Compilers and More: GPU Architecture and Applications

By Michael Wolfe

September 10, 2008

One of the most exciting developments in parallel programming over the past few years has been the availability and advancement of programmable graphics cards. A high end graphics card costs less than a high end CPU and provides tantalizing peak performance approaching, or exceeding, one teraflop. Since microprocessor peak performance tops out at about 25 gigaflops/core (single precision), this potential, at such low cost, is worth exploring. Harnessing this performance, however, is problematic.

It’s important to note that the GPUs powering the graphics cards are designed to do specific jobs very well. They are not designed as general purpose processors, and in fact will do a very poor job on many programs, even highly parallel applications. The key is to determine whether your application can fit into a programming model that maps well onto the GPU. I’m going to discuss the GPU architecture, but I’m going to start with an analogy, and probably stretch the analogy to the breaking point; let’s discuss airline travel.

Suppose your job is to transport several dozen large tour groups between London and Seattle, each group with 30-60 members. Your most likely choice is to use jet aircraft for a flight of about 5,000 miles or 7,700 km. Going for the most parallelism, you could use a new Airbus A380 to move 600 people in about 9 hours. One problem you have with these jumbo jets is they don’t fit at the main terminal, so you have to take an airport train out to the remote terminal where the plane is parked; the train can only carry so many people at a time, but let’s be optimistic and say it will take 90 minutes to move everyone out to the remote gate and load them, and another 90 minutes at the other end for unloading. This gives us a 24 hour round trip. If you regularly fill the plane, this could be a good investment (though, at $300M, a bit more than a GPU). However, if you only have a half-load, it doesn’t get them there twice as fast or at half the cost, though it does reduce the load/unload time. Measuring performance as passenger-miles, the performance comes from parallelism (many passengers), not from latency reduction.

Alternatively, you could opt for a few smaller planes, say two to four Boeing 787s. Each can carry about 200-250 passengers at the same speed, so you can move the same number of people at the same rate. You also have the advantage of parking at the main terminal, so you can reduce the passenger load/unload time to about half an hour, so your total round trip time is only 20 hours. Of course, you have to arrange for more crew, more landing slots at the airports, and so forth. Your total capital investment is about the same, but it gives you some flexibility. If you only have 200 people to move, you can leave all but one of the planes behind, saving on fuel and crew costs.

Or, perhaps you could invest in the (future) hypersonic transport, which some believe may be able to travel at speeds of Mach 6, 7-8 times faster than the current subsonic airliners. Assuming it takes time to get up to speed and to slow down for landing, the total flight time might drop from 9 hours to 2.5 hours, or 7 hours round trip. If the capacity of your hypothetical hypersonic transport is 200 passengers, you can transport 600 people in each direction in just 21 hours. Even better, if you only have to transport 400 people, you can get the same work done in two-thirds the time. Of course, you are buying a higher cost transport, probably paying more for fuel, and so on.

Very fast CPUs are more like the hypersonic transport; they are designed to provide extremely high performance for small tasks, but still offer adequate performance for large tasks. Multicore processors are more like the middle option, several smaller, lower-capacity devices, each quite capable, and you can save power by shutting down one or more of them. The GPU is more like the super-jumbo jet; it only gets high performance (passenger-miles) when you have lots of passengers. It doesn’t do so well getting just one or a few passengers across the ocean.

Thread Execution Control Unit

So now to GPU architecture. GPUs were originally hardwired for specific tasks; as transistor budgets and demand for flexibility grew, the hardware became more programmable. They still contain special hardware and functional units specific to graphics tasks, but I’ll ignore those and view today’s GPUs as compute accelerators. A typical design, shown in the figure above, is abstracted from information put out by NVIDIA. I’ll use both NVIDIA’s terms and more standard computer architecture descriptions of the various parts. The key to the performance is all those thread processors; in the figure, there are eight thread processors in each of 16 multiprocessors, for 128 TPs total. NVIDIA delivers GPUs with up to 30 multiprocessors and 240 TPs. In each clock, each TP can produce a result, giving this design a very high peak performance rating.

