James Laudon
James Laudon is a member of the Google Brain team, whose mission is to develop deep learning technologies and deploy them throughout Google. His research interests focus on hardware and software co-design for high-performance systems and he's currently working on domain-specific computer architectures for machine learning and applying machine learning to system design. Before joining the Brain team in 2017, James was founder and site director for the Google Madison office. Prior to joining Google in 2007 he contributed to the architecture and implementation of multiple computer systems including the Stanford DASH, SGI Origin 2000, and Sun UltraSPARC T1. James has a B.S. in Electrical Engineering from the University of Wisconsin – Madison and a M.S. and Ph.D. in Electrical Engineering from Stanford University.
Authored Publications
Google Publications
Other Publications
Sort By
Graph Transformer: A Generalized Method for Computation Graph Optimizations
Amirali Abdolrashidi
Anna Darling Goldie
Azalia Mirhoseini
Daniel Wong
Hanxiao Liu
Qiumin Xu
Shen Wang
Sudip Roy
(2020)
Preview abstract
Runtime and scalability of neural networks can be significantly affected by computational graph optimization during compilation. Most existing automated graph optimizations are impractical for deployment due to the significant amount of compute required and their inability to generalize to new, previously held-out graphs. To address both limitations, we propose an end-to-end deep reinforcement learning method named Graph Transformer (GTf), based on a scalable sequential attention mechanism over an inductive graph neural network that is transferable to new, unseen graphs. GTf generates decisions on the entire graph in a single-shot fashion, rather than on each individual node progressively, drastically speeding up the search compared to prior methods. Moreover, we propose recurrent attention layers to jointly optimize dependent graph optimization tasks and demonstrate 33%-60% speedup of three graph optimization tasks compared to Tensorflow default optimizations. On a diverse set of representative graphs consisting of 1k-80k nodes, including Inception-v3, Transformer-XL, and WaveNet, GTf achieves an average 21% improvement over human experts and 18% improvement over the prior art with 15x faster convergence, on a device placement task evaluated in real systems.
View details
Apollo: Transferable Architecture Exploration
Albin Jones
Ravi Narayanaswami
Sat Chatterjee
ML for Systems Workshop at NeurIPS 2020
Preview abstract
The looming end of Moore's Law and ascending use of deep learning drives the design of custom accelerators that are optimized for specific neural architectures.
Accelerator design forms a challenging constrained optimization problem over a complex, high-dimensional and structured input space with a costly to evaluate objective function. Existing approaches for accelerator design are sample-inefficient do not transfer knowledge between related optimizations tasks with different design constraints (e.g. area budget) or neural architecture configurations. In this work, we propose a transferable architecture exploration framework, dubbed Apollo, that leverages recent advances in black-box function optimization for sample-efficient accelerator design. We use Apollo to optimize accelerator configurations of a diverse set of neural architectures with alternative design constraints. We show that Apollo finds optimal design configurations more sample-efficiently than baseline approaches. We further show that transferring knowledge between target architectures with different design constraints helps to find optimal configurations faster. This encouraging outcome portrays a promising path forward in shortening the timeline for accelerator design.
View details
In-Datacenter Performance Analysis of a Tensor Processing Unit
Norman P. Jouppi
Cliff Young
Nishant Patil
Gaurav Agrawal
Raminder Bajwa
Sarah Bates
Suresh Bhatia
Nan Boden
Al Borchers
Rick Boyle
Pierre-luc Cantin
Clifford Chao
Chris Clark
Jeremy Coriell
Mike Daley
Matt Dau
Ben Gelb
Tara Vazir Ghaemmaghami
Rajendra Gottipati
William Gulland
Robert Hagmann
C. Richard Ho
Doug Hogberg
John Hu
Dan Hurt
Julian Ibarz
Aaron Jaffey
Alek Jaworski
Alexander Kaplan
Harshit Khaitan
Andy Koch
Naveen Kumar
Steve Lacy
James Law
Diemthu Le
Chris Leary
Zhuyuan Liu
Kyle Lucke
Alan Lundin
Gordon MacKean
Adriana Maggiore
Maire Mahony
Kieran Miller
Rahul Nagarajan
Ravi Narayanaswami
Ray Ni
Kathy Nix
Thomas Norrie
Mark Omernick
Narayana Penukonda
Andy Phelps
Jonathan Ross
ISCA(2017) (to appear)
Preview abstract
Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC---called a Tensor Processing Unit (TPU)---deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs (caches, out-of-order execution, multithreading, multiprocessing, prefetching, ...) that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X - 30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X - 80X higher. Moreover, using the GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.
