company is on their archived webpage.

Other Media Coverage:


Page last modified on May 24, 2023