[Future Technology Research Index] [SGI Tech/Advice Index] [Nintendo64 Tech Info Index]

[WhatsNew] [P.I.] [Indigo] [Indy] [O2] [Indigo2] [Crimson] [Challenge] [Onyx] [Octane] [Origin] [Onyx2]

Ian's SGI Depot: FOR SALE! SGI Systems, Parts, Spares and Upgrades

(check my current auctions!)

Octane Architecture

Last Change: 20/Mar/2006

Like O2, Octane does not use a bus design, but there the similarities end. Octane employs the technology used in the high-end Origin systems and its features reflect this.

A multiprocessor system, Octane supports one or two R10000, R12000 or R14000 CPUs. There is a range of possible clock speeds and cache sizes. The ones I know of are:

R10000 175MHz (1MB L2)
R10000 195MHz (1MB L2)
R10000 225MHz (1MB L2)
R10000 250MHz (1MB L2)
R10000 250MHz (2MB L2)
R12000 270MHz (2MB L2)
R12000 300MHz (2MB L2)
R12000 360MHz (2MB L2)
R12000 400MHz (2MB L2)
R14000 550MHz (2MB L2)
R14000 600MHz (2MB L2)

I don't think the R16K/700, or anything after that, was made available for Octane, though many have had success clocking-up existing modules, eg. a 360 can be increased to run at 425.

Original early Octanes typically used either a 175 or 195MHz R10K. Such systems often had SI graphics and were very popular for CAD design. Later, such systems typically had an R10K/250 with SE graphics instead. Note that for dual processor systems, both processors always run at the same clock speed.

Later, SGI released a newer motherboard to support faster processors and increased RAM capacity (up to 8GB instead of the original 2GB limit).

Importantly, unlike many dual processor systems (especially PCs), Octane's design permits maximum exploitation of both CPUs at the same time, ie. two processes running in parallel will not interfere with each other due to the large memory, subsystem and I/O bandwidth that is available in the system. Both can be running flat out without having to compete for resources. Later in Octane's life cycle, SGI increased the speed of the frontplane board in order to ensure this would still be possible even when using the faster CPUs and top-end VPro graphics.

The heart of Octane is an 8-port non-blocking high-speed crossbar switch, completely replacing the old bus design. This crossbar chip is the same device used in the high-end Origin line, the only difference being that Octane does not have support for connecting multiple Octanes together CrayLink-style. In Origin, one port of the crossbar is connected to a second HUB chip on another node board (a node board has one or two CPUs) and is partly the means by which Origin can offer more than two CPUs in a system. Thus, in Octane, this port is not used.

Note: the above means that if you want larger multiprocessing capability combined with fast graphics, you should use Tezro or suitable Onyx systems. Remember that a 2nd-hand quad-CPU Onyx2 might be cheaper than a Tezro.

Here is a diagram describing the internal structure of Octane, though note this diagram does not reflect the later CPUs and graphics options:

[Octane Structural Diagram]

Earlier graphics boards produced for Octane used IMPACT-based (MGRAS) technology, but the new architecture combined with improvements in key ASICs provides great performance improvements and enhancements over Indigo2 IMPACT systems, especially in certain areas. In the first half of 1998, SGI released new versions of the geometry and texturing ASICs, giving 15% faster texture performance, 40% faster geometry performance, and faster texture data loading. These options were given 'E' versions of the original names to reflect the improvements as follows:

Original Option      Improved Option

      SI                    SE
 SI + Texture          SE + Texture
     SSI                   SSE
     MXI                   MXE

