PC Desktop Video Revolution: 
    SGI Visual Computers (11/99)

    by Douglas Dixon

    Processor, Bus, Graphics, Throughput

It is an axiom that computers are getting faster, cheaper, and more powerful. And yet working with video on a PC is still much too painful. Video is big and fast and endless: each single frame is like a large image file, and the frames stream in thirty times a second, pouring in like an endless flood, so fast that the processor can't keep up, swamping the bandwidth of the PC's bus, and filling up the capacity of the hard disk.

But there is hope for the future, and we can get a glimpse of it from the new Silicon Graphics (SGI) NT-based Visual Workstations. With these new products, SGI has brought its expertise in high-performance Unix workstations for 3-D graphics and visualization to the PC / Windows NT market. And in these products we can see the kinds of performance enhancements and video processing capabilities that hold promise for the future of mass-market PC's for the rest of us.

While SGI is primarily known for its Unix workstations with hot 3-D graphics performance, especially for special effects in the movies, it has also developed a strong reputation for scientific visualization of both 3-D graphics and imagery. The challenge for SGI in moving to the PC platform was to adapt to the requirements of the PC hardware and software architecture, where backward compatibility and legacy design impose fundamental limitations on the ability to boost performance.

We'll examine SGI's design approaches in four major areas to see the promise for future mainline PC's: processor performance, bus architecture, graphics performance, and video throughput. Just to give you a taste of the future, in our testing we were able to run video processing applications and benchmarks at about 4X the speed of typical business PC's, which means that instead of having to work with 320 x 240 video for reasonable performance, we were able to process 640 x 480 video at around the same speed.

The SGI Visual Workstations

The current SGI Visual Workstation product line includes two models, the Silicon Graphics 320, which supports up to two co-processors, and the Silicon Graphics 540, which supports up to four co-processors. Both share the same basic design, with an integrated video / graphics memory architecture running under Microsoft Windows NT, but the 540 is more expandable.

The 320 starts at $3695 (list price) with a single Intel Pentium III 450 MHz processor, 128 MB RAM, and a 6.4 GB Ultra ATA hard disk (but no monitor). The 540 starts at $6495, but with a single Pentium III Xeon 500 MHz processor, 128 MB RAM, and a 9.1 GB Ultra SCSI hard disk (but no monitor). For more power, you can step up to a dual-processor 540 at $9952, with dual Pentium III Xeon 550 MHz processors, 512 MB RAM, 9.1 GB Ultra2 SCSI hard disk at 10,000 RPM, and a 17" Triniton monitor.

    SGI Visual Workstation

What sets these systems apart from standard PC's which similar specifications is their integrated design for high-performance multimedia processing. Both systems have the same "Integrated Visual Computing" architecture with the SGI Cobalt graphics chipset, integrated 10/100 Fast Ethernet, and a plethora of I/O ports for additional devices, audio, and video, including serial, parallel, USB for mouse and keyboard, and an additional USB connector (to be supported in a future release of Windows NT).

The Visual Workstations also support a cool new flat panel display, the 17.3" Silicon Graphics 16000SW digital flat panel monitor, for an additional $2494. The monitor provides wide 1600 x 1200 resolution with bright colors and strong contrast, and, of course, is a lot lighter and cooler than a traditional CRT monitor. SGI also includes a built-in ColorLock color calibration system that manages the adjustment, measurement, and setting of the display to ensure consistency between screen displays and printed output. This monitor is also available separately for any PC, bundled with a Number Nine digital graphics card for around $2549 (street price).

Processor Performance

The first step in accelerating graphics and digital media performance in the PC architecture is the main processor. You want raw speed for running instructions, an efficient instruction set for the operations you are trying to perform, and the ability to scale up performance with more processors.

Raw speed comes in two ways, through the clock speed, and from fast access to memory for instructions and data. In this area, PC and workstation vendors like SGI can just ride the performance curve driven by Intel and its competitors. Typical processor speeds have increased from 200 to 450 MHz (mega hertz) in 1988, to 500, 550 and then to 600 MHz and beyond in 1999, and are threatening to go over 1000 MHz in 2000. Along with the speed increases have come integrated L1 and L2 caches for fast memory access, now growing from 512 KB up to 1 to 2 MB of memory for the cache alone.

