An evolving solution to address hallucination and enhance accuracy in large language models (LLMs) is Retrieval-Augmented Generation (RAG), which involves augmenting LLMs with information retrieved from an external knowledge source, such as the web. This paper profiles several RAG execution pipelines and demystifies the complex interplay between their retrieval and generation phases. We demonstrate that while exact retrieval schemes are expensive, they can reduce inference time compared to approximate retrieval variants because an exact retrieval model can send a smaller but more accurate list of documents to the generative model while maintaining the same end-to-end accuracy. This observation motivates the acceleration of the exact nearest neighbor search for RAG.

Trusted Executing Environments (TEEs) in hardware designs protect program execution from other untrusted software programs in the processor as well as untrusted off-chip hardware components. Meanwhile, Near-Memory Processing (NMP) has shown performance and energy benefits on memory-intensive workloads. Recently, novel memory encryption schemes have been proposed to allow TEEs to leverage the benefits of NMP without requiring trust in the NMP components. In this paper, we present a system design of confidential computing with NMP that can be directly used in Intel SGX, a TEE platform available in commercial processors today. We develop the full software stack and evaluate the results on commercial processors with the emulated AxDIMM, an FPGA-based NMP platform. In our case study on personalized Deep Learning Recommendation Model (DLRM) inference, the proposed confidential computing in NMP achieves up to 1.51× latency reduction and up to 2.57× throughput improvement.

Large language models (LLMs) have emerged due to their capability to generate high-quality content across diverse contexts. To reduce their explosively increasing demands for computing resources, a mixture of experts (MoE) has emerged. The MoE layer enables exploiting a huge number of parameters with less computation. Applying state-of-the-art continuous batching increases throughput; however, it leads to frequent DRAM access in the MoE and attention layers. We observe that conventional computing devices have limitations when processing the MoE and attention layers, which dominate the total execution time and exhibit low arithmetic intensity (Op/B). Processing MoE layers only with devices targeting low-Op/B such as processing-in-memory (PIM) architectures is challenging due to the fluctuating Op/B in the MoE layer caused by continuous batching, To address these challenges, we propose Duplex, which comprises xPU tailored for high-Op/B and Logic-PIM to effectively perform low-Op/B operation within a single device. Duplex selects the most suitable processor based on the Op/B of each layer within LLMs. As the Op/B of the MoE layer is at least 1 and that of the attention layer has a value of 4–8 for grouped query attention, prior PIM architectures are not efficient, which place processing units inside DRAM dies and only target extremely low-Op/B (under one) operations. Based on recent trends, Logic-Pimadds more through-silicon vias (TSVs) to enable high-bandwidth communication between the DRAM die and the logic die and place powerful processing units on the logic die, which is best suited for handling low-Op/B operations ranging from few to a few dozens. To maximally utilize the xPU and Logic-Pim,we propose expert and attention co-processing. By exploiting proper processing units for MoE and attention layers, Duplex shows up to 2.67 × higher throughput and consumes 42.0% less energy compared to GPU systems for LLM inference.

In this paper, we implemented a world-first RNN-T inference accelerator using FPGA with PIM-HBM that can multiply the internal bandwidth of the memory. The accelerator offloads matrix-vector multiplication (GEMV) operations of LSTM layers in RNN-T into PIM-HBM, and PIM-HBM reduces the execution time of GEMV significantly by exploiting HBM internal bandwidth. To ensure that the memory commands are issued in a pre-defined order, which is one of the most important constraints in exploiting PIM-HBM, we implement a direct memory access (DMA) module and change configuration of the on-chip memory controller by utilizing the flexibility and reconfigurability of the FPGA. In addition, we design the other hardware modules for acceleration such as non-linear functions (i.e., sigmoid and hyperbolic tangent), element-wise operation, and ReLU module, to operate these compute-bound RNN-T operations on FPGA. For this, we prepare FP16 quantized weight and MLPerf input datasets, and modify the PCIe device driver and C++ based control codes. On our evaluation, our accelerator with PIM-HBM reduces the execution time of RNN-T by 2.5 × on average with 11.09% reduced LUT size and improves energy efficiency up to 2.6 × compared to the baseline.

