Murali Annavaram
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Murali Annavaram has a total of 44 co-authors. Here are the top co-authors by number of shared publications:
- Abdulla Alshabanah
University of Southern California, Los Angeles, CA, USA
5 shared publications - Yongqin Wang
Roblox, San Mateo, USA
4 shared publications - Keshav Balasubramanian
University of Southern California, Los Angeles, CA, USA
3 shared publications - Rachit Rajat
NVIDIA, Santa Clara, USA
3 shared publications - Chaoyi Jiang
University of Southern California
2 shared publications - Elan Markowitz
University of Southern California
2 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 876-891
FLASH: Fault-Led, Attested, Secure Hybrid-Key Checkpoint for Confidential GPU
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 876-891
Open AccessFLASH: Fault-Led, Attested, Secure Hybrid-Key Checkpoint for Confidential GPU
- 10 September 2026
- Published by Springer Nature
- p. 624-640
MARché: Fast Masked Autoregressive Image Generation with Cache-Aware Attention
- 10 September 2026
- Published by Springer Nature
- p. 624-640
MARché: Fast Masked Autoregressive Image Generation with Cache-Aware Attention
- Open Access
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 3848-3853
HuffmanEmbed: Using Huffman Coding for Embedding Table Compression in Deep Learning Recommendation Models
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 3848-3853
Open AccessHuffmanEmbed: Using Huffman Coding for Embedding Table Compression in Deep Learning Recommendation Models
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 287-290
SoftmaxPIM: An HBM-Based PIM Architecture for Accelerating GeMV–Softmax Execution Pipeline
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 25 (2) , 287-290
SoftmaxPIM: An HBM-Based PIM Architecture for Accelerating GeMV–Softmax Execution Pipeline
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-5
Icy-Hot: Decoupled Compute Paradigm Toward a General-Purpose Superconducting CPU Design
The decline of Dennard scaling [1] has driven unsustainable power consumption, motivating exploration of post- CMOS technologies. Among these, Josephson Junction (JJ) based Single Flux Quantum (SFQ) logic offers ultrafast switching (tens of GHz) [2], [3] and extremely low switching energy (on the order of $10^{-19}$ J) [2], [4]. Even accounting for cooling, SFQ circuits can be up to $1000\times$ more energy-efficient than CMOS [4], [5], making them an attractive candidate for post-Moore computing. However, SFQ circuit designs are limited by JJ density and thus building SFQ CPUs must be cognizant of this limitation. Due to various SFQ circuit design limitations, such as limited fan out and path balancing, building a SFQ CPU requires spending majority of the JJ budget on control logic. This work proposes Icy-Hot, a novel granular decoupling of the pipeline stages in an SFQ CPU design. Icy-Hot exploits the ultra-fast SFQ computation in the superconducting zone while minimizing the JJ usage for memory and control logic by executing those pipeline stages in existing CMOS technology at 77K. Our evaluation of the proposed design achieves around a 38% power benefit while consuming only about 220,000 JJs.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-5
Icy-Hot: Decoupled Compute Paradigm Toward a General-Purpose Superconducting CPU Design
The decline of Dennard scaling [1] has driven unsustainable power consumption, motivating exploration of post- CMOS technologies. Among these, Josephson Junction (JJ) based Single Flux Quantum (SFQ) logic offers ultrafast switching (tens of GHz) [2], [3] and extremely low switching energy (on the order of $10^{-19}$ J) [2], [4]. Even accounting for cooling, SFQ circuits can be up to $1000\times$ more energy-efficient than CMOS [4], [5], making them an attractive candidate for post-Moore computing. However, SFQ circuit designs are limited by JJ density and thus building SFQ CPUs must be cognizant of this limitation. Due to various SFQ circuit design limitations, such as limited fan out and path balancing, building a SFQ CPU requires spending majority of the JJ budget on control logic. This work proposes Icy-Hot, a novel granular decoupling of the pipeline stages in an SFQ CPU design. Icy-Hot exploits the ultra-fast SFQ computation in the superconducting zone while minimizing the JJ usage for memory and control logic by executing those pipeline stages in existing CMOS technology at 77K. Our evaluation of the proposed design achieves around a 38% power benefit while consuming only about 220,000 JJs.