The basic Intel processor instruction set provided both integer and floating-point operations for typical arithmetic operations. With the advent of audio and video applications, and the desire to reduce hardware costs by performing more processing on the host processor, came the demand to provide DSP (Digital Signal Processor) operations like compression and filtering, and to perform them efficiently and in parallel. In the Pentium II design, Intel integrated an additional MMX instruction set, which provided 57 new integer instructions designed for this kind of processing of image pixels and sound samples. Because the MMX instructions can process up to four data values at the same time, they can theoretically provide a 4 times speed-up in throughput.

The next bottleneck was then the 3-D geometry operations for interactively calculating the appearance of 3-D models by manipulating their geometry and perspective from the current point of view. In the Pentium III design, Intel introduced a new set of Streaming SIMD Extensions for floating-point arithmetic. SIMD stands for Single Instruction Multiple Data, which means the processor also can perform the same operation on multiple data values at the same time.

Of course, you only get these speed-ups if your application performs these types of operations, if the rest of the processing is not slowing you down more, and if your software is written to take advantage of these new instructions when they are available. In developing these new workstations, SGI developed graphics drivers so that Windows applications could take advantage of the higher performance from both the Intel processors and the SGI Cobalt graphics chip.

Bus Architecture

The next step in achieving high performance on the PC platform is to work on the memory and bus architecture to make sure that data can flow rapidly through the system. For compatibility reasons, the PC bus speed has been a major drag on system performance, moving slowly from 66 MHz to 100 MHz and recently to 133 MHz. Intel has attempted to work around this by introducing the AGP (Accelerated Graphics Port) interface for faster transfers between the processor and video display memory. But AGP provides only a 2X speed-up, moving to 4X this year, which still slows down the processor. In the Visual Workstations, SGI has introduced a high-speed bus which runs at 3.2 GB per second, and accesses main memory six times faster than an AGP 2X bus.

But it's not enough to have data flowing quickly between the processor and memory, you also need to move data to hard disk and to other devices. Older PC's used the ISA (Industry Standard Architecture) bus, which was only 8 bits wide and transferred data at only 8 MB/sec. Newer PC's have moved to the PCI (Peripheral Component Interconnect) bus, which is 32-bit and transfers at 133 MB/sec. Most current PC's still have one or two ISA expansion slots for compatibility reasons.

The Visual Workstations use a 64-bit PCI bus design that transfers data at 200 MB/sec. For additional performance, the PCI bus is split into two independent groups, supporting high-speed disks, networks, and external devices without bus contention. In these workstations, SGI abandoned any ISA compatibility, and provided six PCI slots for add-in cards in the 540, and three slots in the 320. These are the new PCI v2.1 Universal card slots at 3.3V, which may not be compatible with existing PCI cards.

On the 540, SGI also moved from ATA hard disks to the SCSI Ultra2 interface, providing an expandable high-speed interface at up to 80 MB/sec, and supporting disks of sizes up to 18 GB, spinning at 7,200 to 10,000 RPM (revolutions per minute).

Graphics Performance

Graphics performance on mainline PC's is driven by video games, and especially 3-D point-of-view games in which you guide your virtual character through a texture-mapped 3-D environment. Hot graphics performance requires graphics accelerator cards, from companies like ATI, Diamond, and 3DFX, with names like Rage Fury, Viper, Stealth and Voodoo. These cards off-load the graphics processing to a separate board with a custom graphics processor with built-in 2-D and 3-D graphics drawing capabilities, including texture mapping and video processing, at rates around 6 million triangles per second. These cards also include their own fast video memory, growing from 8 to 16 to 32 MB, to support high-resolution deep-color displays of up to around 1900 x 1200. However, the data to be displayed must still be shoveled to these cards over the relatively slow PCI or AGP bus, so their performance depends on how well the application software has been optimized to take advantage of their capabilities.

For the Visual Workstations, SGI used a scalable Integrated Visual Computing (IVC) architecture that shares the system memory for graphics textures and display. Instead of having a fixed dedicated graphics memory, system memory can be dynamically allocated for graphics operations to accommodate specific applications and different formats. Based on its experience with the custom graphics chips used in its high-end Unix workstations, SGI designed the Cobalt graphics chipset, which combines graphics with the memory and I/O controller for fast graphics and multimedia performance. The Cobalt chipset provides hardware-accelerated 3-D primitives, with texture, lighting, clipping, fog, and blending, in a variety of pixel and color formats. It draws filled, shaded, z-buffered 3-D triangles at a rate of 7.4 million triangles per second, and textured pixels at up to 176 million per second.

