Nvidia losing influence due to Ps3 involvement?

[hstewarth]

I actually surprise NVidia is doing this deal with PS3. Didn't they get a bad deal with Microsoft on the XBox with making a lot of chips for cheap...

Microsoft had the bad deal, they are paying a fixed price on a chip that got cheaper to make as time went on... Nvidia, on the other hand, is still getting $80-100 million per quarter revenue on a now outdated chip.

It really depends on your point of view - you can also think it means that NVidia must keep up and maintain a production of outdated chip because Microsoft uses an outdate chip.. The chip in XBox is bases on GeForce 3/4 series which NVidia has advance 2 technologies ahead of it. Not to even mention the production process has change which means they have to keep the manufaction lines up.

 

[Richthofen]

It really depends on your point of view - you can also think it means that NVidia must keep up and maintain a production of outdated chip because Microsoft uses an outdate chip.. The chip in XBox is bases on GeForce 3/4 series which NVidia has advance 2 technologies ahead of it. Not to even mention the production process has change which means they have to keep the manufaction lines up.

So where is the problem with that? They have nothing to do with manufacturing. Once production ramp up is completed and yields are good, all they do is ordering a certain amount of wafers @ their foundry partner.
They sold TNT2M64, GF2MX and GF4MX for such a long period even after several next generation parts have been out.
That is not a problem at all. There isn't any change of the production process. Those GPUs are still 0.15 micron based and this process is getting cheaper and cheaper because demand for it is getting lower unlike 0.13, 0.11 and 0.09.

 

[Bouncing Zabaglione Bros.]

This was the main row between Nvidia and MS a couple of years back. Nvidia took the contract on a fixed fee for development and certain numbers of chips. If things had gone wrong and Nvidia overspent, then MS would have not had to pay any more money.

In fact what happened is that Nvidia did very nicely, and even managed to use a lot of the tech in their Nforce chipsets. Microsoft didn't shift as many X-boxes as they expected, and were forced into a price war with Sony (who's PS2 was a lot cheaper to make and selling a lot more units).

MS wanted Nvidia to drop the price of their chips to help them, especially once the initial runs had been completed and Nvidia was getting back profit. Nvidia wanted to stick to the contract, thank you very much.

 

[KimB]

Furthermore, there were rumors that partial precision itself was a late addition to the DX9 spec, which would seem to indicate that nVidia didn't originally intend to support partial precision at all with the NV30. But I don't think we can know this for certain.

Did the rumor say if they were looking at supporting just FP16 or FP32?

Erm, partial precision is the only way FP16 could be supported.

Additionally, on all NV3x parts, all FP processing units are also FP32. I really think that the register pressure that the NV3x parts suffer from was unexpected. There was an interview after the announcement of the NV40 parts wherein it was stated that the compiler for the NV3x series was the last thing they did. I think that they expected the compiler to be able to switch registers around efficiently enough that there would be no problem. But, as it turned out, this utterly failed.

So, with the NV40, they fixed this compiler issue by putting a portion of the compiler in the hardware. But they decided it made sense to also use partial precision as an option, and decided to this time add in some extra functional units.

Anyway, for once, I think BZB is right. The main stink between nVidia and Microsoft was over the price of the X-Box chips. But I don't think that should necessarily put nVidia "out of the loop" when it comes to designing future architectures.

 

[KimB]

I thought the rumours were the exact opposite. Nvidia built 32 bit for quality (remember Cinematic Computing?) and then tried to claw back speed with 16 bit. MS went the ATI route of a 24 bit midpoint compromise between speed, quality and transistor budget. Nvidia then had to lobby madly to get PP included in one of the DX9 revisions.

Utterly impossible. FP32 is not inherently any slower than FP24 (at least, not in a deeply-pipelined architecture where latency of the processing isn't important). Just consider that without adding any additional per-pipeline processing power, the transistor count more than doubled from the NV25 to the NV30. Since when should an improvement in featureset cause that kind of change? The transistor count changes from the NV30 to the NV35, and later to the NV40 are a further testament as to how FP32 was not the problem.

 

[hstewarth]

I would expect with Manufactoring of GPU's, it would be like CPU - like Intel will desired all manufactoring to go to one process so that its cheaper in the long one - also it means that you don't have maintain and staff the older facilitys.. This is main reason why Intel came out with 478 Prescotts.

As far as contracts go.. why should NVidia make it cheaper - just because Microsoft demanded it.

I also expect that ATI will benifit with XBox 2, they will use the technology that comes out of XBox in GPU / Motherboard chipset to build future GPU snad Motherboards.. I think it would be a veryy expensive contract to say that the technology being developer can only be the XBox series.

It is likely that NVidia / XBox deal was not good for both Microsoft and NVidia. Also Microsoft is using the GPU war between the two companies to their advantage.

Well thats my opinion.. I have absolutely no relationship with any of the companies except that I use there productes including ATI which is on one of my notebooks.

 

[KimB]

Microsoft would probably just buy ATI or Nvidia, which they can easily afford. I doubt MS would though, as it would probably leave them open to monopoly investigations.

Or that they could. If a company doesn't want to be purchased, it's nigh impossible to do so. The only way would be to buy a controlling interest in the stock, but it should be pretty easy for either nVidia or ATI to disallow this.

The only time hostile takeovers really become possible is when a company is really going downhill, and sells lots of stock in a paniced attempt to recover.

Similarly, the only time when a company would really want to sell would be when said company is not doing very well.

