Posts
Feb, 24
The AI CUDA Engineer: Agentic CUDA Kernel Discovery, Optimization and Composition
Recent advances in Large Language Models have driven large-scale deployment, resulting in ever-growing inference time and energy demand. While manual optimization of low-level code implementations is feasible, it is an arduous task that requires deep expertise to balance the complex interplay of algorithmic, software, and hardware bottlenecks. This report presents the first comprehensive agentic framework […]
Feb, 24
KernelBench: Can LLMs Write Efficient GPU Kernels?
Efficient GPU kernels are crucial for building performant machine learning architectures, but writing them is a time-consuming challenge that requires significant expertise; therefore, we explore using language models (LMs) to automate kernel generation. We introduce KernelBench, an open-source framework for evaluating LMs’ ability to write fast and correct kernels on a suite of 250 carefully […]
Feb, 16
cuSZp2: A GPU Lossy Compressor with Extreme Throughput and Optimized Compression Ratio
Existing GPU lossy compressors suffer from expensive data movement overheads, inefficient memory access patterns, and high synchronization latency, resulting in limited throughput. This work proposes CUSZP2, a generic single-kernel error-bounded lossy compressor purely on GPUs designed for applications that require high speed, such as large-scale GPU simulation and large language model training. In particular, CUSZP2 […]
Feb, 16
Leveraging LLVM OpenMP GPU Offload Optimizations for Kokkos Applications
OpenMP provides a cross-vendor API for GPU offload that can serve as an implementation layer under performance portability frameworks like the Kokkos C++ library. However, recent work identified some impediments to performance with this approach arising from limitations in the API or in the available implementations. Advanced programming concepts such as hierarchical parallelism and use […]
Feb, 16
InfiniteHiP: Extending Language Model Context Up to 3 Million Tokens on a Single GPU
In modern large language models (LLMs), handling very long context lengths presents significant challenges as it causes slower inference speeds and increased memory costs. Additionally, most existing pre-trained LLMs fail to generalize beyond their original training sequence lengths. To enable efficient and practical long-context utilization, we introduce InfiniteHiP, a novel, and practical LLM inference framework […]
Feb, 16
Teaching An Old Dog New Tricks: Porting Legacy Code to Heterogeneous Compute Architectures With Automated Code Translation
Legacy codes are in ubiquitous use in scientific simulations; they are well-tested and there is significant time investment in their use. However, one challenge is the adoption of new, sometimes incompatible computing paradigms, such as GPU hardware. In this paper, we explore using automated code translation to enable execution of legacy multigrid solver code on […]
Feb, 16
Vortex: Overcoming Memory Capacity Limitations in GPU-Accelerated Large-Scale Data Analytics
Despite the high computational throughput of GPUs, limited memory capacity and bandwidth-limited CPU-GPU communication via PCIe links remain significant bottlenecks for accelerating large-scale data analytics workloads. This paper introduces Vortex, a GPU-accelerated framework designed for data analytics workloads that exceed GPU memory capacity. A key aspect of our framework is an optimized IO primitive that […]
Feb, 10
Optimizing the optimizer increasing performance efficiency of modern compilers
A long-standing goal, which is increasingly important in the post-Moore era, is to augment system performance by building more intelligent compilers. One of our motivating hypotheses is that much of the capability needed to advance compiler optimization is already present: state-of-the-art compilers not only provide a large set of code transformations, but also (by-and-large) correctly […]
Feb, 10
Demystifying Cost-Efficiency in LLM Serving over Heterogeneous GPUs
Recent advancements in Large Language Models (LLMs) have led to increasingly diverse requests, accompanied with varying resource (compute and memory) demands to serve them. However, this in turn degrades the cost-efficiency of LLM serving as common practices primarily rely on homogeneous GPU resources. In response to this problem, this work conducts a thorough study about […]
Feb, 10
Compiler Support for Speculation in Decoupled Access/Execute Architectures
Irregular codes are bottlenecked by memory and communication latency. Decoupled access/execute (DAE) is a common technique to tackle this problem. It relies on the compiler to separate memory address generation from the rest of the program, however, such a separation is not always possible due to control and data dependencies between the access and execute […]
Feb, 10
Towards autonomous resource management: Deep learning prediction of CPU-GPU load balancing
The demand of data centers has increased due to the latest improvements of Artificial Intelligence. These data centers are composed of thousands of servers with cooling systems that consume high amounts of energy. The servers usually contain several processing units that can cooperate for solving computational tasks. When making a proper partitioning of the entire […]
Feb, 10
Ilargi: a GPU Compatible Factorized ML Model Training Framework
The machine learning (ML) training over disparate data sources traditionally involves materialization, which can impose substantial time and space overhead due to data movement and replication. Factorized learning, which leverages direct computation on disparate sources through linear algebra (LA) rewriting, has emerged as a viable alternative to improve computational efficiency. However, the adaptation of factorized […]