Kartik Prabhu

Kartik Prabhu

Auto-generated profile, based on publicly available information

Kartik Prabhu

Department of Electrical Engineering, Stanford University, Stanford, CA, USA

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
24 January 2022 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
KPKartik Prabhu
Kartik Prabhu
AGAlbert Gural
Albert Gural
ZKZainab F. Khan
Zainab F. Khan
RRRobert M. Radway
Robert M. Radway
MGMassimo Giordano
Massimo Giordano
PRPriyanka Raina
Priyanka Raina
Implementing edge artificial intelligence (AI) inference and training is challenging with current memory technologies. As deep neural networks (DNNs) grow in size, this problem is only getting worse. This article presents CHIMERA, the first non-volatile DNN chip for both edge AI training and inference using foundry on-chip resistive RAM (RRAM) macros and no off-chip memory, fabricated in 40-nm CMOS. CHIMERA’s DNN accelerator is specifically optimized for RRAM and achieves 0.92-TOPS peak performance and 2.2-TOPS/W energy efficiency. We scale inference up to $6\times $ larger DNNs by connecting six CHIMERAs in an illusion system with just 4% overhead in measured execution time and 5% in energy, enabled by communication-sparse DNN mappings that exploit RRAM non-volatility through quick chip wake-up and shutdown ( $ < 33 ~\mu \text{s}$ ). Our incremental edge AI training algorithm, called low-rank training, overcomes RRAM write energy, speed, and endurance challenges and achieves the same accuracy as traditional algorithms with up to $283\times $ fewer RRAM weight update steps and $340\times $ better energy-delay product. Combined with ENDUrance REsiliency using random Remapping (ENDURER), a hardware module that provides resilience to write endurance failures, we enable ten years of 20-samples/min incremental edge AI training.

27 April 2024
Verified
Conference Paper
Open access
JYJeffrey Yu
Jeffrey Yu
KPKartik Prabhu
Kartik Prabhu
YUYonatan Urman
Yonatan Urman
RRRobert M. Radway
Robert M. Radway
EHEric Han
Eric Han
PRPriyanka Raina
Priyanka Raina
Transformer models achieve state-of-the-art accuracy on natural language processing (NLP) and vision tasks, but demand significant computation and memory resources, which makes it difficult to perform inference and training (fine-tuning) on edge accelerators. Quantization to lower precision data types is a promising way to reduce computation and memory resources. Prior work has employed 8-bit integer (int8) quantization for Transformer inference, but int8 lacks the precision and range required for training. 8-bit floating-point (FP8) quantization has been used for Transformer training, but prior work only quantizes the inputs to matrix multiplications and leaves the rest of the operations in high precision. This work conducts an in-depth analysis of Transformer inference and fine-tuning at the edge using two 8-bit floating-point data types: FP8 and 8-bit posit (Posit8). Unlike FP8, posit has variable length exponent and fraction fields, leading to higher precision for values around 1, making it well suited for storing Transformer weights and activations. As opposed to prior work, we evaluate the impact of quantizing all operations in both the forward and backward passes, going beyond just matrix multiplications. Specifically, our work makes the following contributions: (1) We perform Transformer inference in FP8 and Posit8, achieving less than 1% accuracy loss compared to BFloat16 through operation fusion, without the need for scaling factors. (2) We perform Transformer fine-tuning in 8 bits by adapting low-rank adaptation (LoRA) to Posit8 and FP8, enabling 8-bit GEMM operations with increased multiply-accumulate efficiency and reduced memory accesses. (3) We design an area- and power-efficient posit softmax, which employs bitwise operations to approximate the exponential and reciprocal functions. The resulting vector unit in the Posit8 accelerator, that performs both softmax computation and other element-wise operations in Transformers, is on average 33% smaller and consumes 35% less power than the vector unit in the FP8 accelerator, while maintaining the same level of accuracy. Our work demonstrates that both Posit8 and FP8 can achieve inference and fine-tuning accuracy comparable to BFloat16, while reducing accelerator's area by 30% and 34%, and power consumption by 26% and 32%, respectively.