View details
Throughput-Oriented Multicore Processors
Robert Golla
Greg Grohoski
Multicore Processors and Systems, Springer(2009), pp. 205-230
Preview abstract
Many important commercial server applications are throughput-oriented. Chip multiprocessors (CMPs) are ideally suited to handle these workloads, as the multiple processors on the chip can independently service incoming requests. To date, most CMPs have been built using a small number of high-performance superscalar processor cores. However, the majority of commercial applications exhibit high cache miss rates, larger memory footprints, and low instruction-level parallelism, which leads to poor utilization on these CMPs. An alternative approach is to build a throughput-oriented, multithreaded CMP from a much larger number of simpler processor cores. This chapter explores the tradeoffs involved in building such a simple-core CMP. Two case studies, the Niagara and Niagara 2 CMPs from Sun Microsystems, are used to illustrate how simple-core CMPs are built in practice and how they compare to CMPs built from more traditional high-performance superscalar processor cores. The case studies show that simple-core CMPs can have a significant performance/watt advantage over complex-core CMPs.
View details
Virtual Private Caches
Kyle J. Nesbit
James E. Smith
Proceedings of the 34th Annual International Symposium on Computer Architecture(2007), pp. 57-68
Preview abstract
Virtual Private Machines (VPM) provide a framework for Quality of Service (QoS) in CMP-based computer systems. VPMs incorporate microarchitecture mechanisms that allow shares of hardware resources to be allocated to executing threads, thus providing applications with an upper bound on execution time regardless of other thread activity. Virtual Private Caches (VPCs) are an important element of VPMs. VPC hardware consists of two major components: the VPC Arbiter, which manages shared cache bandwidth, and the VPC Capacity Manager, which manages the cache storage. Both the VPC Arbiter and VPC Capacity Manager provide minimum service guarantees that, when combined, achieve QoS for the cache subsystem. Simulation-based evaluation shows that conventional cache bandwidth management policies allow concurrently executing threads to affect each other significantly in an uncontrollable manner. The evaluation targets cache bandwidth because the effects of cache capacity sharing have been studied elsewhere. In contrast with the conventional policies, the VPC Arbiter meets its QoS performance objectives on all workloads studied and over a range of allocated bandwidth levels. The VPC Arbiter’s fairness policy, which distributes leftover bandwidth, mitigates the effects of cache preemption latencies, thus ensuring threads a high-degree of performance isolation. Furthermore, the VPC Arbiter eliminates negative bandwidth interference which can improve aggregate throughput and resource utilization.
View details
The Coming Wave of Multithreaded Chip Multiprocessors
Lawrence Spracklen
International Journal of Parallel Programming, 35(2007), pp. 299-330
Preview abstract
The performance of microprocessors has increased exponentially for over 35 years. However, process technology challenges, chip power constraints, and difficulty in extracting instruction-level parallelism are conspiring to limit the performance of future individual processors. To address these limits, the computer industry has embraced chip multiprocessing (CMP), predominately in the form of multiple high-performance superscalar processors on the same die. We explore the trade-off between building CMPs from a few high-performance cores or building CMPs from a large number of lower-performance cores and argue that CMPs built from a larger number of lower-performance cores can provide better performance and performance/Watt on many commercial workloads. We examine two multi-threaded CMPs built using a large number of processor cores: Sun's Niagara and Niagara 2 processors. We also explore the programming issues for CMPs with large number of threads. The programming model for these CMPs is similar to the widely used programming model for symmetric multiprocessors (SMPs), but the greatly reduced costs associated with communication of data through the on-chip shared secondary cache allows for more fine-grain parallelism to be effectively exploited by the CMR Finally, we present performance comparisons between Sun's Niagara and more conventional dual-core processors built from large superscalar processor cores. For several key server workloads, Niagara shows significant performance and even more significant performance/Watt advantages over the CMPs built from traditional superscalar processors.