Using PIM to overcome memory bottleneck • Although various bandwidth increase methods have been proposed, it is physically impossible to achieve a breakthrough increase. - Limited by # of PCB wires, # of CPU ball, and thermal constraints • PIM has been proposed to improve performance of bandwidth-intensive workloads and improve energy efficiency by reducing computing-memory data movement.

Emerging applications such as deep neural network demand high off-chip memory bandwidth. However, under stringent physical constraints of chip packages and system boards, it becomes very expensive to further increase the bandwidth of off-chip memory. Besides, transferring data across the memory hierarchy constitutes a large fraction of total energy consumption of systems, and the fraction has steadily increased with the stagnant technology scaling and poor data reuse characteristics of such emerging applications. To cost-effectively increase the bandwidth and energy efficiency, researchers began to reconsider the past processing-in-memory (PIM) architectures and advance them further, especially exploiting recent integration technologies such as 2.5D/3D stacking. Albeit the recent advances, no major memory manufacturer has developed even a proof-of-concept silicon yet, not to mention a product. This is because the past PIM architectures often require changes in host processors and/or application code which memory manufacturers cannot easily govern. In this paper, elegantly tackling the aforementioned challenges, we propose an innovative yet practical PIM architecture. To demonstrate its practicality and effectiveness at the system level, we implement it with a 20nm DRAM technology, integrate it with an unmodified commercial processor, develop the necessary software stack, and run existing applications without changing their source code. Our evaluation at the system level shows that our PIM improves the performance of memory-bound neural network kernels and applications by 11.2× and 3.5×, respectively. Atop the performance improvement, PIM also reduces the energy per bit transfer by 3.5×, and the overall energy efficiency of the system running the applications by 3.2×.

Recurrent Neural Networks (RNNs) spend most of their execution time performing matrix-vector multiplication (MV-mul). Because the matrices in RNNs have poor reusability and the ever-increasing size of the matrices becomes too large to fit in the on-chip storage of mobile/IoT devices, the performance and energy efficiency of MV-mul is determined by those of main-memory DRAM. Therefore, computing MV-mul within DRAM draws much attention. However, previous studies lacked consideration for the matrix sparsity, the power constraints of DRAM devices, and concurrency in accessing DRAM from processors while performing MV-mul. We propose a main-memory architecture called MViD, which performs MV-mul by placing MAC units inside DRAM banks. For higher computational efficiency, we use a sparse matrix format and exploit quantization. Because of the limited power budget for DRAM devices, we implement the MAC units only on a portion of the DRAM banks. We architect MViD to slow down or pause MV-mul for concurrently processing memory requests from processors while satisfying the limited power budget. Our results show that MViD provides 7.2× higher throughput compared to the baseline system with four DRAM ranks (performing MV-mul in a chip-multiprocessor) while running inference of Deep Speech 2 with a memory-intensive workload.

Computer systems using DRAM are exposed to row-hammer (RH) attacks, which can flip data in a DRAM row without directly accessing a row but by frequently activating its adjacent ones. There have been a number of proposals to prevent RH, but they either incur large area overhead, suffer from noticeable performance drop on adversarial memory access patterns, or provide probabilistic protection with no capability to detect attacks.

The advance of DRAM manufacturing technology slows down, whereas the density and performance needs of DRAM continue to increase. This desire has motivated the industry to explore emerging Non-Volatile Memory (e.g., 3D XPoint) and the high-density DRAM (e.g., Managed DRAM Solution). Since such memory technologies increase the density at the cost of longer latency, lower bandwidth, or both, it is essential to use them with fast memory (e.g., conventional DRAM) to which hot pages are transferred at runtime. Nonetheless, we observe that page transfers to fast memory often block memory channels from servicing memory requests from applications for a long period. This in turn significantly increases the high-percentile response time of latency-sensitive applications. In this paper, we propose a high-density managed DRAM architecture, dubbed 3D-XPath for applications demanding both low latency and high capacity for memory. 3D-XPath DRAM stacks conventional DRAM dies with high-density DRAM dies explored in this paper and connects these DRAM dies with 3D-XPath. Especially, 3D-XPath allows unused memory channels to service memory requests from applications when primary channels supposed to handle the memory requests are blocked by page transfers at given moments, considerably increasing the high-percentile response time. This can also improve the throughput of applications frequently copying memory blocks between kernel and user memory spaces. Our evaluation shows that 3D-XPath DRAM decreases high-percentile response time of latency-sensitive applications by ~30% while improving the throughput of an I/O-intensive applications by ~39%, compared with DRAM without 3D-XPath.

