Indigo2/Octane Comparison

Octane's memory latency is 48% less than on Indigo2 (less is better); combined with XIO, this leads to the following basic performance improvements on Octane compared to Indigo2/IMPACT:

Texture loading:  70% faster
Pixel Read/Write: 200% faster
A/W's Maya:       400% faster
Volume Roaming:   500% faster

Note that these improvements do not include the additional speed increases offered by MXE over MXI (40% better geometry system, 15% better texture system), SSE over SSI and SE over SI (40% better geometry system).

Alan Commike of SGI said about volume roaming on Octane:

"You could think of volume roaming as a superset of volume rendering. It means that the whole volume to render is not resident in memory at all times. Each frame, only a Region of Interest is rendered. The roam allows you to change the ROI [Region of Interest] and not drop the frame rate. Kinda like terrain fly-overs, but instead of terrain you might have a 3D seismic survey of the gulf of mexico, a high resolution CT scan, etc.

The Octane architecture allows one CPU to be paging 3D textures in and out of memory, while the other CPU is doing the rendering. The bus bandwidth and I/O capabilities are high enough that you don't take much of a hit moving all the data around the system."

Octane's better memory and I/O design leads to vastly improved SPECfp95 results for the R10000 compared to Indigo2 and newer compiler releases offer better integer performance too (I2 = Indigo2). Here are the single-CPU results:

           I2/175   Octane/175 | I2/195   Octane/195
SPECint95   8.0       8.4 (*)  |  8.9       11.0
SPECfp95    10.3       15.5    |  10.6      17.0

(*) The SPECint95 figure for Octane/175 would actually be higher because of compiler changes, but SGI has not released any updated figures.

Then there are the figures for R10K/250 (as used in Octane/MXE) which has its L2 cache running at 2/3rds core speed:

            Octane/250
SPECint95      13.6
SPECfp95       20.3

There is also the 225MHz R10000 which is used with Octane/SE and Octane/SSE. I do not yet have performance figures for R10K/225. Finally, note the obvious point that Octane can have two CPUs while Indigo2 can only have one.

See my R10000 performance comparison page for further details of processor differences between Indigo2 and Octane. I have a further page on the performance of R10K/250.

The STREAM numbers speak for themselves. Octane leads the field at the moment for desktop workstations.

                        STREAM    
Indigo2 IMPACT/10000:     117  *********
Octane 1x195MHz R10000:   358  ****************************
Sun Ultra 2 (1x200MHz):   210  ****************
Octane 2x195MHz R10000:   515  ****************************************
Sun Ultra 2 (2x200MHz):   305  ***********************
HP C180:                  195  ***************
DEC 500-5 400MHz:         178  **************

Low-end desktop (eg. SolidIMPACT): little difference between Indigo2 and Octane as the task is graphics (gfx) bound (the bottleneck is raw gfx power, not memory bandwidth, etc.) The newer SE graphics option offers improved performance.

High-end desktop (eg. MaxIMPACT): big difference because of XIO and faster memory bandwidth (Octane/SSI is 44% faster than Indigo2 MaxIMPACT - ie. same pixel fill and geometry power, yet Octane is much faster). Hence, SSE and MXE are considerably faster than any Indigo2 configuration.

XIO and memory bandwidth gives Octane a huge lead over all other machines. Octane/SI is 35% faster than Indigo2/SolidIMPACT and Octane/SSI is 55% faster than Indigo2/MaxIMPACT. SE, SSE and MXE offer even greater improvements.

This benchmark uses display lists exclusively and isn't bus bound on Indigo2, hence there is no speed up on Octane SI/SSI/MXI. Rival systems do well on CDRS as a relatively simple wireframe model makes up half the benchmark, though the nature of the CDRS test means rival systems end up with severely skewed and misleading final averages. Note that SPEC has announced that CDRS as a benchmark is now officially out-of-date and will be replaced when an improved benchmark is available.

Obviously, SE/SSE/MXE gives much better results over any equivalent Indigo2 configuration.

As this benchmark uses a relatively small model and relies heavily on lighting and texture mapping, the bottleneck here is graphics and so SI, SSI and MXI are not faster than equivalent Indigo2 configurations (ie. a larger more complex model would certainly show Octane to be faster than Indigo2). Of course, SE, SSE and MXE do offer significant performance improvements.

At the low end of desktop workstations, one is geometry bound and so the speedup on Octane is not great (6% better for Octane/SI+texture compared to Indigo2/HighIMPACT), but at the high-end of the desktop the large models in Design Review can take advantage of Octane's faster memory bandwidth and XIO to give a 32% performance increase with Octane/MXI over Indigo2/MaxIMPACT. MXE offers even better performance.

(these comments refer to SI, SSI and MXI, not SE, SSE and MXE)

For image transfers, pixel read performance on Octane is 200% better than Indigo2, pixel draw performance is 400% better, pixel copy is 300% better and textured image download is 200% better. These improvements are due to the higher XIO bandwidth. This can accelerate some applications by an enormous margin, eg. IL ELT is 50% faster on Octane compared to Indigo2/IMPACT. Anything involving compositing will be faster too.

Dual processor systems will automatically accelerate Performer applications. Some databases have moved from 20Hz update on Indigo2/IMPACT to 30Hz update on Octane. The newer 225MHz and 250MHz R10000 configurations will obviously give improved performance compared to lower-clocked Indigo2 systems.

In general, the larger and more complex a task or model, the greater the improvement of Octane over Indigo2/IMPACT (eg. large models in Maya run at least 67% faster on Octane compared to Indigo2).

Despite the smaller L2 cache in Octane than in Origin2000, Octane's performance on tasks such as engineering analysis (eg. LSDyna) compares very well with Origin: a two-CPU Origin2000 is just 4% faster than two-CPU Octane for LSDyna (single CPU difference is 8%); dual CPU Octane is 60% faster than single CPU Octane for LSDyna.

Other engineering applications on Octane show a 50% improvment over Indigo2. This includes Ansys, Fluent, Nadina and Nastran.

The CPU-intensive aspects of ProEngineer run 42% faster on Octane than on Indigo2/R10K, whilst the graphics aspects of Pro/E are much the same (though SE, SSE and MXE obviously offer improvements). Overall, for the various models used in the Pro/E test, speedups range from 7% to at least 29%. Of course, in the real world, anyone manipulating complex models will get much better performance from Octane compared to Indigo2.

SDRC: Octane/SI runs on average 23% faster than Indigo2/Solid IMPACT, while Octane/MXI runs on average a massive 217% faster compared to Indigo2/MaxIMPACT; larger models show greater performance gains due to XIO, etc.

For CATIA, Octane is between 14% and 72% faster than Indigo2 (I don't have any figures for Octane's newer gfx and CPU options).

For non-parallelised code, dual-CPU Octane gives little improvement over single-CPU Octane (eg. SDRC, CATIA), as one might expect. Bare this in mind when considering a dual-CPU system: don't expect any speedup if your application does not contain support for parallel execution - check with the software vendor. For general compilation, one can take advantage of the Autoparallelising Option, if one is using an appropriate kind of code.

Finally, when SGI hosted their online chat about Octane, I asked the host about software support for dual-processor configurations. This was the reply:

"Most of the applications in engineering analysis (fluid dynamics and structural mechanics), chemistry, visual simulation, and some digital prototyping applications have been parallelized to take advantage of multiple CPUs on Silicon Graphics high-end systems for some time now. OCTANE will inherit all of this work and these applications will take advantage of dual CPUs on OCTANE."