More Like this

1. APOLLO: SGD-like Memory, AdamW-level Performance

   Zhu, H; Zhang, Z; Cong, W; Liu, X; Park, S; Chandra, V; Long, B; Pan, D Z; Wang, Z; Lee, J (February 2025, https://doi.org/10.48550/arXiv.2412.05270)

   Large language models (LLMs) are notoriously memory-intensive during training, particularly with the popular AdamW optimizer. This memory burden necessitates using more or higher-end GPUs or reducing batch sizes, limiting training scalability and throughput. To address this, various memory-efficient optimizers have been proposed to reduce optimizer memory usage. However, they face critical challenges: (i) reliance on costly SVD operations; (ii) significant performance trade-offs compared to AdamW; and (iii) still substantial optimizer memory overhead to maintain competitive performance. In this work, we identify that AdamW's learning rate adaptation rule can be effectively coarsened as a structured learning rate update. Based on this insight, we propose Approximated Gradient Scaling for Memory-Efficient LLM Optimization (APOLLO), which approximates learning rate scaling using an auxiliary low-rank optimizer state based on pure random projection. This structured learning rate update rule makes APOLLO highly tolerant to further memory reductions while delivering comparable pre-training performance. Even its rank-1 variant, APOLLO-Mini, achieves superior pre-training performance compared to AdamW with SGD-level memory costs. Extensive experiments demonstrate that the APOLLO series performs on-par with or better than AdamW, while achieving greater memory savings by nearly eliminating the optimization states of AdamW. These savings provide significant system-level benefits: (1) Enhanced Throughput: 3x throughput on an 8xA100-80GB setup compared to AdamW by supporting 4x larger batch sizes. (2) Improved Model Scalability: Pre-training LLaMA-13B with naive DDP on A100-80GB GPUs without system-level optimizations. (3) Low-End GPU Friendly Pre-training: Pre-training LLaMA-7B on a single GPU using less than 12 GB of memory with weight quantization. 

   more » « less
2. APOLLO: SGD-like Memory, AdamW-level Performance

   Zhu, Hanqing; Zhang, Zhang; Cong, Wenyan; Liu, Xi; Park, Sem; Chandra, Vikas; Long, Bo; Zan, David Z; Wang, Zhangyang; Lee, Jinwon (February 2025, MLSys 2025)

   Large language models (LLMs) are notoriously memory-intensive during training, particularly with the popular AdamW optimizer. This memory burden necessitates using more or higher-end GPUs or reducing batch sizes, limiting training scalability and throughput. To address this, various memory-efficient optimizers have been proposed to reduce optimizer memory usage. However, they face critical challenges: (i) reliance on costly SVD operations; (ii) significant performance trade-offs compared to AdamW; and (iii) still substantial optimizer memory overhead to maintain competitive performance. In this work, we identify that AdamW's learning rate adaptation rule can be effectively coarsened as a structured learning rate update. Based on this insight, we propose Approximated Gradient Scaling for Memory-Efficient LLM Optimization (APOLLO), which approximates learning rate scaling using an auxiliary low-rank optimizer state based on pure random projection. This structured learning rate update rule makes APOLLO highly tolerant to further memory reductions while delivering comparable pre-training performance. Even its rank-1 variant, APOLLO-Mini, achieves superior pre-training performance compared to AdamW with SGD-level memory costs. Extensive experiments demonstrate that the APOLLO series performs on-par with or better than AdamW, while achieving greater memory savings by nearly eliminating the optimization states of AdamW. These savings provide significant system-level benefits: (1) Enhanced Throughput: 3x throughput on an 8xA100-80GB setup compared to AdamW by supporting 4x larger batch sizes. (2) Improved Model Scalability: Pre-training LLaMA-13B with naive DDP on A100-80GB GPUs without system-level optimizations. (3) Low-End GPU Friendly Pre-training: Pre-training LLaMA-7B on a single GPU using less than 12 GB of memory with weight quantization. 

   more » « less
3. Systems-on-Chip with Strong Ordering

   https://doi.org/10.1145/3428153  

   Puthoor, Sooraj; Lipasti, Mikko H. (January 2021, ACM Transactions on Architecture and Code Optimization)

   null (Ed.)

