Modern Graphic Processing Units (GPUs) provide sufficiently flexible programming models that understanding their performance can provide insight in designing tomorrow's manycore processors, whether those are GPUs or otherwise. The combination of multiple, multithreaded, SIMD cores makes studying these GPUs useful in understanding tradeoffs among memory, data, and thread level parallelism. While modern GPUs offer orders of magnitude more raw computing power than contemporary CPUs, many important applications, even those with abundant data level parallelism, do not achieve peak performance. This paper characterizes several non-graphics applications written in NVIDIA's CUDA programming model by running them on a novel detailed microarchitecture performance simulator that runs NVIDIA's parallel thread execution (PTX) virtual instruction set. For this study, we selected twelve non-trivial CUDA applications demonstrating varying levels of performance improvement on GPU hardware (versus a CPU-only sequential version of the application). We study the performance of these applications on our GPU performance simulator with configurations comparable to contemporary high-end graphics cards. We characterize the performance impact of several microarchitecture design choices including choice of interconnect topology, use of caches, design of memory controller, parallel workload distribution mechanisms, and memory request coalescing hardware. Two observations we make are (1) that for the applications we study, performance is more sensitive to interconnect bisection bandwidth rather than latency, and (2) that, for some applications, running fewer threads concurrently than on-chip resources might otherwise allow can improve performance by reducing contention in the memory system.

/pdf/analyzing-cuda-workloads-using-a-detailed-gpu-simulator-lz9yu5o7va.pdf

Analyzing CUDA workloads using a detailed GPU simulator

A key question for the microprocessor research and design community is whether scaling multicores will provide the performance and value needed to scale down many more technology generations. To provide a quantitative answer to this question, a comprehensive study that projects the speedup potential of future multicores and examines the underutilization of integration capacity-dark silicon-is timely and crucial.

/pdf/dark-silicon-and-the-end-of-multicore-scaling-fxij8qeoty.pdf

Dark Silicon and the End of Multicore Scaling

Since 2005, processor designers have increased core counts to exploit Moore's Law scaling, rather than focusing on single-core performance. The failure of Dennard scaling, to which the shift to multicore parts is partially a response, may soon limit multicore scaling just as single-core scaling has been curtailed. This paper models multicore scaling limits by combining device scaling, single-core scaling, and multicore scaling to measure the speedup potential for a set of parallel workloads for the next five technology generations. For device scaling, we use both the ITRS projections and a set of more conservative device scaling parameters. To model single-core scaling, we combine measurements from over 150 processors to derive Pareto-optimal frontiers for area/performance and power/performance. Finally, to model multicore scaling, we build a detailed performance model of upper-bound performance and lower-bound core power. The multicore designs we study include single-threaded CPU-like and massively threaded GPU-like multicore chip organizations with symmetric, asymmetric, dynamic, and composed topologies. The study shows that regardless of chip organization and topology, multicore scaling is power limited to a degree not widely appreciated by the computing community. Even at 22 nm (just one year from now), 21% of a fixed-size chip must be powered off, and at 8 nm, this number grows to more than 50%. Through 2024, only 7.9x average speedup is possible across commonly used parallel workloads, leaving a nearly 24-fold gap from a target of doubled performance per generation.

/pdf/dark-silicon-and-the-end-of-multicore-scaling-goghgnng6o.pdf

Dark silicon and the end of multicore scaling

This work observes that a large fraction of the computations performed by Deep Neural Networks (DNNs) are intrinsically ineffectual as they involve a multiplication where one of the inputs is zero. This observation motivates Cnvlutin (CNV), a value-based approach to hardware acceleration that eliminates most of these ineffectual operations, improving performance and energy over a state-of-the-art accelerator with no accuracy loss. CNV uses hierarchical data-parallel units, allowing groups of lanes to proceed mostly independently enabling them to skip over the ineffectual computations. A co-designed data storage format encodes the computation elimination decisions taking them off the critical path while avoiding control divergence in the data parallel units. Combined, the units and the data storage format result in a data-parallel architecture that maintains wide, aligned accesses to its memory hierarchy and that keeps its data lanes busy. By loosening the ineffectual computation identification criterion, CNV enables further performance and energy efficiency improvements, and more so if a loss in accuracy is acceptable. Experimental measurements over a set of state-of-the-art DNNs for image classification show that CNV improves performance over a state-of-the-art accelerator from 1.24× to 1.55× and by 1.37× on average without any loss in accuracy by removing zero-valued operand multiplications alone. While CNV incurs an area overhead of 4.49%, it improves overall EDP (Energy Delay Product) and ED2P (Energy Delay Squared Product) on average by 1.47× and 2.01×, respectively. The average performance improvements increase to 1.52× without any loss in accuracy with a broader ineffectual identification policy. Further improvements are demonstrated with a loss in accuracy.

Cnvlutin: ineffectual-neuron-free deep neural network computing

GPU architectures are increasingly important in the multi-core era due to their high number of parallel processors. Programming thousands of massively parallel threads is a big challenge for software engineers, but understanding the performance bottlenecks of those parallel programs on GPU architectures to improve application performance is even more difficult. Current approaches rely on programmers to tune their applications by exploiting the design space exhaustively without fully understanding the performance characteristics of their applications.To provide insights into the performance bottlenecks of parallel applications on GPU architectures, we propose a simple analytical model that estimates the execution time of massively parallel programs. The key component of our model is estimating the number of parallel memory requests (we call this the memory warp parallelism) by considering the number of running threads and memory bandwidth. Based on the degree of memory warp parallelism, the model estimates the cost of memory requests, thereby estimating the overall execution time of a program. Comparisons between the outcome of the model and the actual execution time in several GPUs show that the geometric mean of absolute error of our model on micro-benchmarks is 5.4% and on GPU computing applications is 13.3%. All the applications are written in the CUDA programming language.

