We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.AR

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Hardware Architecture

Title: MF-Net: Compute-In-Memory SRAM for Multibit Precision Inference using Memory-immersed Data Conversion and Multiplication-free Operators

Abstract: We propose a co-design approach for compute-in-memory inference for deep neural networks (DNN). We use multiplication-free function approximators based on ell_1 norm along with a co-adapted processing array and compute flow. Using the approach, we overcame many deficiencies in the current art of in-SRAM DNN processing such as the need for digital-to-analog converters (DACs) at each operating SRAM row/column, the need for high precision analog-to-digital converters (ADCs), limited support for multi-bit precision weights, and limited vector-scale parallelism. Our co-adapted implementation seamlessly extends to multi-bit precision weights, it doesn't require DACs, and it easily extends to higher vector-scale parallelism. We also propose an SRAM-immersed successive approximation ADC (SA-ADC), where we exploit the parasitic capacitance of bit lines of SRAM array as a capacitive DAC. Since the dominant area overhead in SA-ADC comes due to its capacitive DAC, by exploiting the intrinsic parasitic of SRAM array, our approach allows low area implementation of within-SRAM SA-ADC. Our 8$\times$62 SRAM macro, which requires a 5-bit ADC, achieves $\sim$105 tera operations per second per Watt (TOPS/W) with 8-bit input/weight processing at 45 nm CMOS.
Subjects: Hardware Architecture (cs.AR); Systems and Control (eess.SY)
Cite as: arXiv:2102.00035 [cs.AR]
  (or arXiv:2102.00035v1 [cs.AR] for this version)

Submission history

From: Shamma Nasrin [view email]
[v1] Fri, 29 Jan 2021 19:35:19 GMT (1702kb,D)

Link back to: arXiv, form interface, contact.