ATTILA GPU Simulator

R

RoOoBo

Regular
ATTILA, the infamous GPU simulator, has (finally!) an official site. I'm not the one designing it so you won't be able to enjoy the VAX themed color scheme of my blog anymore ;)

http://attila.ac.upc.edu

Ignore (accept it) the security certificate warning. No one here seems to know how to disable that thing.
 
I think we'd be happy to toss a long-time valued contributor like yourself a front page news item without the certificate issues, RoOoBo. . . . dunno if that will move anyone there to deal with it. . .
 
Really nice project and work on the Attila simulator. I'm impressed.

What kind of things can it simulate in addition to performance and can you vary different major variables like bus widths, memory speeds, amount of rops, alus etc. to diagnost bottle necks in "real life"?

How about mobile chips? Do you have some stuff there that could for example optimize/approximate power consumption with different configurations. Would really be fun to have bit more tech info about Attila on the project pages to read to better understand what one could do with such a simulator and how the internals work.

Edit: Seems that the info is already there. Perhaps the better question is how well does the simulation match real world when changing major parameters ?
 
Davros said:
quick question - it doesnt actually send anything to the monitor ?
Click to expand...

Well, with a bit of effort it could but I think you would get bored after the first hour ... waiting for the first frame to finish. So, no, it just outputs a file with the rendered image after 'a while'.
 
Geo said:
I think we'd be happy to toss a long-time valued contributor like yourself a front page news item without the certificate issues, RoOoBo. . . . dunno if that will move anyone there to deal with it. . .
Click to expand...

They will look at it.
 
manux said:
Edit: Seems that the info is already there. Perhaps the better question is how well does the simulation match real world when changing major parameters ?
Click to expand...

Taking into account that the current implementation is quite limited compared with modern GPUs, for example the model used for the shader processors doesn't support the complex architecture that real GPUs implement, multilevel ALUs and such (so in extreme cases five or more ARB assembler instructions are executed in a single cycle), there is a significative delta between a 'real' GPU performance and the simulated performance. The open source version is also limited because it has become somewhat outdated. The current feature set is similar to the R3xx generation, even if the architecture model is unified. But as the current OpenGL games don't seem to support any modern features that's not a big problem yet. We will be working in the next months to support D3D9 and D3D10 games.

But the process of working with the simulator and adjusting the configuration while comparing with real GPUs has been quite valuable at obtaining information about how a decent architecture would look. For example initially we had a relatively large amount of bw wasted for texturing, which didn't seem to match some hints we had about the matter, and we discovered that it could be solved with two cache levels. The first level would store uncompressed data and the larger (but less complex) second level would store compressed data. With the source available anyone can experiment with it and discover how well it behaves and work to improve it.

One of our PhDs is looking at the architecture of the Memory Controller and to what kind of scheduling and tricks could be used to optimize the access to memory taking into account the different types of memory consumers in the GPU: compressed depth data from the Z ROPs, color data from the color ROPs, texture data, etc.
 
Geo said:
It actually seems to be the case that doing the index.php/Main_Page bit makes the difference here. Going to the top level domain URL even as HTTP still causes IE to have a hissy about the certificate. Any rate, we'll throw it in the hopper to News.
Click to expand...

That's because the main page redirects to the https site, and the cert is self signed. The project team would have to pay money to get the SSL certificate signed by a certificate authority. The certificate looks pretty well formed, but the browsers are always going to complain about it being self signed.

RoOoBo said:
Ignore (accept it) the security certificate warning. No one here seems to know how to disable that thing.
Click to expand...

If you need help understanding SSL, let me know. I manage a site with a Verisign SSL key at work, I can help you guys figure out how to get the cert. signed if that's what you'd like to do.
 
the maddman said:
If you need help understanding SSL, let me know. I manage a site with a Verisign SSL key at work, I can help you guys figure out how to get the cert. signed if that's what you'd like to do.
Click to expand...
I recently bought a single-root cert for 24USD at http://www.trustico.com/ (reselling Geotrust RapidSSL). Have the server admin provide you with the CSR (certificate request) and you will be OK in a matter of minutes :D
 
RoOoBo said:
....We will be working in the next months to support D3D9 and D3D10 games.

But the process of working with the simulator and adjusting the configuration while comparing with real GPUs has been quite valuable at obtaining information about how a decent architecture would look. For example initially we had a relatively large amount of bw wasted for texturing, which didn't seem to match some hints we had about the matter, and we discovered that it could be solved with two cache levels. The first level would store uncompressed data and the larger (but less complex) second level would store compressed data. With the source available anyone can experiment with it and discover how well it behaves and work to improve it.

One of our PhDs is looking at the architecture of the Memory Controller and to what kind of scheduling and tricks could be used to optimize the access to memory taking into account the different types of memory consumers in the GPU: compressed depth data from the Z ROPs, color data from the color ROPs, texture data, etc.
Click to expand...

Congratulations!

I think ATTILA's excellent for such experiments and development. I'm presently working on something similar and would like to use ATTILA, but as a software I think it needs some more work (compatibility issues, errors, detailed documentation/tutorial :) ) before it hits big. I'm sure it has the potential.
 
anjulpa said:
Congratulations!

I think ATTILA's excellent for such experiments and development. I'm presently working on something similar and would like to use ATTILA, but as a software I think it needs some more work (compatibility issues, errors, detailed documentation/tutorial :) ) before it hits big. I'm sure it has the potential.
Click to expand...

Yes, I know it sucks the lack of compatiblity, documentation and support but we are a small group and we have very limited resources right now so we can do so much. We have other priorities right now. Making the simulator open source was a 'promise' I made to myself at the early stages of development and thus I gave it as a 'gift' to the world last year. But the group as a whole can not properly support the simulator and release new versions as it is now. And I could only work sporadically in the simulator in the last year since my PhD grant expired and had to take another job.
 
You know, several people in the GPU research community (and myself) would be willing to help you in this work. Have you ever thought of trying that?
 
anjulpa said:
You know, several people in the GPU research community (and myself) would be willing to help you in this work. Have you ever thought of trying that?
Click to expand...

I'm not sure what kind of help you mean. In any case the source code is avalaible and with one of the more open licenses that exists. Anyone can use it to start a collaborative work. I would open a SourceForge mirror for the code if I thought there were people interested on updating it. And as far as I'm not limited by my current work I will answer any question about the code, as I have been doing in the YahooGroups mail list this year, and help manage the project.

However anything related with updated 'official' releases, releasing the source code for parts of the framework that have not been released or propper support for the current version is outside my personal scope. That must go through other channels.
 
You must log in or register to reply here.

Similar threads

R
ATTILA GPU Simulator : The NI Version
Replies
2
Views
5K
RoOoBo
R
Albuquerque
Rant mode: I hate utter morons who work in IT
2
Replies
21
Views
5K
Albuquerque
Albuquerque
Cyan
Wholly unique/special ways to play in the next gen! Surprises inside
2
Replies
30
Views
8K
pjbliverpool
P
L
  • Sticky
Game development presentations - a useful reference
53 54 55
Replies
1K
Views
345K
troyan
T
B
  • Locked
Insomniac Games Brian Hastings on why PS3 is the console to beat
5 6 7
Replies
121
Views
16K
StefanS
S
Back
Top