This kind of graphics performance is not just good news for the game players with their 3-D shooters, but it's also great news for video types as well. After all, textures are just images that get mapped onto 3-D geometry, to make more realistic brick walls and carpeted floors, or paintings on the walls. And if you can manipulate texture images to brighten and resize and blend them, then you can do the same things with moving images too, providing high-performance video processing and filtering.

Video Throughput

Finally, even if you have the fastest processor, and high-speed internal buses, and great graphics performance, your system still is not very useful unless you can get video (and audio) in to and out of the box. But these days video comes in lots of different formats: analog and digital, raw and compressed, consumer and professional. To get video into your PC, you basically have to pick a format and buy the right kind of hardware for that format. This may be a PCI card to capture analog video from a camcorder and store it uncompressed on the hard disk; which requires big fast disks to store it all, but gives you the most flexibility in further processing. Or it may be a Motion JPEG (M-JPEG) capture device that captures and compresses analog video in hardware; which gives a big savings in data rate and size, but assumes that you want to work with the M-JPEG compressed format. Or you may have a new DV (Digital Video) camcorder that records directly in the DV digital compressed format; which requires that you install a 1394 (FireWire) interface card to get the data into your computer, and use DV-based video editing software.

The SGI Visual Workstations show the future of PC video by providing support for all these formats, and more. They provide built-in analog video input and output (composite and s-video) for uncompressed capture and playback. They also include a built-in IEEE 1394 (FireWire) port for transfers of up to 400 MB/sec (to be supported in a future release of Windows NT). For M-JPEG use, the workstations support an M-JPEG compression card that provides dual-stream Motion-JPEG compression and/or decompression in real time. Finally, the 540 also supports a Serial Digital Video Option for 2 input and 2 output streams of D1 CCIR 601 digital video (BNC).

    SGI Visual Workstations, standard audio / video connectors

For audio processing, the SGI workstations provide CD-quality (16-bit 44.1 KHz) stereo line input and output (RCA connector), as well as microphone and speaker mini-jacks. SGI also offers a digital audio card that provides up to 10 channels of 24-bit digital audio per card. With multiple cards, the 320 workstation can support up to 30 channels, and the 540 up to 60 channels of synchronized digital I/O.

In contrast to traditional PC architectures that support these kinds of function with add-in cards that must fight for bandwidth on the slower PCI and AGP buses, SGI developed an integrated input / output co-processor that integrates analog and digital video processing on the motherboard, and provides much higher bandwidth, up to 12 times faster than 32-bit PCI systems.

The 540 workstation has the system bandwidth to support up to four streams of uncompressed CCIR-601 professional video in real time (which requires an external RAID disk array for multiple streams), and the 320 can support up to two.

The PC Video Future

The Silicon Graphics Visual Workstations provide a convincing demonstration that real-time processing of full-resolution full-rate video is possible on the PC platform, even with the limitations of the PC platform. By combining fast processors, high-speed memory and bus interconnects, hardware graphics (and video) acceleration, and multiple high-performance video interfaces, these machines can acquire, process, store, and display real-time motion video and audio.

With the continued rapid advance of technology in the highly-competitive PC market, these kinds of advances will soon be tricking down to an off-the-shelf PC near you. As soon as 2000, we will see 1 GHz processors, 400 MHz bus speeds with 3 GB/sec memory throughput, built-in 4X AGP and 1394 video buses, and at least another doubling of disk sizes and access speeds. You can see companies like Sony moving in this direction with products like the VAIO Digital Studio Computer, which includes two built-in iLINK (IEEE 1394) ports.

Of course, with the enhanced ability to process video on a PC will come the desire to do more with it. Instead of 320 resolution, we'll want full 640 or 780 resolution. Instead of a reduced frame rate of 5 or 10 or 15 frames per second, we'll want full 30 fps. Instead of compressing to fit in the available bandwidth and disk size, we'll want to work with clean uncompressed video. And, why limit ourselves to just one stream of video, when we really want to edit together multiple streams, and view the result, all in real time? All of this adds up to more and more demands on our PC, so we'll soon be complaining that we can process only one or two video streams in real time. But that's a nice problem to look forward to.

References

Silicon Graphics Visual Workstations http://visual.sgi.com/