Each multiprocessor executes in SIMD mode, meaning each thread processor in that multiprocessor executes the same instruction simultaneously. If one thread processor is executing a floating point add, they all are; if one is doing a fetch, they all are. Moreover, each instruction is quad-clocked, meaning the SIMD width is 32, even though there are only eight thread processors. Unlike classical SIMD machines, there isn’t a distinction between the scalar and parallel operations, or mono and poly operations, to borrow terms from C* and Dataparallel C, and Cn for the Clearspeed card. Instead, the model is that of many scalar threads that just happen to be executing in SIMD mode, something NVIDIA calls SIMT execution. Careful orchestration of the 32 threads that execute in SIMD mode is necessary for best performance.

Stretching our analogy, think of each thread processor like a seat in our superjumbo jet, and each multiprocessor as a tour group. Imagine that when the flight attendant serves a meal, the whole tour group must be served at once (synchronously). The whole tour group must watch the same movie on the seat-back screens at the same time, or they all must read books at the same time, though some may be napping at any time. When one person wants to use the restroom, they all must go at once, even if not everyone needs to. Note that this is a long latency operation, which will correspond to the long latency memory.

The multiprocessors themselves execute asynchronously, and without communication. This last point is quite important. In a multicore or multiprocessor system, the cores or processors can communicate through the memory. If one thread stores a value in variable A then sets a FLAG, the hardware will guarantee that another thread on the same or another core or processor will not see the updated FLAG without seeing the updated value for A. The hardware supports a memory model that preserves the store order. No such memory model is supported on the GPU; a program could store a set of values on one multiprocessor and read the same locations on another, but there is no guarantee that the value fetched will be consistent (in the formal sense) with the values stored. Relaxing the memory model allows the hardware to reorder the stores from a multiprocessor, allowing more throughput.

In this GPU, each multiprocessor has a special function unit, which handles infrequent, expensive operations, like divide, square root, and so on; it operates more slowly than other operations, but since it’s infrequently used, it doesn’t affect performance. There is a high bandwidth, low latency local memory attached to each multiprocessor. The threads executing on that multiprocessor can communicate among themselves using this local memory. In the current NVIDIA generation, the local memory is quite small (16KB).

There is also a large global device memory (over 1GB in some models). This is physical, not virtual; paging is not supported, so all the data has to fit in the memory. The device memory has very high bandwidth, but high latency to the multiprocessors. The device memory is not directly accessible from the host, nor is the host memory directly visible to the GPU. Data from the host that needs to be processed by the GPU must be moved via DMA (across the IO bus) from the host to the device memory, and the results moved back — like loading jumbo jets using thin airport trains.

So, there is a hierarchy of parallelism on this GPU; threads executing within a multiprocessor can share and communicate using the local memory, while threads executing on different multiprocessors cannot communicate or synchronize. This hierarchy is explicit in the programming model as well. Parallelism comes in two flavors; outer, asynchronous parallelism between thread groups or thread blocks, and inner, synchronous parallelism within a thread block. All the threads of a thread block will always be assigned as a group to a single multiprocessor, while different thread blocks can be assigned to different multiprocessors.

Because of the high latency to the device memory, the multiprocessors are highly multithreaded as well. When one set of SIMD threads executes a memory operation, rather than stall, the multiprocessor will switch to execute another set of SIMD threads. The other SIMD threads may be part of the same thread block, or may come from a different thread block assigned to the same multiprocessor. Think of this as the flight attendant serving another tour group while one group gets up to visit the restrooms.

This GPU is designed as a throughput engine; it’s designed to get a lot of work done on a lot of data, all very quickly, all in parallel (lots of tour groups flying in the same direction). To get high performance, your program needs enough parallelism to keep the thread processors busy (lots of customers). Each of your thread blocks needs enough parallelism to fill all the thread processors in a multiprocessor (big tour groups), and at least as many thread blocks as you have multiprocessors (many big tour groups). In addition, you need even more parallelism to keep the multiprocessors busy when they need to thread-switch past a long latency memory operation (many many big tour groups).