View details
Chip Multiprocessor Architecture: Techniques to Improve Throughput and Latency
Preview abstract
Chip multiprocessors - also called multi-core microprocessors or CMPs for short - are now the only way to build high-performance microprocessors, for a variety of reasons. Large uniprocessors are no longer scaling in performance, because it is only possible to extract a limited amount of parallelism from a typical instruction stream using conventional superscalar instruction issue techniques. In addition, one cannot simply ratchet up the clock speed on today's processors, or the power dissipation will become prohibitive in all but water-cooled systems. Compounding these problems is the simple fact that with the immense numbers of transistors available on today's microprocessor chips, it is too costly to design and debug ever-larger processors every year or two.
CMPs avoid these problems by filling up a processor die with multiple, relatively simpler processor cores instead of just one huge core. The exact size of a CMP's cores can vary from very simple pipelines to moderately complex superscalar processors, but once a core has been selected the CMP's performance can easily scale across silicon process generations simply by stamping down more copies of the hard-to-design, high-speed processor core in each successive chip generation. In addition, parallel code execution, obtained by spreading multiple threads of execution across the various cores, can achieve significantly higher performance than would be possible using only a single core. While parallel threads are already common in many useful workloads, there are still important workloads that are hard to divide into parallel threads. The low inter-processor communication latency between the cores in a CMP helps make a much wider range of applications viable candidates for parallel execution than was possible with conventional, multi-chip multiprocessors; nevertheless, limited parallelism in key applications is the main factor limiting acceptance of CMPs in some types of systems.
After a discussion of the basic pros and cons of CMPs when they are compared with conventional uniprocessors, this book examines how CMPs can best be designed to handle two radically different kinds of workloads that are likely to be used with a CMP: highly parallel, throughput-sensitive applications at one end of the spectrum, and less parallel, latency-sensitive applications at the other. Throughput-sensitive applications, such as server workloads that handle many independent transactions at once, require careful balancing of all parts of a CMP that can limit throughput, such as the individual cores, on-chip cache memory, and off-chip memory interfaces. Several studies and example systems, such as the Sun Niagara, that examine the necessary tradeoffs are presented here. In contrast, latency-sensitive applications - many desktop applications fall into this category - require a focus on reducing inter-core communication latency and applying techniques to help programmers divide their programs into multiple threads as easily as possible. This book discusses many techniques that can be used in CMPs to simplify parallel programming, with an emphasis on research directions proposed at Stanford University. To illustrate the advantages possible with a CMP using a couple of solid examples, extra focus is given to thread-level speculation (TLS), a way to automatically break up nominally sequential applications into parallel threads on a CMP, and transactional memory. This model can greatly simplify manual parallel programming by using hardware - instead of conventional software locks - to enforce atomic code execution of blocks of instructions, a technique that makes parallel coding much less error-prone.
Contents: The Case for CMPs / Improving Throughput / Improving Latency Automatically / Improving Latency using Manual Parallel Programming / A Multicore World: The Future of CMPs
View details
Fair Queuing Memory Systems
Kyle J. Nesbit
Nidhi Aggarwal
James E. Smith
Proceedings of the 39th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO-39), IEEE, Orlando, FL, USA(2006)
Preview abstract
We propose and evaluate a multi-thread memory scheduler that targets high performance CMPs. The proposed memory scheduler is based on concepts originally developed for network fair queuing scheduling algorithms. The memory scheduler is fair and provides Quality of Service (QoS) while improving system performance. On a four processor CMP running workloads containing a mix of applications with a range of memory bandwidth demands, the proposed memory scheduler provides QoS to all of the threads in all of the workloads, improves system performance by an average of 14% (41% in the best case), and reduces the variance in the threads' target memory bandwidth utilization from .2 to .0058.
