CoolAsAMoose
Newcomer
At my previus work, in the beginning of 1999, we invested in a new visualization system for our in-house flight-simulator. The screen was a spherical one with FOV of 150x40, and the projection system was made up of 3 Barco CRT projectors + 1 xtra projector for the HUD. Edge-blending was used in the borders between the images from the different projectors. We also were to replace our aging Onyx RE2 with something more modern. Although I wanted to go for something PC-based (most likely from Quantum3D - to be replaced when better/cheaper stuff came up), we eventually sticked with SGI. And this is what we bought:
Onyx2 InfiniteReality (IR) with, 4 x R10000 @ 195 MHz, 2 IR Pipelines, one with 2 Raster Managers (RM), and the second with 1 RM.
Each IR have 2 display output per default, 8 is optional. Each RM have 64 MB texture memory (256 in the newest IR3) and 80 MB frame buffer.
Theoretical performance of each RM is around 180 Mpixel/s with 8x MSAA and RGBA with 12 bits per component. However, when running at 1280x1024 from each RM, IIRC, you could only get 4x MSAA. You can have up to 4 RM per IR, and all 4 can be used for a single display output.
Theoritical geometry performance of each IR is 11 million triangles/s (when using meshes, basically vertices/s).
We also developed a new terrain database at that time. Initial tests showed us that peak performance of the IR, with around 55x40 degrees FOV, was 2 million triangles/s for our type of application/database. But because we were going to run 2 channels from a single IR, we ended up having less then 1 million triangles/s in the most intense parts of the application. Running 3 channels at 1280x1024 each, we were not able to reach 60 FPS, but had to settle with 30 FPS. I do believe that with some optimizations we could have reached 60 FPS, but as our main objective was on other parts of the simulation but the visual, we didn't try that hard.
The Onyx2 costed us around $175000........ just looking at the specs of the Parhelia-512, that single chip (especially when paired with 256 MB memory) should be able to do everything we used our Onyx2 for. The IR are more flexible, and can do some stuff that Parhelia-512 can do, but the opposite is also true: with the displacement mapping being the most impressive feature.
Creating large-scale terrain databases for flight-simulators is a time-consuming task. It involves combining vector data with elevation data to make up a triangulated description of the terrain, and then map some orthogonal corrected satelitte and flight photos: this is fairly straight forward. One problem though, is to create a good set of different LOD levels, and get smooth borders between different parts of the DB when the parts are using different LOD levels. But now, with displacement mapping....... which is using depth-adaptive tessellation!!
This is sooooo cool. If Matrox' depth-adaptive tessellation displacement mapping lives up to its promises, this will absolutely revolutionize the way large-scale, high-end, terrain DBs are created. Create a height-map of an area, add a mosaic of flight-photos, throw in some trees and buildings (at the right height), send it to our buddy Parhelia-512......... and voilà . Together with the possibility to drive 3 displays, this will be a super-hit within the high-end simulation comminity. Just give me some cheap CRT projectors now (CRT projectors are to prefer over LCD as the picture of the CRT ones can be distorted to fit the spherical screen. With LCD you have to distort the image in the image generator via multi-pass rendering.).
Now give us a way of creating ultra-realistic vegetation in real-time........ which is on its way via fractal-based generation. Check this.
Forgot one thing: the virtual texturing from 3DLabs P10 would be really nice to have in the next-gen Parhelia. That would make that Parhelia even more of a dream. The SGI OnyxIR-series supports texture-paging, but that involves transferring the whole texture - this can be done through application support on the Parhelia-512 as well. The 3DLabs method is so much nicer!!
Onyx2 InfiniteReality (IR) with, 4 x R10000 @ 195 MHz, 2 IR Pipelines, one with 2 Raster Managers (RM), and the second with 1 RM.
Each IR have 2 display output per default, 8 is optional. Each RM have 64 MB texture memory (256 in the newest IR3) and 80 MB frame buffer.
Theoretical performance of each RM is around 180 Mpixel/s with 8x MSAA and RGBA with 12 bits per component. However, when running at 1280x1024 from each RM, IIRC, you could only get 4x MSAA. You can have up to 4 RM per IR, and all 4 can be used for a single display output.
Theoritical geometry performance of each IR is 11 million triangles/s (when using meshes, basically vertices/s).
We also developed a new terrain database at that time. Initial tests showed us that peak performance of the IR, with around 55x40 degrees FOV, was 2 million triangles/s for our type of application/database. But because we were going to run 2 channels from a single IR, we ended up having less then 1 million triangles/s in the most intense parts of the application. Running 3 channels at 1280x1024 each, we were not able to reach 60 FPS, but had to settle with 30 FPS. I do believe that with some optimizations we could have reached 60 FPS, but as our main objective was on other parts of the simulation but the visual, we didn't try that hard.
The Onyx2 costed us around $175000........ just looking at the specs of the Parhelia-512, that single chip (especially when paired with 256 MB memory) should be able to do everything we used our Onyx2 for. The IR are more flexible, and can do some stuff that Parhelia-512 can do, but the opposite is also true: with the displacement mapping being the most impressive feature.
Creating large-scale terrain databases for flight-simulators is a time-consuming task. It involves combining vector data with elevation data to make up a triangulated description of the terrain, and then map some orthogonal corrected satelitte and flight photos: this is fairly straight forward. One problem though, is to create a good set of different LOD levels, and get smooth borders between different parts of the DB when the parts are using different LOD levels. But now, with displacement mapping....... which is using depth-adaptive tessellation!!
This is sooooo cool. If Matrox' depth-adaptive tessellation displacement mapping lives up to its promises, this will absolutely revolutionize the way large-scale, high-end, terrain DBs are created. Create a height-map of an area, add a mosaic of flight-photos, throw in some trees and buildings (at the right height), send it to our buddy Parhelia-512......... and voilà . Together with the possibility to drive 3 displays, this will be a super-hit within the high-end simulation comminity. Just give me some cheap CRT projectors now (CRT projectors are to prefer over LCD as the picture of the CRT ones can be distorted to fit the spherical screen. With LCD you have to distort the image in the image generator via multi-pass rendering.).
Now give us a way of creating ultra-realistic vegetation in real-time........ which is on its way via fractal-based generation. Check this.
Forgot one thing: the virtual texturing from 3DLabs P10 would be really nice to have in the next-gen Parhelia. That would make that Parhelia even more of a dream. The SGI OnyxIR-series supports texture-paging, but that involves transferring the whole texture - this can be done through application support on the Parhelia-512 as well. The 3DLabs method is so much nicer!!
