AMD had previously detailed its exascale research journey from initial targets and requirements to the development and evolution of its vision of a high-performance computing (HPC) accelerated processing unit (APU), dubbed the Exascale Heterogeneous Processor or EHP. At the conclusion of that work, the learnings were integrated into the design of the node architecture that went into the Frontier supercomputer, the world’s first exascale machine. However, while the Frontier node architecture embodied many of the attributes of the EHP concept, advanced heterogeneous integration capabilities at the time were not yet sufficiently mature to realize our vision of a fully-integrated APU for HPC and AI. In this paper, we finish the EHP’s story by digging deeper into why an APU was not the right solution at the time of our first exascale architecture, what the shortcomings were of previous EHP concepts, and how AMD further evolved the concept into the AMD Instinct™ MI300A APU. MI300A is the culmination of years of AMD developments in advanced packaging technologies, its APU hardware and software, and the next step in our highly effective chiplet strategy to not only deliver a groundbreaking design for exascale computing, but to also meet the demands of new large-language model and generative AI applications.

This article consists only of a collection of slides from the author's conference presentation.

Contemporary GPUs support multiple kernels to run concurrently on the same streaming multiprocessors (SMs). Recent studies have demonstrated that such concurrent kernel execution (CKE) improves both resource utilization and computational throughput. Most of the prior works focus on partitioning the GPU resources at the cooperative thread array (CTA) level or the warp scheduler level to improve CKE. However, significant performance slowdown and unfairness are observed when latency-sensitive kernels co-run with bandwidth-intensive ones. The reason is that bandwidth over-subscription from bandwidth-intensive kernels leads to much aggravated memory access latency, which is highly detrimental to latency-sensitive kernels. Even among bandwidth-intensive kernels, more intensive kernels may unfairly consume much higher bandwidth than less-intensive ones. In this article, we first make a case that such problems cannot be sufficiently solved by managing CTA combinations alone and reveal the fundamental reasons. Then, we propose a coordinated approach for CTA combination and bandwidth partitioning. Our approach dynamically detects co-running kernels as latency sensitive or bandwidth intensive. As both the DRAM bandwidth and L2-to-L1 Network-on-Chip (NoC) bandwidth can be the critical resource, our approach partitions both bandwidth resources coordinately along with selecting proper CTA combinations. The key objective is to allocate more CTA resources for latency-sensitive kernels and more NoC/DRAM bandwidth resources to NoC-/DRAM-intensive kernels. We achieve it using a variation of dominant resource fairness (DRF). Compared with two state-of-the-art CKE optimization schemes, SMK [52] and WS [55], our approach improves the average harmonic speedup by 78% and 39%, respectively. Even compared to the best possible CTA combinations, which are obtained from an exhaustive search among all possible CTA combinations, our approach improves the harmonic speedup by up to 51% and 11% on average.

Graphics Processing Units (GPUs) leverage massive thread-level parallelism (TLP) to achieve high computation throughput and hide long memory latency. However, recent studies have shown that the GPU performance does not scale with the GPU occupancy or the degrees of TLP that a GPU supports, especially for memory-intensive workloads. The current understanding points to L1 D-cache contention or off-chip memory bandwidth. In this article, we perform a novel scalability analysis from the perspective of throughput utilization of various GPU components, including off-chip DRAM, multiple levels of caches, and the interconnect between L1 D-caches and L2 partitions. We show that the interconnect bandwidth is a critical bound for GPU performance scalability. For the applications that do not have saturated throughput utilization on a particular resource, their performance scales well with increased TLP. To improve TLP for such applications efficiently, we propose a fast context switching approach. When a warp/thread block (TB) is stalled by a long latency operation, the context of the warp/TB is spilled to spare on-chip resource so that a new warp/TB can be launched. The switched-out warp/TB is switched back when another warp/TB is completed or switched out. With this fine-grain fast context switching, higher TLP can be supported without increasing the sizes of critical resources like the register file. Our experiment shows that the performance can be improved by up to 47% and a geometric mean of 22% for a set of applications with unsaturated throughput utilization. Compared to the state-of-the-art TLP improvement scheme, our proposed scheme achieves 12% higher performance on average and 16% for unsaturated benchmarks.

The high amount of memory requests from massive threads may easily cause cache contention and cache-miss-related resource congestion on GPUs. This paper proposes a simple yet effective performance model to estimate the impact of cache contention and resource congestion as a function of the number of warps/thread blocks (TBs) to bypass the cache. Then we design a hardware-based dynamic warp/thread-block level GPU cache bypassing scheme, which achieves 1.68x speedup on average on a set of memory-intensive benchmarks over the baseline. Compared to prior works, our scheme achieves 21.6% performance improvement over SWL-best [29] and 11.9% over CBWT-best [4] on average.

Many-core architectures such as graphics processing units (GPUs) rely on thread-level parallelism (TLP)to overcome pipeline hazards. Consequently, each core in a many-core processor employs a relatively simple in-order pipeline with limited capability to exploit instruction-level parallelism (ILP). In this paper, we study the ILP impact on the throughput-oriented many-core architecture, including data bypassing, score boarding and branch prediction. We show that these ILP techniques significantly reduce the performance dependency on TLP. This is especially useful for applications, whose resource usage limits the hardware to run a high number of threads concurrently. Furthermore, ILP techniques reduce the demand on on-chip resource to support high TLP. Given the workload-dependent impact from ILP, we propose heterogeneous GPGPU architecture, consisting of both the cores designed for high TLP and those customized with ILPtechniques. Our results show that our heterogeneous GPUarchitecture achieves high throughput as well as high energy and area-efficiency compared to homogenous designs.