View details
Maximizing CMP Throughput with Mediocre Cores
John D. Davis
Kunle Olukotun
Proceedings of the 14th International Conference on Parallel Architectures and Compilation Techniques, IEEE Computer Society, Saint Louis, MO, USA(2005), pp. 51-62
Preview abstract
In this paper we compare the performance of area equivalent small, medium, and large-scale multithreaded chip multiprocessors (CMTs) using throughput-oriented applications. We use area models based on SPARC processors incorporating these architectural features. We examine CMTs with in-order scalar processor cores, 2-way or 4-way in-order superscalar cores, private primary instruction and data caches, and a shared secondary cache. We explore a large design space, ranging from processor-intensive to cache-intensive CMTs. We use SPEC JBB2000, TPC-C, TPC-W, and XML Test to demonstrate that the scalar simple-core CMTs do a better job of addressing the problems of low instruction-level parallelism and high cache miss rates that dominate web-service middleware and online transaction processing applications. For the best overall CMT performance, smaller cores with lower performance, so called "mediocre" cores, maximize the total number of CMT cores and outperform CMTs built from larger, higher performance cores.
View details
The SGI Origin 2000: A ccNUMA Highly Scalable Server
Daniel Lenoski
Proceedings of the 24th Annual International Symposium on Computer Architecture, ACM, Denver, CO, USA(1997), pp. 241-251
Preview abstract
The SGI Origin 2000 is a cache-coherent non-uniform memory access (ccNUMA) multiprocessor designed and manufactured by Silicon Graphics, Inc. The Origin system was designed from the ground up as a multiprocessor capable of scaling to both small and large processor counts without any bandwidth, latency, or cost cliffs. The Origin system consists of up to 512 nodes interconnected by a scalable Craylink network. Each node consists of one or two R10000 processors, up to 4 GB of coherent memory, and a connection to a portion of the XIO IO subsystem. This paper discusses the motivation for building the Origin 2000 and then describes its architecture and implementation. In addition, performance results are presented for the NAS Parallel Benchmarks V2.2 and the SPLASH2 applications. Finally, the Origin system is compared to other contemporary commercial ccNUMA systems.
View details
Interleaving: A Multithreading Technique Targeting Multiprocessors and Workstations
Anoop Gupta
Mark Horowitz
Proceedings of the Sixth International Conference on Architectural Support for Programming Languages and Operating Systems, ACM, San Jose, CA, USA(1994), pp. 308 - 318
Preview abstract
There is an increasing trend to use commodity microprocessors as the compute engines in large-scale multiprocessors. However, given that the majority of the microprocessors are sold in the workstation market, not in the multiprocessor market, it is only natural that architectural features that benefit only multiprocessors are less likely to be adopted in commodity microprocessors. In this paper, we explore multiple-context processors, an architectural technique proposed to hide the large memory latency in multiprocessors. We show that while current multiple-context designs work reasonably well for multiprocessors, they are ineffective in hiding the much shorter uniprocessor latencies using the limited parallelism found in workstation environments. We propose an alternative design that combines the best features of two existing approaches, and present simulation results that show it yields better performance for both multiprogrammed workloads on a workstation and parallel applications on a multiprocessor. By addressing the needs of the workstation environment, our proposal makes multiple contexts more attractive for commodity microprocessors.