Octane has three 3.5" SCSI bays (for disks or other devices), 2 to 3 high-speed XIO slots and - if the PCI card cage option is installed - 3 separate PCI64 slots (2 full-height, 1 half-height). This allows for enormous I/O and connectivity support. Also supplied as standard are 2 high speed serial ports (460Kbaud each), two Ultra FastWide SCSI channels (one internal, one external, each 40MB/sec), a parallel port, 10/100Mbit autosensing Ethernet and analogue, AES serial digital plus ADAT optical digital audio connections. Video capabilities are available as XIO option cards (analogue, digital and compression). Most minor peripherals (CDROM, floppy, DAT, DLT, etc.) are housed in new standardised external cases, although an internal DAT option is also available. Note that although Octane's internal SCSI system is limited to 40MB/sec per channel, one can use an U160 QLOGIC PCI card (model QLA12160) to give much faster bandwidth if required. Various QLOGIC FibreChannel (FC) cards are also supported.

Bandwidth and I/O Reliability

The crossbar in Octane has been designed so that software applications can be guaranteed the bandwidth they need for critical/priority data streams if required, eg. CPU from/to GFX, video from/to RAM. Multiple simultaneous connections through the crossbar can occur and these can be continuously altered. A programmer has the ability to define a data stream as 'priority' - the system then ensures that nothing interferes with the required XIO bandwidth allocation.

This results in better performance, smoother graphics, more steady frame rates, improved interactivity and ensures that CPU and graphics pipelines are always kept busy to their maximum potential. Further, when applications are not using the time slices they have reserved, these spare connection opportunities can still be utilised for other tasks, ensuring that no bandwidth capacity is wasted. The end result is an incredibly responsive system. I was once told that animators describe such good responsiveness as 'snappyness'; fast feedback is essential for animation design, so Octane is ideal for such tasks.

Octane also allows processing subsystems to stream data directly to one another, eg. a video subsystem can stream data straight to disk without having to go via main RAM first. This feature enables Octane to fully take advantage of 'smart' peripheral devices.

The importance of Octane's crossbar is that no subsystem (eg. CPU, graphics, etc.) has to compete with any other subsystem for the bandwidth it needs, ie. each subsystem has its own sustainable 1.2GB/sec connection (1.6GB/sec peak) to the crossbar switch. Actually, these connections consist of two paths, both 800MB/sec (peak): a From link and a To link, rather like separate lanes on a 2-lane highway. This means that data traffic heading in one direction does not conflict with traffic heading in the other direction.

Greatly improved memory latency (about 40% better than Indigo2) combined with the new XIO system results in enormous improvements in memory bandwidth as typical benchmarks such as STREAM can demonstrate, and note that the CRC error checking used by the XIO system is better than ECC or parity. Also, R10000 in Indigo2 only supported one outstanding cache miss, whereas R10000 in Octane supports four - the maximum possible with the current R10000 design. I'm not sure if this was increased again with R12000.


To deliniate from Indigo2, Octane uses a different naming convention when describing the original graphics options. Compared to Indigo2, the names are:
      Indigo2          Older Octanes        Newer octanes

    Solid IMPACT       SI                   SE
    High IMPACT        SI + Texture         SE + Texture
    (not available)    SSI                  SSE
    Max IMPACT         MXI                  MXE

The above should make it obvious that Octane offers a graphics configuration that was not available on Indigo2, namely SSE (formerly SSI when using the older ASICs). SSE offers more than twice the geometry and pixel fill power of Solid IMPACT, but does not include hardware texturing; a hardware texture option can be added to any Octane configuration which does not already have it. Note that 4MB of texture RAM is standard on all Octane configurations that employ the texture option, ie. the 1MB option as used in Indigo2 is not used in Octane at all, which is logical since the TRAM option boards for Octane include the texture mapping hardware aswell as the memory. A minor point: the texture memory used in IMPACT technology is the same kind of memory that's used in the Nintendo64, namely Rambus. It is very high speed RAM (500MHz system event speed, ie. a 250MHz clock driving transfers on both clock edges) and as such one can think of it more as texture cache than merely memory.

With the improved architecture, Octane with SE/SSE/MXE provides for a significant performance improvement over Indigo2/IMPACT systems, a fact which can easily be seen by examining the various benchmark figures.