- Open Access
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 16-26
SAGERec: Sampling and Gating for Enhanced Long-Tail Item Recommendations
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 16-26
Open AccessSAGERec: Sampling and Gating for Enhanced Long-Tail Item Recommendations
- Open Access
- 8 October 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Recommender Systems
- Vol. 4 (4) , 1-27
The Upside of Bias: Personalizing Long-Tail Item Recommendations with Biased Sampling
Recommendation systems drive user engagement across social media, streaming platforms, and e-commerce by learning from past interactions. The relevance of a recommended item depends on the quality of the user and item embeddings learned by the model. However, interactions across items follow a highly skewed distribution. Many items receive very few interactions, while a small subset dominates user engagement, which is known as the long-tail problem. Since recommendation models learn from user-item interactions, this imbalance reinforces model bias, limiting exposure for tail items and reducing recommendation diversity. To address this challenge, we introduced Biased User History Synthesis in an earlier work, a training algorithm designed to address this issue while enhancing personalization. Our method leverages a sampling strategy that prioritizes tail items in user interaction histories, then use them to augment user representations. Our approach increases tail item visibility and generates more personalized user embeddings, improving recommendation performance across both head and tail items. This article extends our earlier work with additional details on the models and related work, and new contributions including broader evaluation on another dataset, deeper fairness and efficiency analysis, and enhanced ablation and visualization results. The source code is available at https://github.com/lkp411/BiasedUserHistorySynthesis.
- 8 October 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Recommender Systems
- Vol. 4 (4) , 1-27
Open AccessThe Upside of Bias: Personalizing Long-Tail Item Recommendations with Biased Sampling
Recommendation systems drive user engagement across social media, streaming platforms, and e-commerce by learning from past interactions. The relevance of a recommended item depends on the quality of the user and item embeddings learned by the model. However, interactions across items follow a highly skewed distribution. Many items receive very few interactions, while a small subset dominates user engagement, which is known as the long-tail problem. Since recommendation models learn from user-item interactions, this imbalance reinforces model bias, limiting exposure for tail items and reducing recommendation diversity. To address this challenge, we introduced Biased User History Synthesis in an earlier work, a training algorithm designed to address this issue while enhancing personalization. Our method leverages a sampling strategy that prioritizes tail items in user interaction histories, then use them to augment user representations. Our approach increases tail item visibility and generates more personalized user embeddings, improving recommendation performance across both head and tail items. This article extends our earlier work with additional details on the models and related work, and new contributions including broader evaluation on another dataset, deeper fairness and efficiency analysis, and enhanced ablation and visualization results. The source code is available at https://github.com/lkp411/BiasedUserHistorySynthesis.
- Open Access
- 22 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 217-225
LEAF: Lightweight, Efficient, Adaptive and Flexible Embedding for Large-Scale Recommendation Models
- 22 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 217-225
Open AccessLEAF: Lightweight, Efficient, Adaptive and Flexible Embedding for Large-Scale Recommendation Models
- Open Access
- 8 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 189-199
Biased User History Synthesis for Personalized Long-Tail Item Recommendation
Recommendation systems connect users to items and create value chains in the internet economy. Recommendation systems learn from past user-item interaction histories. As such, items that have short interaction histories, either because they are new or not popular, have been shown to be disproportionately under-recommended. This long-tail item problem can exacerbate model bias, and reinforce poor recommendation of tail items. In this paper, we propose biased user history synthesis, to not only address this problem but also achieve better personalization in recommendation systems. As a result, we concurrently improve tail and head item recommendation performance. Our approach is built on a tail item biased User Interaction History (UIH) sampling strategy and a synthesis model that produces an augmented user representation from the sampled user history. We provide a theoretical justification for our approach using information theory and demonstrate through extensive experimentation, that our model outperforms state-of-the-art baselines on tail, head, and overall recommendation. The source code is available at https://github.com/lkp411/BiasedUserHistorySynthesis.
