ATi Nanosecond Presentation

mr_angry

Has anyone else heard about the leaked ATi presentation from the GDC? Apparently it was posted on their site yesterday, but whoever posted it forgot to remove some of the slide notes. These notes give quite a lot of information on where the R420 will beat the NV40, and vice-versa.

I'm going to tentatively presume that its acceptable to post this here, considering ATi had it freely available on their site for some time, and its already available all over the Net.

If there is a need to remove it, people can feel free to e-mail me about any completely un-related matter, if you get my drift.

http://pwp.netcabo.pt/0239863201/Save_The_Nanosecond.ppt

BloodBath

Damn I don't have powerpoint. Any chance you can write the details?


BloodBath

mr_angry

For all Power-Point deficient, here is the complete list of presenter notes:

Slide 1
•This talk should be fun and engaging.
•This is developers‘ chance to know what the landscape will be like for the next 6 years.
•But we‘re not announcing dates or products – just where we want tobe and why.

Slide 21
Key Messages:
•ATI is on the forefront of driving the visual experience in many different business areas
•We have moved beyond just a PC graphics company with a 3D focus
•This hits many market segments: consumer, commercial, SOHO & enterprise – to name just a few
•Also mention – higher integrated chipset volume than NVIDIA, P4 license, imageon, xilleon etc
•Setting targets for what features to aim for with games for late 2004.
•Note that many developers treat all DX9 NV hardware as “low end DX9” hardware because that’s the best way.
•But be careful not to trash NVIDIA too hard here – your audience will help you out later...
•Also game genres obviously affect this and can drawn in low end DX7 hardware too

Slide 22
•IMpress upon the audience that targetting high end gamers is EQUIVALENT to targetting ATI hardware.
•These figures are from Mercury Research Q4 2003 (most recent figures at time of writting)
•Be careful not to mention Xbox 2 – just let everyone know that‘s what you mean...

Slide 23
•This slide is to get PC developers to understand that these related markets are relatively close behind their main market. And they‘re going to be big!

Slide 24
•This is to let PC developers know that they usually aim too low. PC’s only last on average about 2 years as games machines before they get heavily upgraded – top end graphics cards in retail last on average only about 18 months before being replaced.
•All of which means that you can and should target higher than your publisher thinks. Aim high and succeed, aim low and spend the rest of your life in scalability and back-compatibility hell.

Slide 25
•Setting targets for what features to aim for with games for late 2004.
•Note that many developers treat all DX9 NV hardware as “low end DX9” hardware because that’s the best way.
•But be careful not to trash NVIDIA too hard here – your audience will help you out later...
•Also game genres obviously affect this and can drawn in low end DX7 hardware too

Slide 26
•True of everyone’s hardware

Slide 27
•True only of ATI’s hardware.
•NVIDIA’s hardware has 2-3 times the shader perf available in FF mode. But then that kills alls the useful features.
•Hence treat all NV DX9 hardware as DX8... (See previous slide)

Slide 28
•You get lots of questions on this, so know the answers well!
•Sort by Render-Target, Draw Front to Back, Sub-Sort by shader.
•State change groups include matrices etc. Group them for best perf.
•Quote MS who say that current Xbox games average 150K polys per frame at refresh rate. It’s doable – you just need to (a) try and (b) talk to people like us who’ll help.

Slide 29
•DO NOT be specific about which we will adress first.
•THis is all about faster pixel shader throughput, not a statement about how many pixel engines we‘ll have

Slide 31
•It’s generally agreed that SM3.0 requires a move to 32 bit floats (some still think it’s not)
•But - The spec requires it...
•Blend etc will be on NV40 and on our own SM3.0 parts
•Steer people away from flow control in ps3.0 because we expect it to hurt badly. [Also it’s the main extra feature on NV40 vs R420 so let’s discourage people from using it until R5xx shows up with decent performance...]

Slide 32
•Mention the absence of support for overflow and underflow too – since it makes you look knowledgeable.
•64 bit may well be required in the first new DX after Longhorn.

Slide 33
•Describe how good they are – but they’re not very flexible, and you can be confident that some silicon will always be idle with this model.

Slide 34
•Explain that we have lots (literally hundreds of sub-processors) – and therefore any inefficiency is magnified

Slide 35
•A pool of vector units, each at least 4 units wide (x, y, z, w or rgba).
•It’s not really the driver which allocates resources – it’s innate to the hardware.
•The hardware may reconfigure its-self part way through a large triangle to make use of the spare compute power.

Slide 36
•People smile when they see this – because they think we‘re converging with the CPU guys
•“ Cache Interaction“ is about the multiple caches on the VPU (pixel, Z, texture, vertex) which know nothing of one another. CPU‘s have to produce cache consistency (and for examp0le this allows self altering code). We don‘t do it – and have no short term plans to do it.
•But you can simply flush any cache if you want to be able to get at it from another unit.
•Braching – CPUs are great at predictable braching. We don‘t like any branching.
•Also mention that our VPUs ae 5D. The vertex units can paralellise a 4D op with a 1D op – giving 5 paralell results! [NV can‘t do this!]

Slide 37
•Be careful to stress that “DX Next” is not Microsoft nomenclature. We’re just talking about what we need to do, we’re not announcing Longhorn’s graphics.
•The audience will allow you to be vague here – so take that chance. Much of this is not finally decided at this time..

Slide 38
•Mention that you expect to see a 512MB board in the consumer space this year. Don’t say that it’ll be R420 based!
•Don’t go into detail about the granularity of the VM system – but just press home the fact that this solves important problems.
•Don’t give any detail on the write capabilities at all.

Slide 39
•Explain that the date is speculative – this is not MS’s formal timetable.
•On the issue of small batches ask for samples which show bad perf because of small batches. We plan to solve this – but we need samples with plenty of real world data so that we can be confident.
•This is their chance to offer help to us – so ask for it. Point out that we’re here to listen as well as to speak.
•What gets built into Longhorn will be the basis for their next games. That’s games - plural. DX Next will last at least as long as DX9 (almost 4 years).
•If the geometry shader is “done right” it’s part of the vector processor pool – and therefore can have a vertex shader before it – and after it! Which is useful...

Slide 40
•Point out that you have an animated logo! How cool is that?
•This talk should be fun and engaging. It‘s there to muster enthusiasm and persuade developers that we know where we‘re going, we‘re going to a good place, and we need their help and enthusiasm to make this all work.


You could always download the OpenOffice version of Powerpoint for free...