6 March 2025 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
KPKartik Prabhu
Kartik Prabhu
RRRobert M. Radway
Robert M. Radway
JYJeffrey Yu
Jeffrey Yu
KBKai Bartolone
Kai Bartolone
MGMassimo Giordano
Massimo Giordano
PRPriyanka Raina
Priyanka Raina
Transformer models have revolutionized natural language processing (NLP) and enabled many new applications, but are challenging to deploy on resource-constrained edge devices due to their high computation and memory demands. We present MINOTAUR, an edge system-on-chip (SoC) for inference and fine-tuning of Transformer models with all memory on the chip. MINOTAUR utilizes a configurable 8-bit posit-based accelerator to achieve highly accurate and efficient inference and fine-tuning. To minimize memory power, MINOTAUR employs fine-grained spatiotemporal power gating of on-chip resistive-RAM (RRAM). MINOTAUR enables on-chip fine-tuning through full-network low-rank adaptation (LoRA). MINOTAUR fabricates in a 40-nm CMOS process, achieves ResNet-18 inference in 8.1 mJ and MobileBERTTINY inference in 8.2 mJ, and performs on-chip fine-tuning with an accuracy that is within 1.7% of offline training.

26 February 2026 in Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences
Verified
Journal Article
Open access
PDF file
KPKartik Prabhu
Kartik Prabhu
JYJeffrey Yu
Jeffrey Yu
XPXinyuan Allen Pan
Xinyuan Allen Pan
PRPriyanka Raina
Priyanka Raina
As neural network models continue to grow in scale and complexity, specialized hardware accelerators have emerged to meet the increased demand for compute and memory. These accelerators employ a wide range of architectural innovations, making it challenging to perform fair comparisons and isolate the impact of specific design decisions. Traditional evaluation metrics, such as tera operations per second (TOPS) and TOPS per watt (TOPS/W), are heavily influenced by external factors such as technology node, clock frequency, the scale of the design and workload variations, limiting their effectiveness for meaningful analysis. In this work, we propose a methodology for benchmarking neural network accelerators using Voyager, a high-level synthesis (HLS)-based accelerator generator. Voyager enables the creation of baseline accelerators matched in compute scale and technology node capable of running identical workloads. This enables fair, apples-to-apples comparisons across diverse accelerator architectures. We showcase this methodology with a range of case studies, including those on technology scaling and involving comparisons with state-of-the-art digital accelerators, in-memory computing-based accelerators, and sparsity-aware designs. Our results demonstrate that Voyager-generated designs serve as well-optimized baselines that enable systematic evaluation of accelerator architectures. This article is part of the discussion meeting issue ‘Bits, neurons and qubits for sustainable AI’.

26 July 2026
Verified
Conference Paper
Open access
KPKartik Prabhu
Kartik Prabhu
JYJeffrey Yu
Jeffrey Yu
XPXinyuan Allen Pan
Xinyuan Allen Pan
ZXZhouhua Xie
Zhouhua Xie
AAAbigail Aleshire
Abigail Aleshire
PRPriyanka Raina
Priyanka Raina

12 March 2026 in IEEE Solid-State Circuits Letters
Verified
Journal Article
Closed access
KFKathleen Feng
Kathleen Feng
KBKai Bartolone
Kai Bartolone
KPKartik Prabhu
Kartik Prabhu
BMBrianna McColm
Brianna McColm
XNXingyu Ni
Xingyu Ni
PRPriyanka Raina
Priyanka Raina
Birch is a system-on-chip that efficiently and accurately accelerates the multi-task multi-domain extended reality (XR) perception pipeline, with workloads such as visual inertial odometry (VIO), eye gaze tracking, and scene understanding. Birch features vision modules with cascaded line buffers, in-step feature sorting, and double-buffered optical flow to extract and track quality features for VIO, a configurable matrix unit and vector unit for acceleration of deep neural network (DNN)-based tasks like eye gaze tracking and scene understanding, and reuse of the DNN vector unit for acceleration of matrix multiplications in VIO to save area in small-form-factor XR devices. Birch extracts and tracks features in 7.14 ms and 2.29 mJ per frame. For DNNs, Birch runs ResNet-18 at 109 frames-per-second (FPS) and eye gaze extraction at 985 FPS, 3.5× and 1.6× faster than prior work, respectively. ResNet-50 also runs in real time at 46.5 FPS.