However, lots of parallelism isn’t quite sufficient. The GPU is designed for structured data accesses, which graphic processing naturally uses. (Here’s where my analogy falls apart.) The memory is designed for efficient access to contiguous blocks of memory, meaning adjacent threads using adjacent data. Your program will run noticeably slower if it does random memory fetches, or (shades of vector processing) simultaneous accesses to the same memory bank. But if your program is structured to use contiguous data, the memory can run at full bandwidth.

The program model reflects this structure as well. Consider a simple graphics shader; you want to compute the color at each pixel on the screen. Generally, the color at each pixel is more or less independent of the color at every other pixel (natural parallelism), and the problem has a natural index set (the two-dimensional pixel coordinate). So the program model is a scalar program to be executed at each pixel coordinate, replicated and parallelized over the index set. The scalar program is a kernel and the index set is the domain. The indices for each instance of the kernel determine which pixel is being processed and are used to fetch and store data.

When you write a general purpose program for a GPU, you must currently follow this model as well. You must split out the computational kernels and the corresponding domains. You have to identify which data will live in the local memory, and move data between the local and device memory, keeping in mind that the local memory data only has a lifetime of a thread block.

Each kernel will complete its entire domain before the next kernel starts. In our airline analogy, you have to complete the eastbound flight for all the tour groups on board before starting the next flight westbound. The parallelism comes from the domain of a single kernel (data parallelism, all tour groups flying East at the same time), not from running many different kernels at once (task parallelism, lots of tour groups flying all over the country or world at once). This naturally limits, or focuses users on, the applications for which a GPU is appropriate.

Another key difference between GPUs and more general purpose multicore processors is hardware support for parallelism. GPUs don’t try to address all the possible forms of parallelism, but they do solve their target range quite well. We’ve already mentioned the SIMD instruction sets within a multiprocessor and hardware multithreading. There is also a hardware thread control unit that manages the distribution and assignment of thread blocks to multiprocessors. There is additional hardware support for synchronization within a thread block. Your common multicore processor depends on software and the OS to provide these features, so GPU has an advantage.

Programming GPUs Today

So far I haven’t really discussed how these highly parallel GPU architectures are programmed. Past accelerators were often programmed by offloading functions or subroutines; the user or compiler would marshall the arguments, send them to the accelerator, launch the function on the accelerator, and wait for completion, perhaps doing other useful work in the meantime. GPUs don’t fit this model; they aren’t fully functional, separately programmable devices. They really can only execute kernels, comprising a scalar kernel program and an index domain over which to apply it. If your function were that simple (matrix multiplication, SGEMM), it could look like a subroutine engine. Anything more complex, and the host has to manage the GPU execution by selecting and ordering a sequence of kernels to execute, and performing any scalar operations and conditionals along the way. In my next column, I’ll discuss the sometimes superhuman efforts necessary to decompose a program into kernels and manage the data movement between the host and the GPU.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

HPE to Acquire Cray for $1.3B

May 17, 2019

Venerable supercomputer pioneer Cray Inc. will be acquired by Hewlett Packard Enterprise for $1.3 billion under a definitive agreement announced this morning. The news follows HPE’s acquisition nearly three years ago o Read more…

By Doug Black & Tiffany Trader

China Establishes Seventh National Supercomputing Center

May 16, 2019

Chinese media is reporting that China will construct a new National Supercomputer Center in Zhengzhou, in central China's Henan Province. The new Zhengzhou facility will house a 100-petaflops supercomputer and will be ta Read more…

By Staff report

Interview with 2019 Person to Watch Ken King

May 16, 2019

Today, as the final installment of our HPCwire People to Watch focus series, we present our interview with Ken King, general manager of OpenPOWER for the IBM Systems Group. Ken is responsible for building and managing t Read more…

By HPCwire Editorial Team

HPE Extreme Performance Solutions

HPE and Intel® Omni-Path Architecture: How to Power a Cloud

Learn how HPE and Intel® Omni-Path Architecture provide critical infrastructure for leading Nordic HPC provider’s HPCFLOW cloud service.