   Sequential consistency (SC) is the most intuitive memory consistency model and the easiest for programmers and hardware designers to reason about. However, the strict memory ordering restrictions imposed by SC make it less attractive from a performance standpoint. Additionally, prior high-performance SC implementations required complex hardware structures to support speculation and recovery. In this article, we introduce the lockstep SC consistency model (LSC), a new memory model based on SC but carefully defined to accommodate the data parallel lockstep execution paradigm of GPUs. We also describe an efficient LSC implementation for an APU system-on-chip (SoC) and show that our implementation performs close to the baseline relaxed model. Evaluation of our implementation shows that the geometric mean performance cost for lockstep SC is just 0.76% for GPU execution and 6.11% for the entire APU SoC compared to a baseline with a weaker memory consistency model. Adoption of LSC in future APU and SoC designs will reduce the burden on programmers trying to write correct parallel programs, while also simplifying the implementation and verification of systems with heterogeneous processing elements and complex memory hierarchies. 1 

   more » « less
4. EQ-ViT: Algorithm-Hardware Co-Design for End-to-End Acceleration of Real-Time Vision Transformer Inference on Versal ACAP Architecture

   Dong, Peiyan; Zhuang, Jinming; Yang, Zhuoping; Ji, Shixin; Li, Yanyu; Xu, Dongkuan; Huang, Heng; Hu, Jingtong; Jones, Alex K; Shi, Yiyu; et al (October 2024, IEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS)

   While Vision Transformers (ViTs) have shown consistent progress in computer vision, deploying them for real-time decision-making scenarios (< 1 ms) is challenging. Current computing platforms like CPUs, GPUs, or FPGA-based solutions struggle to meet this deterministic low-latency real-time requirement, even with quantized ViT models. Some approaches use pruning or sparsity to reduce model size and latency, but this often results in accuracy loss. To address the aforementioned constraints, in this work, we propose EQ-ViT, an end-to-end acceleration framework with novel algorithm and architecture co-design features to enable real-time ViT acceleration on AMD Versal Adaptive Compute Acceleration Platform (ACAP). The contributions are four-fold. First, we perform in-depth kernel- level performance profiling & analysis and explain the bottlenecks for existing acceleration solutions on GPU, FPGA, and ACAP. Second, on the hardware level, we introduce a new spatial and heterogeneous accelerator architecture, EQ-ViT architec- ture. This architecture leverages the heterogeneous features of ACAP, where both FPGA and artificial intelligence engines (AIEs) coexist on the same system-on-chip (SoC). Third, On the algorithm level, we create a comprehensive quantization-aware training strategy, EQ-ViT algorithm. This strategy concurrently quantizes both weights and activations into 8-bit integers, aiming to improve accuracy rather than compromise it during quanti- zation. Notably, the method also quantizes nonlinear functions for efficient hardware implementation. Fourth, we design EQ- ViT automation framework to implement the EQ-ViT architec- ture for four different ViT applications on the AMD Versal ACAP VCK190 board, achieving accuracy improvement with 2.4%, and average speedups of 315.0x, 3.39x, 3.38x, 14.92x, 59.5x, 13.1x over computing solutions of Intel Xeon 8375C vCPU, Nvidia A10G, A100, Jetson AGX Orin GPUs, and AMD ZCU102, U250 FPGAs. The energy efficiency gains are 62.2x, 15.33x, 12.82x, 13.31x, 13.5x, 21.9x. 

   more » « less
5. Fine-grain Quantitative Analysis of Demand Paging in Unified Virtual Memory

   https://doi.org/10.1145/3632953  

   Allen, Tyler; Cooper, Bennett; Ge, Rong (March 2024, ACM Transactions on Architecture and Code Optimization)

   The abstraction of a shared memory space over separate CPU and GPU memory domains has eased the burden of portability for many HPC codebases. However, users pay for ease of use provided by system-managed memory with a moderate-to-high performance overhead. NVIDIA Unified Virtual Memory (UVM) is currently the primary real-world implementation of such abstraction and offers a functionally equivalent testbed for in-depth performance study for both UVM and future Linux Heterogeneous Memory Management (HMM) compatible systems. The continued advocacy for UVM and HMM motivates improvement of the underlying system. We focus on UVM-based systems and investigate the root causes of UVM overhead, a non-trivial task due to complex interactions of multiple hardware and software constituents and the desired cost granularity. In our prior work, we delved deeply into UVM system architecture and showed internal behaviors of page fault servicing in batches. We provided quantitative evaluation of batch handling for various applications under different scenarios, including prefetching and oversubscription. We revealed that the driver workload depends on the interactions among application access patterns, GPU hardware constraints, and host OS components. Host OS components have significant overhead present across implementations, warranting close attention. This extension furthers our prior study in three aspects: fine-grain cost analysis and breakdown, extension to multiple GPUs, and investigation of platforms with different GPU-GPU interconnects. We take a top-down approach to quantitative batch analysis and uncover how constituent component costs accumulate and overlap, governed by synchronous and asynchronous operations. Our multi-GPU analysis shows reduced cost of GPU-GPU batch workloads compared to CPU-GPU workloads. We further demonstrate that while specialized interconnects, NVLink, can improve batch cost, their benefits are limited by host OS software overhead and GPU oversubscription. This study serves as a proxy for future shared memory systems, such as those that interface with HMM, and the development of interconnects. 

   more » « less