Embedded FPGA (eFPGA) can reduced memory bandwidth requirements by more than 10x and allows efficient execution of future operators and activation functions
Flex Logix Technologies, Inc., the leading supplier of embedded FPGA (eFPGA) IP and reconfigurable DSP/SDR/AI solutions, announced additional applications for embedded FPGA to improve the value proposition for AI Accelerators.
Also Read: Beta Systems Unveils Lighthouse Program for Data Center Solutions
First, memory bandwidth – the scarcest resource in AI Accelerators, especially in the cloud, where model weights/parameters exceed 100 billion and HBM memory is expensive and scarce. Techniques for saving memory bandwidth are evolving faster than hardware.
“Embedded FPGA (eFPGA) can enable innovations in sub-INT4 data and weight representations (e.g. ternary, 2 bit, 3 bit, mixed or mat-mul free) to be converted on the fly by eFPGA into existing TPUs,” said Cheng Wang, Flex Logix CTO & SVP Software + Architecture. “This can also be mixed with innovations in sparsity that can further reduce the memory bandwidth requirements. Aggregate memory bandwidth reduction can be up to 16x.”
Also Read: Kyndryl Appoints New Chief Information Officer and Announces Key Practice Leadership Transitions
Second, higher performance. AI models are rapidly evolving. With most TPUs, new operators and activation functions must be handled by a much slower processor. eFPGA can be used to run the new operators and activation functions at much higher performance.
Flex Logix is already using these concepts in its own InferX AI optimized for edge vision AI models and DSP.
Also Read: Trimble Launches End-To-End Asset Lifecycle Management Software Suite
[To share your insights with us as part of editorial or sponsored content, please write to psen@itechseries.com]