For decades, HPE has been at the forefront of high-performance computing, and we’ve powered some of the fastest and most robust supercomputers in the world. Read more…

IBM Accelerated Insights

Autonomous Vehicles: New challenges for the CAE Data Center

Managing infrastructure complexity in the age of AI

When most of us hear the term autonomous vehicles, we conjure up images of driverless Waymos or robotic transport trucks driving long-haul highway routes. Read more…

What’s New in HPC Research: Image Classification, Crowd Computing, Genome Informatics & More

May 15, 2019

In this bimonthly feature, HPCwire highlights newly published research in the high-performance computing community and related domains. From parallel programming to exascale to quantum computing, the details are here. Read more…

By Oliver Peckham

HPE to Acquire Cray for $1.3B

May 17, 2019

Venerable supercomputer pioneer Cray Inc. will be acquired by Hewlett Packard Enterprise for $1.3 billion under a definitive agreement announced this morning. T Read more…

By Doug Black & Tiffany Trader

Deep Learning Competitors Stalk Nvidia

May 14, 2019

There is no shortage of processing architectures emerging to accelerate deep learning workloads, with two more options emerging this week to challenge GPU leader Nvidia. First, Intel researchers claimed a new deep learning record for image classification on the ResNet-50 convolutional neural network. Separately, Israeli AI chip startup Hailo.ai... Read more…

By George Leopold

CCC Offers Draft 20-Year AI Roadmap; Seeks Comments

May 14, 2019

Artificial Intelligence in all its guises has captured much of the conversation in HPC and general computing today. The White House, DARPA, IARPA, and Departmen Read more…

By John Russell

Cascade Lake Shows Up to 84 Percent Gen-on-Gen Advantage on STAC Benchmarking

May 13, 2019

The Securities Technology Analysis Center (STAC) issued a report Friday comparing the performance of Intel's Cascade Lake processors with previous-gen Skylake u Read more…

By Tiffany Trader

Nvidia Claims 6000x Speed-Up for Stock Trading Backtest Benchmark

May 13, 2019

A stock trading backtesting algorithm used by hedge funds to simulate trading variants has received a massive, GPU-based performance boost, according to Nvidia, Read more…

By Doug Black

ASC19: NTHU Returns to Glory

May 11, 2019

As many of you Student Cluster Competition fanatics know by now, Taiwan’s National Tsing Hua University (NTHU) won the gold medal at the recently concluded AS Read more…

By Dan Olds

Intel 7nm GPU on Roadmap for 2021, OneAPI Coming This Year

May 8, 2019

At Intel's investor meeting today in Santa Clara, Calif., the company filled in details of its roadmap and product launch plans and sought to allay concerns about delays of its 10nm chips. In laying out its 10nm and 7nm timelines, Intel revealed that its first 7nm product would be... Read more…

By Tiffany Trader

Ten Great Reasons to Build the 1.5 Exaflops Frontier

May 7, 2019

It’s perhaps obvious that the fundamental reason for building expensive exascale computers is to drive science and industry forward, realizing the resulting b Read more…

By John Russell

Cray, AMD to Extend DOE’s Exascale Frontier

May 7, 2019

Cray and AMD are coming back to Oak Ridge National Laboratory to partner on the world’s largest and most expensive supercomputer. The Department of Energy’s Read more…

By Tiffany Trader

Graphene Surprises Again, This Time for Quantum Computing

May 8, 2019

Graphene is fascinating stuff with promise for use in a seeming endless number of applications. This month researchers from the University of Vienna and Institu Read more…

By John Russell

Why Nvidia Bought Mellanox: ‘Future Datacenters Will Be…Like High Performance Computers’

March 14, 2019

“Future datacenters of all kinds will be built like high performance computers,” said Nvidia CEO Jensen Huang during a phone briefing on Monday after Nvidia revealed scooping up the high performance networking company Mellanox for $6.9 billion. Read more…

By Tiffany Trader

ClusterVision in Bankruptcy, Fate Uncertain

February 13, 2019

ClusterVision, European HPC specialists that have built and installed over 20 Top500-ranked systems in their nearly 17-year history, appear to be in the midst o Read more…

