GDC Europe

I look forward to seeing the presentations, especially yours Deano. It will help give us an idea on a few things that this forum has been arguing about for a while now.
 
Oooh I didn't know Deano's giving a lecture. Lucky me then for being at GDC :) Today's mobile thing was quite a bore though, but the real biz begins today anyways.

Well, time to hit the pub. Cheers!
 
DeanoC, don't know if you can correct this but there is a typo in slide 10:
An in-order processor can't work around cache stalls that well, Cell does have some help in this regard but don't expect it to work as well an OOOE processor in this regard
Shouldn't that be "don't expect it to work as well as an OOOE processor"
Nitpicking I know. :oops:
 
Squeak said:
DeanoC, don't know if you can correct this but there is a typo in slide 10:

Shouldn't that be "don't expect it to work as well as an OOOE processor"
Nitpicking I know. :oops:

Thanks, I fixed my version I will use on Thursday and I'll see about getting the online one updated once I've done it. I need to improve the references section, that was pre BPA docs release, I was going to add them once they were release but haven't done so yet...
 
Last edited by a moderator:
This post right here is slightly off topic, but it's in the context of your presentation tomorrow Deano.

I guess I'm wondering firstly: with Cell recently being 'opened up,' and with the majority of hands-on experience for some time to come probably lying in the hands of the game dev community, would it possibly be the case that a lot of the non-gaming apps we do eventually see come out for Cell being created at the hands of game devs, simply because they will have the most experience and the urge might strike?

Secondly Deano there's been some controversy in the recent past as to how many GFLops the PPE in Cell can push; since your slides back up the E3 218 figure, does that then verify the PPE at 38.8 GFLops?
 
How can I enter in the industry?

My Skills are:

-I Know C, ObjC, Python, Delphi and RealBasic.

-I have some knowledge of OpenGL from playing with Quake2 Engine and Quake3 Engine but I never created an application from 0 using OpenGL.

-I am studying the career of multimedia (image and sound).
 
Urian said:
How can I enter in the industry?

My Skills are:

-I Know C, ObjC, Python, Delphi and RealBasic.

-I have some knowledge of OpenGL from playing with Quake2 Engine and Quake3 Engine but I never created an application from 0 using OpenGL.

-I am studying the career of multimedia (image and sound).

OFF TOPIC@@!!!!!11111:devilish::devilish::devilish:
 
xbdestroya said:
Secondly Deano there's been some controversy in the recent past as to how many GFLops the PPE in Cell can push; since your slides back up the E3 218 figure, does that then verify the PPE at 38.8 GFLops?

This is how I work it out...

PPE can dual issue VMX FMADDs
SPE can single issue a SIMD FMADD
1 * 3.2 * 2 +
7 * 3.2 * 1 = 230.4 billion FLOPs

Now I'm not sure if a GFLOP is 1 billion or 2^30 FLOPs so I'll go with a billion...

Thats the maximum flops you can get (I think), its irrelevant if the PPE FPU's could do some work, as you have filled all the instruction slots with VMX instruction so you can't issue any...

I think I'll update the slide as well.
 
Something that confused me at first in your PPT was the slide entitled 'PPU Threads'. It reads...
For the current implementation of Cell it's best to think about it as two separate PPUs.
Which reads as 'Think of Cell as two half-speed cores' and not 'Think of the PPU part of Cell as two half-speed cores'. Though that'll be obvious from the talk the slide might want an edit for clarity.
 
Shifty Geezer said:
Something that confused me at first in your PPT was the slide entitled 'PPU Threads'. It reads...
Which reads as 'Think of Cell as two half-speed cores' and not 'Think of the PPU part of Cell as two half-speed cores'. Though that'll be obvious from the talk the slide might want an edit for clarity.

Thanks, updated it.

That whole section underwent a lot of changes... originally I had lot more details on why its best to think about it like that but then that information hasn't been publically disclosed so I removed some of the low-level details...
 
Nice presentation DeanoC.

I would recommend adding graphics/drawings that show what you will be talking about, and move some of the detailed explanations to the notes section.

It will help to allow people to listen to you, rather than try reading what you wrote.

Great info in there though, not sure if you'll get an opportunity to add diagrams.

Good luck,
Speng.
 
DeanoC said:
This is how I work it out...

PPE can dual issue VMX FMADDs
SPE can single issue a SIMD FMADD
1 * 3.2 * 2 +
7 * 3.2 * 1 = 230.4 billion FLOPs

Now I'm not sure if a GFLOP is 1 billion or 2^30 FLOPs so I'll go with a billion...

Thats the maximum flops you can get (I think), its irrelevant if the PPE FPU's could do some work, as you have filled all the instruction slots with VMX instruction so you can't issue any...

I think I'll update the slide as well.

shouldn't that be:
2 * 3.2 * 4 * 2 = 51.2
7 * 3.2 * 4 * 2 = 179.2
+ = 230.4 GFLOPS

BUT...

The PPE is only 2 issue according to available data, so even in a daxpy loop, the max sustained flops for the PPE is actually 25.6 since on insruction slot will be needed to load data.

So sustained is actually: 204.8 GFLOPs

and flops are 10^9. Only memory is 2^x.


Aaron Spink
speaking for myself inc.
 
Back
Top