nVidia Conference Call

[PaulS]

The nVidia Conference Call for this quarter just finished, and these were the main points that I can remember:

- Complete family of NV4x to launch in April, with a bridge chip providing PCI-E compatibility
- Bridge chip apparently isn't causing bandwidth problems (i.e. it's performing better than AGP 8x)
- Native PCI-E solutions to appear in June
- Refresh chips to appear in September (or at least, that's how I read it)
- New nForce 3's to support both 939-pin A64 and PCI-E
- Mainstream and Budget offers are unquestionably better than those from competitors. High-end is merely "competitive".
- Happy with IBM and 5700 margins/volumes
- Continuing to split manufacturing between IBM and TSMC, with 0.11 products at TSMC mentionned.
- Increased cycle times and costs are associated with Low-K, and they're not convinced it's right for high-volume parts. They say they're still looking at the possibility of using it somewhere in their line-up at sometime this year.
- Effectively confirmed they were working on an Intel chipset (when pressed, they simply issued an awkward "no comment", saying they'd only be happy commenting on their AMD chipsets at the moment)
- XBox revenues are down, and the trend is going to keep going down over the next quarter
- Q1 2004 is expected to be flat or down 5%

If this is the wrong forum for this, feel free to move it.

 

[Thunderbird]

How can the bridge chip be performing better than AGP 8x, doesn't it have to end up sending out only the same amount of bandwidth that 8x supports?

 

[rendezvous]

From what i heard (i was only listening @ one ear) the GPU's are capable of sending and reciving more data than APG 8x, i think he gave some sort of short explanation of how it worked, but the sound was all crappy and i wasn't really listening. So the GPU's are not limited to AGP 8X and they will recive a performance gain when a PCI-Express bride is slapped on. At least the hi-end boards.

Let's just hope that someone will give a better explanation soon, becouse i'm not too happy with this one... ^_^

 

[incurable]

How can the bridge chip be performing better than AGP 8x, doesn't it have to end up sending out only the same amount of bandwidth that 8x supports?

Simply put: It can't.

cu

incurable

 

[Mariner]

How can the bridge chip be performing better than AGP 8x

Marketing?

 

[T2k]

- Mainstream and Budget offers are unquestionably better than those from competitors. High-end is merely "competitive".

Is this just me - I feel contradictionhere - or really weird stuff going on?

Main and budget cards are derivated from its high-end part - so how can you be 'unquestionably better' here if your high-end part doesn't perform too shiny (even though you are the King of PR )?

Edit: grammar

 

[Tim Murray]

Should we make RV420 predictions yet? Or is it too early?

I'm going to say 128-bit memory bus, GDDR2 or 3, 600-650Mhz raw, 4x2.

 

[GraphixViolence]

Even if they got the AGP bus running faster than 8X, you could still only ever get half the total bandwidth of PCI Express at best. PCI Express x16 has separate independent upload and download links both running at 2.5 GHz and 4 GB/sec. AGP has a single link that can upload or download, but not both at the same time. And that's assuming they could double the AGP clocks of graphics cores that were designed for AGP 8X speeds, which is highly unlikely.

I think the main benefit of PCI Express is going to be the GPU->CPU bandwidth anyway. All of the traffic in games goes the other way (CPU->GPU), and we've seen that AGP 4X already provides more than enough bandwidth in this direction (i.e. going to 8X provides no performance improvement). If you ever want to send GPU output back to the VPU, AGP is limited to much lower speeds on current hardware. I remember seeing an article on www.techreport.com I think, way back when. Something to do with a video editing benchmark.

In general, I'd say it sounds like someone is back on the hallucinogens.

 

[Richthofen]

Is this just me - I feel contradictionhere - or really weird stuff going on?

Main and budget cards are derivated from its high-end part - so how can you be 'unquestionably better' here if your high-end part performs isn't too shiny (even though you are the King of PR )?

because in the low end there is no competition. The competition still does not offer a low end DX9 sollution.
Mainstream seems to be fine - mainly because the competition crippled its own mainstream products a littlebit by smaller memory bandwith.

Looking @ PCI Express i have to say that i definatly like the move with the bridge chip. They will be able to supply the market with PCI express solutions from top to bottom in a very short timeframe. Of course the bridge chip version is slower than a native PCI express version.

But does that matter? Don't think so. Low end and mainstream won't profit from PCI express anyway. High End maybe will a little but overall current game titles are designed to work on AGP2x, 4x and 8x. I think we won't see any significant performance gains with a native PCI express sollution. The bridge chip version enables them to supply PCI express compatible hardware without redesigning the GPU. in case of 3 to 4 different GPUs in Nvidias NV3x roadmap right now - that safes a lot of work.

 

[Mark]

after-hours trading not looking so good for NVDA atm. Down almost 8%.
http://finance.yahoo.com/q?s=nvda

 

[AlphaWolf]

because in the low end there is no competition. The competition still does not offer a low end DX9 sollution.

9600SE cards are going for as low as $70. That's not low enough?