The wide availability and the Single-Instruction Multiple-Thread (SIMT)-style programming model have made graphics processing units (GPUs) a promising choice for high performance computing. However, because of the SIMT style processing, an instruction will be executed in every thread even if the operands are identical for all the threads. To overcome this inefficiency, the AMD's latest Graphics Core Next (GCN) architecture integrates a scalar unit into a SIMT unit. In GCN, both the SIMT unit and the scalar unit share a single SIMT style instruction stream. Depending on its type, an instruction is issued to either a scalar or a SIMT unit. In this paper, we propose to extend the scalar unit so that it can either share the instruction stream with the SIMT unit or execute a separate instruction stream. The program to be executed by the scalar unit is referred to as a scalar program and its purpose is to assist SIMT-unit execution. The scalar programs are either generated from SIMT programs automatically by the compiler or manually developed by expert developers. We make a case for our proposed flexible scalar unit through three collaborative execution paradigms: data prefetching, control divergence elimination, and scalar-workload extraction. Our experimental results show that significant performance gains can be achieved using our proposed approaches compared to the state-of-art SIMT style processing.

State-of-art graphics processing units (GPUs) employ the single-instruction multiple-data (SIMD) style execution to achieve both high computational throughput and energy efficiency. As previous works have shown, there exists significant computational redundancy in SIMD execution, where different execution lanes operate on the same operand values. Such value locality is referred to as uniform vectors. In this paper, we first show that besides redundancy within a uniform vector, different vectors can also have the identical values. Then, we propose detailed architecture designs to exploit both types of redundancy. For redundancy within a uniform vector, we propose to either extend the vector register file with token bits or add a separate small scalar register file to eliminate redundant computations as well as redundant data storage. For redundancy across different uniform vectors, we adopt instruction reuse, proposed originally for CPU architectures, to detect and eliminate redundancy. The elimination of redundant computations and data storage leads to both significant energy savings and performance improvement. Furthermore, we propose to leverage such redundancy to protect arithmetic-logic units (ALUs) and register files against hardware errors. Our detailed evaluation shows that our proposed design has low hardware overhead and achieves performance gains, up to 23.9% and 12.0% on average, along with energy savings, up to 24.8% and 12.6% on average, as well as a 21.1% and 14.1% protection coverage for ALUs and register files, respectively.

No abstract available.

On-chip shared memory (a.k.a. local data share) is a critical resource to many GPGPU applications. In current GPUs, the shared memory is allocated when a thread block (also called a workgroup) is dispatched to a streaming multiprocessor (SM) and is released when the thread block is completed. As a result, the limited capacity of shared memory becomes a bottleneck for a GPU to host a high number of thread blocks, limiting the otherwise available thread-level parallelism (TLP). In this paper, we propose software and/or hardware approaches to multiplex the shared memory among multiple thread blocks.

This paper presents a novel approach to utilize the CPU resource to facilitate the execution of GPGPU programs on fused CPU-GPU architectures. In our model of fused architectures, the GPU and the CPU are integrated on the same die and share the on-chip L3 cache and off-chip memory, similar to the latest Intel Sandy Bridge and AMD accelerated processing unit (APU) platforms. In our proposed CPU-assisted GPGPU, after the CPU launches a GPU program, it executes a pre-execution program, which is generated automatically from the GPU kernel using our proposed compiler algorithms and contains memory access instructions of the GPU kernel for multiple thread-blocks. The CPU pre-execution program runs ahead of GPU threads because (1) the CPU pre-execution thread only contains memory fetch instructions from GPU kernels and not floating-point computations, and (2) the CPU runs at higher frequencies and exploits higher degrees of instruction-level parallelism than GPU scalar cores. We also leverage the prefetcher at the L2-cache on the CPU side to increase the memory traffic from CPU. As a result, the memory accesses of GPU threads hit in the L3 cache and their latency can be drastically reduced. Since our pre-execution is directly controlled by user-level applications, it enjoys both high accuracy and flexibility. Our experiments on a set of benchmarks show that our proposed pre-execution improves the performance by up to 113% and 21.4% on average.

Given the extraordinary computational power of modern graphics processing units (GPUs), general purpose computation on GPUs (GPGPU) has become an increasingly important platform for high performance computing. To better understand how well the GPU resource has been utilized by application developers and then to facilitate them to develop high performance GPGPU code, we conduct an empirical study on GPGPU programs from ten open-source projects. These projects span a wide range of disciplines and many are designed as high performance libraries. Among these projects, we found various performance 'bugs', i.e., code segments leading to inefficient use of GPU hardware. We characterize these performance bugs, and propose the bug fixes. Our experiments confirm both significant performance gains and energy savings from our fixes and reveal interesting insights on different GPUs.

This article presents a novel optimizing compiler for general purpose computation on graphics processing units (GPGPU). It addresses two major challenges of developing high performance GPGPU programs: effective utilization of GPU memory hierarchy and judicious management of parallelism. The input to our compiler is a naïve GPU kernel function, which is functionally correct but without any consideration for performance optimization. The compiler generates two kernels, one optimized for global memories and the other for texture memories. The proposed compilation process is effective for both AMD/ATI and NVIDIA GPUs. The experiments show that our optimized code achieves very high performance, either superior or very close to highly fine-tuned libraries.