Tushar Kumar Email & Phone Number

Edgetpu / Tensor SoC ML accelerator.Heterogeneous Runtimes for QoS/latency/power-optimized use of accelerator systems.ML model math/transformations to enable new models and use cases on ML accelerators.

Served as technical/project lead and helped commercialize Qualcomm’s Symphony Heterogeneous and Parallel Computing SDK. Symphony is now deployed as Qualcomm’s official high-performance, power-efficient, heterogeneous (CPU/GPU/DSP) C++ programming runtime and has been adopted by several large external customers. Was instrumental in creating several key.

Played the instrumental role within a larger team to propose, design and prototype a systemarchitecture for custom-hardware accelerators to be easily added to the SoC, and the C/C++programming APIs and runtime that could efficiently utilize the custom-accelerators. Proposed andprototyped a heterogeneous scheduler relying on ideas of statistical learning.

Coreopsys was a startup company out of Georgia Tech’s technology incubator that addressed thereal-world performance challenges of writing applications for multi-core processors. The centralpieces of Coreopsys’s technology were a result of Tushar’s PhD research.

Santa Clara, California, US

Designed and implemented a Semantic Unification Framework (SUF) that tracked and correlatedthe simulation-states of a multi-processor functional simulator and a detailed uncore performancesimulator for multi-core chips, inside an execution-driven-simulation environment.The SUF allowed the behavior of user-defined application-level constructs (threads.

Santa Clara, California, US

The startup Proceler’s technology provided a software development flow that enabled ANSI Capplications to be automatically partitioned and extensively optimized across closely coupledCPUs and FPGAs on embedded system platforms. The compiler would compose a collection ofparallelized, application-specific custom-hardware datapaths for the FPGA and.

Ms, Ece

Phd, Ece

B. Tech, Electrical Engineering