- 8 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 189-199
Open AccessBiased User History Synthesis for Personalized Long-Tail Item Recommendation
Recommendation systems connect users to items and create value chains in the internet economy. Recommendation systems learn from past user-item interaction histories. As such, items that have short interaction histories, either because they are new or not popular, have been shown to be disproportionately under-recommended. This long-tail item problem can exacerbate model bias, and reinforce poor recommendation of tail items. In this paper, we propose biased user history synthesis, to not only address this problem but also achieve better personalization in recommendation systems. As a result, we concurrently improve tail and head item recommendation performance. Our approach is built on a tail item biased User Interaction History (UIH) sampling strategy and a synthesis model that produces an augmented user representation from the sampled user history. We provide a theoretical justification for our approach using information theory and demonstrate through extensive experimentation, that our model outperforms state-of-the-art baselines on tail, head, and overall recommendation. The source code is available at https://github.com/lkp411/BiasedUserHistorySynthesis.
- Open Access
- 14 September 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 21 (3) , 1-32
Cross-core Data Sharing for Energy-efficient GPUs
Graphics Processing Units (GPUs) are the accelerator of choice in a variety of application domains, because they can accelerate massively parallel workloads and can be easily programmed using general-purpose programming frameworks such as CUDA and OpenCL. Each Streaming Multiprocessor (SM) contains an L1 data cache (L1D) to exploit the locality in data accesses. L1D misses are costly for GPUs for two reasons. First, L1D misses consume a lot of energy as they need to access the L2 cache (L2) via an on-chip network and the off-chip DRAM in case of L2 misses. Second, L1D misses impose performance overhead if the GPU does not have enough active warps to hide the long memory access latency. We observe that threads running on different SMs share 55% of the data they read from the memory. Unfortunately, as the L1Ds are in the non-coherent memory domain, each SM independently fetches data from the L2 or the off-chip memory into its L1D, even though the data may be currently available in the L1D of another SM. Our goal is to service L1D read misses via other SMs, as much as possible, to cut down costly accesses to the L2 or the off-chip DRAM. To this end, we propose a new data-sharing mechanism, called Cross-Core Data Sharing (CCDS) . CCDS employs a predictor to estimate whether the required cache block exists in another SM. If the block is predicted to exist in another SM’s L1D, then CCDS fetches the data from the L1D that contain the block. Our experiments on a suite of 26 workloads show that CCDS improves average energy and performance by 1.30× and 1.20×, respectively, compared to the baseline GPU. Compared to the state-of-the-art data-sharing mechanism, CCDS improves average energy and performance by 1.37× and 1.11×, respectively.
- 14 September 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
- Vol. 21 (3) , 1-32
Open AccessCross-core Data Sharing for Energy-efficient GPUs
Graphics Processing Units (GPUs) are the accelerator of choice in a variety of application domains, because they can accelerate massively parallel workloads and can be easily programmed using general-purpose programming frameworks such as CUDA and OpenCL. Each Streaming Multiprocessor (SM) contains an L1 data cache (L1D) to exploit the locality in data accesses. L1D misses are costly for GPUs for two reasons. First, L1D misses consume a lot of energy as they need to access the L2 cache (L2) via an on-chip network and the off-chip DRAM in case of L2 misses. Second, L1D misses impose performance overhead if the GPU does not have enough active warps to hide the long memory access latency. We observe that threads running on different SMs share 55% of the data they read from the memory. Unfortunately, as the L1Ds are in the non-coherent memory domain, each SM independently fetches data from the L2 or the off-chip memory into its L1D, even though the data may be currently available in the L1D of another SM. Our goal is to service L1D read misses via other SMs, as much as possible, to cut down costly accesses to the L2 or the off-chip DRAM. To this end, we propose a new data-sharing mechanism, called Cross-Core Data Sharing (CCDS) . CCDS employs a predictor to estimate whether the required cache block exists in another SM. If the block is predicted to exist in another SM’s L1D, then CCDS fetches the data from the L1D that contain the block. Our experiments on a suite of 26 workloads show that CCDS improves average energy and performance by 1.30× and 1.20×, respectively, compared to the baseline GPU. Compared to the state-of-the-art data-sharing mechanism, CCDS improves average energy and performance by 1.37× and 1.11×, respectively.