By Tiffany Trader

It’s Official: Aurora on Track to Be First US Exascale Computer in 2021

March 18, 2019

The U.S. Department of Energy along with Intel and Cray confirmed today that an Intel/Cray supercomputer, "Aurora," capable of sustained performance of one exaf Read more…

By Tiffany Trader

Intel Reportedly in $6B Bid for Mellanox

January 30, 2019

The latest rumors and reports around an acquisition of Mellanox focus on Intel, which has reportedly offered a $6 billion bid for the high performance interconn Read more…

By Doug Black

Looking for Light Reading? NSF-backed ‘Comic Books’ Tackle Quantum Computing

January 28, 2019

Still baffled by quantum computing? How about turning to comic books (graphic novels for the well-read among you) for some clarity and a little humor on QC. The Read more…

By John Russell

The Case Against ‘The Case Against Quantum Computing’

January 9, 2019

It’s not easy to be a physicist. Richard Feynman (basically the Jimi Hendrix of physicists) once said: “The first principle is that you must not fool yourse Read more…

By Ben Criger

Leading Solution Providers

SC 18 Virtual Booth Video Tour

Advania @ SC18 AMD @ SC18
ASRock Rack @ SC18
DDN Storage @ SC18
HPE @ SC18
IBM @ SC18
Lenovo @ SC18 Mellanox Technologies @ SC18
NVIDIA @ SC18
One Stop Systems @ SC18
Oracle @ SC18 Panasas @ SC18
Supermicro @ SC18 SUSE @ SC18 TYAN @ SC18
Verne Global @ SC18

Deep500: ETH Researchers Introduce New Deep Learning Benchmark for HPC

February 5, 2019

ETH researchers have developed a new deep learning benchmarking environment – Deep500 – they say is “the first distributed and reproducible benchmarking s Read more…

By John Russell

Deep Learning Competitors Stalk Nvidia

May 14, 2019

There is no shortage of processing architectures emerging to accelerate deep learning workloads, with two more options emerging this week to challenge GPU leader Nvidia. First, Intel researchers claimed a new deep learning record for image classification on the ResNet-50 convolutional neural network. Separately, Israeli AI chip startup Hailo.ai... Read more…

By George Leopold

IBM Bets $2B Seeking 1000X AI Hardware Performance Boost

February 7, 2019

For now, AI systems are mostly machine learning-based and “narrow” – powerful as they are by today's standards, they're limited to performing a few, narro Read more…

By Doug Black

Arm Unveils Neoverse N1 Platform with up to 128-Cores

February 20, 2019

Following on its Neoverse roadmap announcement last October, Arm today revealed its next-gen Neoverse microarchitecture with compute and throughput-optimized si Read more…

By Tiffany Trader

Intel Launches Cascade Lake Xeons with Up to 56 Cores

April 2, 2019

At Intel's Data-Centric Innovation Day in San Francisco (April 2), the company unveiled its second-generation Xeon Scalable (Cascade Lake) family and debuted it Read more…

By Tiffany Trader

France to Deploy AI-Focused Supercomputer: Jean Zay

January 22, 2019

HPE announced today that it won the contract to build a supercomputer that will drive France’s AI and HPC efforts. The computer will be part of GENCI, the Fre Read more…

By Tiffany Trader

In Wake of Nvidia-Mellanox: Xilinx to Acquire Solarflare

April 25, 2019

With echoes of Nvidia’s recent acquisition of Mellanox, FPGA maker Xilinx has announced a definitive agreement to acquire Solarflare Communications, provider Read more…

By Doug Black

Nvidia Claims 6000x Speed-Up for Stock Trading Backtest Benchmark

May 13, 2019

A stock trading backtesting algorithm used by hedge funds to simulate trading variants has received a massive, GPU-based performance boost, according to Nvidia, Read more…

By Doug Black

  • arrow
  • Click Here for More Headlines
  • arrow
Do NOT follow this link or you will be banned from the site!
Share This