View details
The DASH Prototype: Logic Overhead and Performance
Daniel Lenoski
Truman Joe
David Nakahira
Luis Stevens
Anoop Gupta
John Hennessy
IEEE Transactions on Parallel and Distributed Systems, 4(1993), pp. 41-61
Preview abstract
The fundamental premise behind the DASH project is that it is feasible to build large-scale shared-memory multiprocessors with hardware cache coherence. The hardware overhead of directory-based cache coherence in a 48-processor is examined. The data show that the overhead is only about 10-15%, which appears to be a small cost for the ease of programming offered by coherent caches and the potential for higher performance. The performance of the system is discussed, and the speedups obtained by a variety of parallel applications running on the prototype are shown. Using a sophisticated hardware performance monitor, the effectiveness of coherent caches and the relationship between an application's reference behavior and its speedup are characterized. The optimizations incorporated in the DASH protocol are evaluated in terms of their effectiveness on parallel applications and on atomic tests that stress the memory system.
View details
The Stanford Dash Multiprocessor
Daniel Lenoski
Kourosh Gharachorloo
Anoop Gupta
John L. Hennessy
Mark Horowitz
Monica S. Lam
IEEE Computer, 25(1992), pp. 63-79
The DASH Prototype: Implementation and Performance
Daniel Lenoski
Truman Joe
David Nakahira
Luis Stevens
Anoop Gupta
John Hennessy
Proceedings of the 19th Annual International Symposium on Computer Architecture, ACM, Queensland, Australia(1992), pp. 92-103
Preview abstract
The fundamental premise behind the DASH project is that it is feasible to build large-scale shared-memory multiprocessors with hardware cache coherence. While paper studies and software simulators are useful for understanding many high-level design trade-offs, prototypes are essential to ensure that no critical details are overlooked. A prototype provides convincing evidence of the feasibility of the design allows one to accurately estimate both the hardware and the complexity cost of various features, and provides a platform for studying real workloads. A 16-processor prototype of the DASH multiprocessor has been operational for the last six months. In this paper, the hardware overhead of directory-based cache coherence in the prototype is examined. We also discuss the performance of the system, and the speedups obtained by parallel applications running on the prototype. Using a sophisticated hardware performance monitor, we characterize the effectiveness of coherent caches and the relationship between an application's reference behavior and its speedup.
View details
Overview and Status of the Stanford DASH Multiprocessor
Daniel Lenoski
Kourosh Gharachorloo
Anoop Gupta
John Hennessy
Proceedings of the International Symposium on Shared Memory Multiprocessing, Tokyo, Japan(1991)
Memory Consistency and Event Ordering in Scalable Shared-Memory Multiprocessors
Kourosh Gharachorloo
Daniel Lenoski
Phillip Gibbons
Anoop Gupta
John Hennessy
Proceedings of the 17th Annual International Symposium on Computer Architecture, ACM, Seattle, WA, USA(1990), pp. 15-26
The Directory-Based Cache Coherence Protocol for the DASH Multiprocessor
Daniel Lenoski
Kourosh Gharachorloo
Anoop Gupta
John Hennessy
Proceedings of the 17th Annual International Symposium on Computer Architecture, ACM, Seattle, WA, USA(1990), pp. 148-159
The ZS-1 Central Processor
James E. Smith
Greg E. Dermer
Brian D. Vanderwarn
Steve D. Klinger
Chris M. Rozewski
Dan L. Fowler
Keith R. Scidmore
Proceedings of the Second International Conference on Architectural Support for Programming Languages and Operating Systems, IEEE, Palo Alto, CA, USA(1987), pp. 199-204
Preview abstract
The Astronautics ZS-1 is a high speed, 64-bit computer system designed for scientific and engineering applications. The ZS-1 central processor uses a decoupled architecture, which splits instructions into two streams---one for fixed point/memory address computation and the other for floating point operations. The two instruction streams are then processed in parallel. Pipelining is also used extensively throughout the ZS-1.This paper describes the architecture and implementation of the ZS-1 central processor, beginning with some of the basic design objectives. Descriptions of the instruction set, pipeline structure, and virtual memory implementation demonstrate the methods used to satisfy the objectives. High performance is achieved through a combination of static (compile-time) instruction scheduling and dynamic (run-time) scheduling. Both types of scheduling are illustrated with examples.
View details