Distributed in-memory key-value stores (KVSs), such as memcached, have become a critical data serving layer in modern Internet-oriented data center infrastructure. Their performance and efficiency directly affect the QoS of web services and the efficiency of data centers. Traditionally, these systems have had significant overheads from inefficient network processing, OS kernel involvement, and concurrency control. Two recent research thrusts have focused on improving key-value performance. Hardware-centric research has started to explore specialized platforms including FPGAs for KVSs; results demonstrated an order of magnitude increase in throughput and energy efficiency over stock memcached. Software-centric research revisited the KVS application to address fundamental software bottlenecks and to exploit the full potential of modern commodity hardware; these efforts also showed orders of magnitude improvement over stock memcached. We aim at architecting high-performance and efficient KVS platforms, and start with a rigorous architectural characterization across system stacks over a collection of representative KVS implementations. Our detailed full-system characterization not only identifies the critical hardware/software ingredients for high-performance KVS systems but also leads to guided optimizations atop a recent design to achieve a record-setting throughput of 120 million requests per second (MRPS) (167MRPS with client-side batching) on a single commodity server. Our system delivers the best performance and energy efficiency (RPS/watt) demonstrated to date over existing KVSs including the best-published FPGA-based and GPU-based claims. We craft a set of design principles for future platform architectures, and via detailed simulations demonstrate the capability of achieving a billion RPS with a single server constructed following our principles.

Although aggressive technology scaling has allowed manufacturers to integrate Giga bits of cells into a cost-sensitive main memory DRAM device, these cells have become more defect-prone. With increased cell failure rates, conventional solutions such as populating spare DRAM rows and relying on error-correcting codes (ECCs) have shown limited success due to high area overhead, the latency penalties of data coding, and interference between ECC within a device (in-DRAM ECC) and other ECC across devices (rank-level ECC). In this paper, we propose CiDRA, a cache-inspired DRAM resilience architecture, which substantially reduces the area and latency overheads of correcting bit errors on random locations due to these faulty cells. We put a small SRAM cache within a DRAM device to replace accesses to the addresses including the faulty cells with ones that correspond to the cache data array. This CiDRA cache is paired with a Bloom filter to minimize the energy overhead of accessing the cache tags for every DRAM access and is also partitioned into small pieces, each being associated with the I/O pads for better area efficiency. Both the cache and DRAM banks are accessed in parallel while the banks are much slower. Consequently, the cache and filter are not in the critical path for normal DRAM accesses and incur no latency overhead. We also enhance the traditional in-DRAM ECC with error position bits and the appropriate error detecting capability while preventing interference with the traditional rank-level ECC scheme. By combining this enhanced in-DRAM ECC with the cache and Bloom filter, CiDRA becomes more area efficient because the in-DRAM ECC corrects most bit errors that are sporadic while the cache deals with the remaining relatively few pathological cases.

Distributed in-memory key-value stores (KVSs), such as memcached, have become a critical data serving layer in modern Internet-oriented datacenter infrastructure. Their performance and efficiency directly affect the QoS of web services and the efficiency of datacenters. Traditionally, these systems have had significant overheads from inefficient network processing, OS kernel involvement, and concurrency control. Two recent research thrusts have focused upon improving key-value performance. Hardware-centric research has started to explore specialized platforms including FPGAs for KVSs; results demonstrated an order of magnitude increase in throughput and energy efficiency over stock memcached. Software-centric research revisited the KVS application to address fundamental software bottlenecks and to exploit the full potential of modern commodity hardware; these efforts too showed orders of magnitude improvement over stock memcached.