25 March 2024 in IEEE Access
Conventional DNN inference accelerators are designed with a few (up to four) large systolic arrays. As such a scale-up architecture often suffers from low utilization, a scale-out architecture, in which a single accelerator has tens of pods and each pod has a small systolic array, has been proposed. While the scale-out architecture is promising, it still incurs increasing off-chip memory access as the pods are supposed to access the duplicate tiles of tensors. Prior work has proposed a shared buffer structure to address the problem, but those architectures suffer from performance degradation due to shared buffer access latency. We make an observation that all the pods access the same rows of input and weights within a short time window. With the observation, we propose a new inference accelerator architecture, called Vectored Systolic Arrays (SAVector). SAVector consists of a new two-level on-chip buffer architecture and a tensor tile scheduling technique. In the new buffer architecture, global buffers are shared by all the pods and they keep the rows shared by the pods. And each pod has a tiny dedicated buffer. SAVector monitors the memory access behavior and timely determines to prefetch the data and flush it. In our evaluation, SAVector exhibits a very similar off-chip memory access count to the scale-up architecture and achieves 52% energy-delay-product (EDP) reduction. Also, SAVector achieves 27% EDP reduction over prior work by mitigating performance degradation from global buffer access latency.
17 June 2025 in IEEE Computer Architecture Letters
Deep Neural Network (DNN) training demands large memory capacities that exceed the limits of current GPU onboard memory. Expanding GPU memory with SSDs is a cost-effective approach. However, the low bandwidth of SSDs introduces severe performance bottlenecks in data management, particularly for Unified Virtual Memory (UVM)-based systems. The default on-demand migration mechanism in UVM causes frequent page faults and stalls, exacerbated by memory oversubscription and eviction processes along the critical path. To address these challenges, this paper proposes Memory Oversubscription-aware Scheduling for Tensor Migration (MOST), a software framework designed to improve data migration in UVM environments. MOST profiles memory access behavior and quantifies the impact of memory oversubscription stalls and schedules tensor migrations to minimize overall training time. With the profiling results, MOST executes newly designed pre-eviction and prefetching instructions within DNN kernel code. MOST effectively selects and migrates tensors that can mitigate memory oversubscription stalls, thus reducing training time. Our evaluation shows that MOST achieves an average speedup of 22.9% and 12.8% over state-of-the-art techniques, DeepUM and G10, respectively.
14 November 2024 in IEEE Embedded Systems Letters
This paper introduces a novel software technique to optimize thread allocation for merged and fused kernels in multi-tenant inference systems on embedded Graphics Processing Units (GPUs). Embedded systems equipped with GPUs face challenges in managing diverse deep learning workloads while adhering to Quality-of-Service (QoS) standards, primarily due to limited hardware resources and the varied nature of deep learning models. Prior work has relied on static thread allocation strategies, often leading to suboptimal hardware utilization. To address these challenges, we propose a new software technique called TLP Balancer. TLP Balancer automatically identifies the best-performing number of threads based on performance modeling. This approach significantly enhances hardware utilization and ensures QoS compliance, outperforming traditional fixed-thread allocation methods. Our evaluation shows that TLP Balancer improves throughput by 40% compared to the state-of-the-art automated kernel merge and fusion techniques.
28 January 2026 in IEEE Transactions on Computers
Retrieval-augmented generation pipelines store large volumes of embedding vectors in vector databases for semantic search. In Compute Express Link (CXL)-based tiered memory systems, page-level placement often promotes pages containing both hot and cold vectors, leaving many hot vectors in the long-latency, lower-bandwidth CXL tier. This imbalance increases memory access time, reduces throughput, and worsens tail latency. This paper presents Bauhaus, a software-based memory management technique that increases the share of hot vectors in host memory, thereby improving access efficiency, reducing backend stalls, and lowering tail latency. Bauhaus monitors and reorders embedding vectors at vector-level granularity using Processor Event-Based Sampling (PEBS) and Virtual Memory Area (VMA) metadata, clustering hot vectors into contiguous pages to raise hot-page density and enhance the effectiveness of page promotion. Bauhaus, integrated into the Heterogeneous Memory Software Development Kit (HMSDK) developed and publicly released by SK hynix, is evaluated on a production-grade CXL platform. Across various embedding models, datasets, and CXL-to-host memory ratios, Bauhaus achieves 42.0% higher throughput over the baseline and 15.8% over HMSDK. Bauhaus sustains high RAG performance under constrained host memory capacity, offering a practical solution for large-scale vector database deployments in modern datacenter environments.
3 November 2025 in ACM Transactions on Storage
DRAM accounts for a large fraction of the total cost of ownership of memory systems in deep learning acceleration systems. To achieve sustainable scalability, tiered memory systems with denser technologies become critical. Prior work has proposed various tiered memory systems, but no matter how a system is designed, data movements between tiers consume substantial energy. In particular, as model sizes and memory capacity demands grow, the data movements between memory tiers become more frequent, posing a challenge that tiered memory systems may reduce deployment costs but suffer from low energy efficiency. If a memory system proactively places data into a tier and timely fetches it from the tier, then the excessive data movement between tiers can be mitigated. We find that a system can statically anticipate such behaviors for all pages and localities. With this insight, we propose a new DNN acceleration system called TM-Training using flash memory. TM-Training capitalizes on the repetitive nature of the same computational patterns during execution, enabling the static establishment of optimal data placement for subsequent operations. Moreover, TM-Training employs a new data-splitting scheme to enable precise memory management. Our evaluation demonstrates that TM-Training reduces inter-tier data traffic by 64% and achieves a 55% higher throughput per watt in training than prior work.