/pdf/an-analytical-model-for-a-gpu-architecture-with-memory-level-4930mwitke.pdf

An analytical model for a GPU architecture with memory-level and thread-level parallelism awareness

Recent advances in graphics processing units (GPUs) have resulted in massively parallel hardware that is easily programmable and widely available in commodity desktop computer systems. GPUs typically use single-instruction, multiple-data (SIMD) pipelines to achieve high perfor- mance with minimal overhead incurred by control hard- ware. Scalar threads are grouped together into SIMD batches, sometimes referred to as warps. While SIMD is ideally suited for simple programs, recent GPUs include control flow instructions in the GPU instruction set archi- tecture and programs using these instructions may experi- ence reduced performance due to the way branch execution is supported by hardware. One approach is to add a stack to allow different SIMD processing elements to execute dis- tinct program paths after a branch instruction. The occur- rence of diverging branch outcomes for different processing elements significantly degrades performance. In this paper, we explore mechanisms for more efficient SIMD branch ex- ecution on GPUs. We show that a realistic hardware im- plementation that dynamically regroups threads into new warps on the fly following the occurrence of diverging branch outcomes improves performance by an average of 20.7% for an estimated area increase of 4.7%.

/pdf/dynamic-warp-formation-and-scheduling-for-efficient-gpu-4okrlhtvtn.pdf

Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow

Manycore accelerators such as graphics processor units (GPUs) organize processing units into single-instruction, multiple data “cores” to improve throughput per unit hardware cost. Programming models for these accelerators encourage applications to run kernels with large groups of parallel scalar threads. The hardware groups these threads into warps/wavefronts and executes them in lockstep-dubbed single-instruction, multiple-thread (SIMT) by NVIDIA. While current GPUs employ a per-warp (or per-wavefront) stack to manage divergent control flow, it incurs decreased efficiency for applications with nested, data-dependent control flow. In this paper, we propose and evaluate the benefits of extending the sharing of resources in a block of warps, already used for scratchpad memory, to exploit control flow locality among threads (where such sharing may at first seem detrimental). In our proposal, warps within a thread block share a common block-wide stack for divergence handling. At a divergent branch, threads are compacted into new warps in hardware. Our simulation results show that this compaction mechanism provides an average speedup of 22% over a baseline per-warp, stack-based reconvergence mechanism, and 17% versus dynamic warp formation on a set of CUDA applications that suffer significantly from control flow divergence.

/pdf/thread-block-compaction-for-efficient-simt-control-flow-27mzr8ylan.pdf

Thread block compaction for efficient SIMT control flow

While scalable coherence has been extensively studied in the context of general purpose chip multiprocessors (CMPs), GPU architectures present a new set of challenges. Introducing conventional directory protocols adds unnecessary coherence traffic overhead to existing GPU applications. Moreover, these protocols increase the verification complexity of the GPU memory system. Recent research, Library Cache Coherence (LCC) [34, 54], explored the use of time-based approaches in CMP coherence protocols. This paper describes a time-based coherence framework for GPUs, called Temporal Coherence (TC), that exploits globally synchronized counters in single-chip systems to develop a streamlined GPU coherence protocol. Synchronized counters enable all coherence transitions, such as invalidation of cache blocks, to happen synchronously, eliminating all coherence traffic and protocol races. We present an implementation of TC, called TC-Weak, which eliminates LCC's trade-off between stalling stores and increasing L1 miss rates to improve performance and reduce interconnect traffic. By providing coherent L1 caches, TC-Weak improves the performance of GPU applications with inter-workgroup communication by 85% over disabling the non-coherent L1 caches in the baseline GPU. We also find that write-through protocols outperform a writeback protocol on a GPU as the latter suffers from increased traffic due to unnecessary refills of write-once data.

/pdf/cache-coherence-for-gpu-architectures-3gv0u1ukzc.pdf

Cache coherence for GPU architectures

Graphics processor units (GPUs) are designed to efficiently exploit thread level parallelism (TLP), multiplexing execution of 1000s of concurrent threads on a relatively smaller set of single-instruction, multiple-thread (SIMT) cores to hide various long latency operations. While threads within a CUDA block/OpenCL workgroup can communicate efficiently through an intra-core scratchpad memory, threads in different blocks can only communicate via global memory accesses. Programmers wishing to exploit such communication have to consider data-races that may occur when multiple threads modify the same memory location. Recent GPUs provide a form of inter-block communication through atomic operations for single 32-bit/64-bit words. Although fine-grained locks can be constructed from these atomic operations, synchronization using locks is prone to deadlock. In this paper, we propose to solve these problems by extending GPUs to support transactional memory (TM). Major challenges include supporting 1000s of concurrent transactions and committing non-conflicting transactions in parallel. We propose KILO TM, a novel hardware TM design for GPUs that scales to 1000s of concurrent transactions. Without cache coherency hardware to depend on, it uses word-level, value-based conflict detection to avoid broadcast communication and reduce on-chip storage overhead. It employs speculative validation using a novel bloom filter organization to increase transaction commit parallelism. For a set of TM-enhanced GPU applications, KILO TM captures 59% of the performance of fine-grained locking, and is on average 128x faster than executing all transactions serially, for an estimated hardware area overhead of 0.5% of a commercial GPU.

/pdf/hardware-transactional-memory-for-gpu-architectures-3r7pqlxsfu.pdf

Wilson W. L. Fung

Papers

Analyzing CUDA workloads using a detailed GPU simulator

Dynamic Warp Formation and Scheduling for Efficient GPU Control Flow

Thread block compaction for efficient SIMT control flow

Cache coherence for GPU architectures

Hardware transactional memory for GPU architectures