 

[demalion]

The low end comment seems like pure marketing BS. DX 9 featureset checkbox does not "unquestionably better" make, especially with how DX 8.1 (on an 8500-9200) trades advantages and disadvantages in general featureset and usable shader length/performance with the "DX 9" implementation of FP mixed with FX12 that its competition utilizes.

The mainstream comment, OTOH, has to be about the most accurate PR comment coming out of nVidia in the last year or two (even if there are still some significant disadvantages for the 5900SE), but not because the "competition crippled" its products (at least in comparison to what others have put out for the market segment). nVidia simply ended up putting out an outstanding value for consumers by pricing a 256-bit card with "a high end transistor budget" for its core in the mainstream bracket, where its disadvantages tend to be actually outweighed by its real competitive strengths. Give credit where credit is due, and for as long as the situation lasts.

You can, however, find stock of 9700 and 9700 Pros for similar prices (or better) sometimes, but that can quite reasonably be said not be the same thing from the IHV perspective of offering a product.

 

[Xmas]

How can the bridge chip be performing better than AGP 8x, doesn't it have to end up sending out only the same amount of bandwidth that 8x supports?

Simply put: It can't.

cu

incurable

Not necessarily. It may be unlikely, but we don't know what the AGP interface of the chips is really capable of. AGP clock is very low in comparison to the core clock, so it could very well support significantly faster rates. When NVidia designed their AGPx8 interface, they were surely aware of PCI-Express. And the bridge chip has certainly been on the roadmap for more than the last few months.

 

[thatdude90210]

In last quarter's conference call, the CEO predicted "record revenue" for this quarter. That turned out to be wrong, the revenue was flat. This time he predicts a flat to down next quarter, wonder how the market will react to that.

 

[Dave Baumann]

Not necessarily. It may be unlikely, but we don't know what the AGP interface of the chips is really capable of. AGP clock is very low in comparison to the core clock, so it could very well support significantly faster rates. When NVidia designed their AGPx8 interface, they were surely aware of PCI-Express. And the bridge chip has certainly been on the roadmap for more than the last few months.

Sustained throughput of AGP is considerably less than its theoretical rates (about 1.2GB AFAIK) whereas 16X PCI-Express is a little little higher than that - if this is an AGP limitation on the chipset side there may be some small benefits over standard AGP, but not close to PCI-Express.

 

[Brandon]

While it may be fun to debate the merits of AGP 8X, PCI Express, etc, personally I think it's all much about nothing. As long as ATI and NVIDIA continue to pair the graphics processors with more memory we'll never see the total benefits of these newer interfaces anyway.

 

[reever]

Wow, this CC was exceptionally bad, sorta reminds me of the time the Nvidia guys had trouble forming words and coherent sentences as their stock dropped 50 percent a couple quarters ago.

Some of the comments I like alot, like the low-k comments, and how they are so proud of IBM, yet still IBM has no low-k and as they said, no 0.11, both of which TSMC has

 

[GraphixViolence]

While it may be fun to debate the merits of AGP 8X, PCI Express, etc, personally I think it's all much about nothing. As long as ATI and NVIDIA continue to pair the graphics processors with more memory we'll never see the total benefits of these newer interfaces anyway.

I wouldn't be so quick to jump to this conclusion... PCI Express is certainly more than just "AGP 16x". The fast GPU->CPU transfers could open up some interesting new possibilities for game developers, like offloading collision detection, physics calculations, etc. to the GPU to do load balancing. It probably will have minimal impact on existing games, but this could present more performance optimization opportunities for future games regardless of how much memory is on the graphics card. The more programmable and powerful GPUs become, the more beneficial this could become.

Another big advantage could come when more IHVs start moving to virtualized resource models, where you don't have to keep all of your textures, models, shaders, etc. in graphics memory to achieve good performance. I think 3DLabs already offers something like this, and it looks like the way DX10 will be moving, so it could start driving up the importance of bus bandwidth.

 

[T2k]

Wow, this CC was exceptionally bad, sorta reminds me of the time the Nvidia guys had trouble forming words and coherent sentences as their stock dropped 50 percent a couple quarters ago.

Some of the comments I like alot, like the low-k comments, and how they are so proud of IBM, yet still IBM has no low-k and as they said, no 0.11, both of which TSMC has

Exactly: I can definitely smell some fire between NV and TSMC...

 

[Ailuros]

Not necessarily. It may be unlikely, but we don't know what the AGP interface of the chips is really capable of. AGP clock is very low in comparison to the core clock, so it could very well support significantly faster rates. When NVidia designed their AGPx8 interface, they were surely aware of PCI-Express. And the bridge chip has certainly been on the roadmap for more than the last few months.

Sustained throughput of AGP is considerably less than its theoretical rates (about 1.2GB AFAIK) whereas 16X PCI-Express is a little little higher than that - if this is an AGP limitation on the chipset side there may be some small benefits over standard AGP, but not close to PCI-Express.

You may excuse if there's any brainfart involved, but isn't buswidth highly relevant to CPU and host ram bandwidths? I can understand the substantionally higher transfer bandwidth of PCI-E, what I haven't been able to understand so far is what good it would do a system if the difference between CPU and host ram bandwidth is hypothetically still in the 600MB/s or 1GB/s league.

What exactly am I missing here?