Octane's design removes the memory and I/O bottlenecks in graphics applications; this is an important step in improving performance, consistency, efficiency and responsiveness for users.

One note on SI, SSI and MXI: it can be easy to assume that these physical graphics boards used in Octane are actually the same as for Indigo2 IMPACT systems, but this isn't true. In fact, the board layouts are different, the voltage levels, track lengths, capacitances, and revisions of various ASICs are all different. It is better to say that SI, SSI and MXI use the same 'level' of graphics technology as their IMPACT equivalents, but Octane can get more out of them.

Some years after initial launch, SGI released a new generation of graphics options for Octane, called VPro. A single-chip GPU design, VPro graphics options offer significant speed advantages over the older MGRAS boards, aswell as more advanced features, higher resolution support and a large increase in available memory. The VPro boards use a single memory pool for both texture data and general video buffers.

The initial VPro boards to be released were V6 and V8. Both run at the same speed, but the latter has considerably more combined total memory (128MB vs 32MB), and also supports more resolutions than V6.

Note that V6/V8 do not support 1280x1024 at 72Hz - a bizarre 'feature' (caused by a timing bug) which means that older monitors, if running at 1280x1024, will have to operate at 60Hz with the board using a reduced frame buffer depth (8 bytes per pixel). To get the most out of a V6 board, use a monitor that can run at 96Hz in order to allow the use of the maximum frame buffer depth, eg. the SGI GDM5411 monitor, or (for V8) use 1600x1200. The later V10/V12 boards do not have this restriction though, ie. they can use 72Hz or other frequencies just fine. When my own system was a V8, I just ran it at 1600x1200 at 75Hz.

In order to use a VPro board, an Octane must have at least XBow 1.3, though even if this is present, the system must also have the later Cherokee PSU. Any Octane which has all its main subsystems of the later type (PSU, XBow and motherboard) can effectively be regared as an Octane2. Thus, any older Octane can be upgraded to an Octane2 by simple part swaps.

SGI later doubled the speed of the geometry system in their VPro GPU design and added some more features and other abilities, especially with respect to dual-channel/dual-display. The later boards, called V10 and V12, run at the same speed, but share the same difference in memory allocation as V6/V8.

Thus, the boards can be summarised as follows:

                     CAD-oriented, fewer    GIS, imaging, heavy texture
                     hw features, not so    oriented, complex animation,
                     much VRAM or TRAM.     more features, dual-head, etc.

Old Series VPro         V6 (32MB RAM)               V8 (128MB RAM)
Original GE speed

New Series VPro         V10 (32MB RAM)              V12 (128MB RAM)
2X Faster GE speed

V6 and V10 can have up to 8MB RAM allocated to textures (2X more than the textured-enabled MGRAS options), while V8 and V12 can have up to 108MB RAM used for textures. As with the MGRAS boards, all VPro boards support the OpenGL ARB imaging extensions, allowing for hardware acceleration of numerous imaging operations at real-time rates. It's also worth noting that the freeware mplayer application can use any VPro board to hardware accelerate video playback, thus allowing fullscreen playback of a full-size/rate PAL DivX file - something O2 cannot do since relevant codecs for ICE were never written for this.

Another simple difference: Quake3 plays very nicely on VPro systems, but is far too slow on an older MGRAS system such as MXI.

Lastly, since V12 is the best graphics option available for Octane,

System Expansion

Octane provides for enormous expansion and I/O capabilities. For example, using XIO and/or PCI cards, Octane can support up to forty digital audio streams (8 AES serial digital and 32 ADAT optical digital audio I/O).

Octane uses XIO-based expansion cards to provide analogue and digital video processing capabilities, thus allowing better price points for those who do not need such features. Available options include Digital Video, Personal Video and Compression.

Octane Digital Video (ODV) provides two 10bit CCIR-601 video I/O streams. These can be configured as real-time connections to/from graphics screen/memory/etc. This interface allows a single stream of full frame-rate video with an alpha or key channel, or two streams of field-rate video to be texture mapped onto a polygonal surface with real-time mipmapping for high image quality. Also included is a high-quality hardware colour space converter, incorporating new constant-hue technology to minimise distortion when converting internally generated RGB images to YUV-based video.

Octane Personal Video (OPV) is a high quality analogue/digital video board. Offering many of the features of the digital video option but at a lower cost (one digital stream instead of two), OPV includes composite I/O connectors, a digital input for use with the bundled O2Cam (this input can be converted to full CCIR 601 I/O via 3rd-party adapator from Miranda Technologies), an audio channel, a reference-timing input, plus the same hardware colour space converter found in the ODV option to provide for conversion of the entire 1280x1024 screen display, or just a portion of it, down to video resolution which can then be streamed to video out and/or main memory.

Octane Compression (OC) provides 2 streams of motion JPEG with compression ratios as low as 2:1 for very high image quality (again employing a hardware colour space converter), a composite/S-Video I/O port and an analogue genlock loop-through, allowing one to compress incoming video or data resident in memory (eg. rendered frames) in real-time. The analogue video hardware can be used to playback these streams for preview or broadcast. OC has a direct connection to ODV (if installed), so two streams of digital 601 video can be independently compressed or decompressed in real-time. Note that one can use MediaRecorder and MediaPlayer to capture or playback full-size/rate video with Octane Compression, but MovieMaker does not know about Octane Compression (or any of the other Cosmo boards) and so cannot be used to do real-time editing (if you want built-in editing with the supplied tools, use O2; eg. one could capture with Octane and then edit with O2).

Video/Audio referencing enables video and audio to be locked for accurate recording, playback, synchronisation and editing.

Octane's PCI expansion slots provide for FDDI, extra 10/100BaseTX Ethernet, ISDN, ATM, HIPPI, Token Ring, Fibre Channel, UltraSCSI and extra audio systems (refer to www.sgi.com for availability).

Octane's XIO slots can take advantage of the XIO cards used on Origin2000 and Onyx2 via the use of a simple mechanical fastener. The XIO slots can utilise various video expansion options, dual head configurations, presenter panels, multi-Ethernet banks (4-port), multi-UltraSCSI banks (4-port), dual-FibreChannel, HIPPI (1-port), multi-ATM OC3 (4-port) and ATM OCM12 (1-port) expansion options.


I think many people were expecting Octane to be brand new in every respect (architecture, graphics, maybe even new CPUs) but with hindsight that was too much to ask and would have been too big a change at the time. It cannot be emphasised enough that Octane is an important step in workstation development. When one looks at older systems that use CPUs such as the R10000, it is very clear that memory bandwidth, I/O bandwidth and memory latency have become critical in real-world application performance as opposed to synthetic benchmark results.

Looking at other vendors' systems then, it's obvious that older bus-based workstations containing fast main-CPUs cannot deliver the kind of performance that the main CPU is theoretically capable of providing (eg. 500MHz Alpha) - at the end of the day, this amounts to wasted money. It's a key reason why, even though the R10000 is not the best CPU available for any single SPEC95 test, SGI consistently does so well on multi-CPU metrics.

Octane has shown that polygons/sec and main CPU power aren't enough when it comes to delivering high-performance graphics workstations. The foundations that the graphics and CPU systems rely on are just as important too.

Ian's SGI Depot: FOR SALE! SGI Systems, Parts, Spares and Upgrades

(check my current auctions!)
[WhatsNew] [P.I.] [Indigo] [Indy] [O2] [Indigo2] [Crimson] [Challenge] [Onyx] [Octane] [Origin] [Onyx2]
[Future Technology Research Index] [SGI Tech/Advice Index] [Nintendo64 Tech Info Index]