 

[hstewarth]

What MS could do is more of what they are already accused of doing: listening to a single IHV on individual rendering feature specifications, and ignoring the competitors. In the past it has reportedly been pixel shader precisions, texture formats, and such. It's much more specific. They might even go so far as to specify how exactly an implementation of a feature such as texture filtering must work (that is, a required equation to produce a specific output). General architecture should always remain up to the IHV, though.

So you agree that this might further (or create) a bias towards one IHV at the expense of another? What originally got me to think this was this quote by Dave...

Last I had heard, NVIDIA were still pressuring MS not to have anything in the API that forces a unified approach at the hardware level (i.e. whilst the VS and PS have a unified instruction set, the hardware could still be distinct if they wanted to).

I would expect it would be bad for industry if one approach to hard designed is desided - it may make it simpler for Microsoft to designed its support, but it lessen competition in the GPU industry.

 

[Bouncing Zabaglione Bros.]

Utterly impossible. FP32 is not inherently any slower than FP24 (at least, not in a deeply-pipelined architecture where latency of the processing isn't important). Just consider that without adding any additional per-pipeline processing power, the transistor count more than doubled from the NV25 to the NV30. Since when should an improvement in featureset cause that kind of change? The transistor count changes from the NV30 to the NV35, and later to the NV40 are a further testament as to how FP32 was not the problem.

And yet Nvidia still tells everyone to use PP wherever possible to gain speed over their 32bit full precision, and has been doing so for the last few years. Why do you suppose Nvidia do that then? Why does NV3x gain performance when forcing PP, even though it's at the expense of IQ?

 

[DemoCoder]

Because the issue isn't the speed of individual FP operations like FADD or FMUL, it's pipeline hiccups that are the problem because FP32 registers take up more context space. NVidia did not provide enough space on the NV30 core to hold more than 2 FP temporary registers per pixel being shaded without spillage (stall)

For the NV40 they doubled the amount of register space, and reduced the penalty for exceeding it. The result is that FP32 runs at full speed.

The NV3x gains performance because FP16 allows the chip to pack in 2 values per register instead of 1, decreasing the pipeline hiccups. It is not because FP16 operations themselves run faster than FP32. All FP16->FP32 does is increase latency, not throughput.

 

[radar1200gs]

This was the main row between Nvidia and MS a couple of years back. Nvidia took the contract on a fixed fee for development and certain numbers of chips. If things had gone wrong and Nvidia overspent, then MS would have not had to pay any more money.

In fact what happened is that Nvidia did very nicely, and even managed to use a lot of the tech in their Nforce chipsets. Microsoft didn't shift as many X-boxes as they expected, and were forced into a price war with Sony (who's PS2 was a lot cheaper to make and selling a lot more units).

MS wanted Nvidia to drop the price of their chips to help them, especially once the initial runs had been completed and Nvidia was getting back profit. Nvidia wanted to stick to the contract, thank you very much.

nVidia and Microsoft were in a legally binding contract. MS wanted to break that contraxt, nVidia didn't. Unsuprisingly, the courts favored nVidia. This didn't go down well with MS.

 

[radar1200gs]

Furthermore, there were rumors that partial precision itself was a late addition to the DX9 spec, which would seem to indicate that nVidia didn't originally intend to support partial precision at all with the NV30. But I don't think we can know this for certain.

Did the rumor say if they were looking at supporting just FP16 or FP32?

Erm, partial precision is the only way FP16 could be supported.

Additionally, on all NV3x parts, all FP processing units are also FP32. I really think that the register pressure that the NV3x parts suffer from was unexpected. There was an interview after the announcement of the NV40 parts wherein it was stated that the compiler for the NV3x series was the last thing they did. I think that they expected the compiler to be able to switch registers around efficiently enough that there would be no problem. But, as it turned out, this utterly failed.

So, with the NV40, they fixed this compiler issue by putting a portion of the compiler in the hardware. But they decided it made sense to also use partial precision as an option, and decided to this time add in some extra functional units.

Anyway, for once, I think BZB is right. The main stink between nVidia and Microsoft was over the price of the X-Box chips. But I don't think that should necessarily put nVidia "out of the loop" when it comes to designing future architectures.

The only bit I never got with the NV30 launch was why nVidia insisted on bringing nV30 to market after R300 and the failed tapeout around that time. They should have gone straight to nV35 right then and there and saved a lot of embarrasment.

 

[Bouncing Zabaglione Bros.]

it's pipeline hiccups that are the problem because FP32 registers take up more context space. NVidia did not provide enough space on the NV30 core to hold more than 2 FP temporary registers per pixel being shaded without spillage (stall)

I know FP32 is not intrinsicly slower in theory, but in *practice* FP32 is slower than FP16 on NV3x due to the design of the hardware. The hardware that Nvidia designed and heavily marketed to support FP32. Which is why Nvidia heavily lobbied MS for PP.

 

[DegustatoR]

I know FP32 is not intrinsicly slower in theory, but in *practice* FP32 is slower than FP16 on NV3x due to the design of the hardware. The hardware that Nvidia designed and heavily marketed to support FP32. Which is why Nvidia heavily lobbied MS for PP.

FP16 is a way of increasing performance without quality loss on NV3x and NV4x hardware. So tell us why it's so wrong to have it in DX and to lobby it in MS?

 

[anaqer]

FP16 is a way of increasing performance without quality loss on NV3x and NV4x hardware.

Is that so?

 

[Bouncing Zabaglione Bros.]

FP16 is a way of increasing performance without quality loss on NV3x and NV4x hardware. So tell us why it's so wrong to have it in DX and to lobby it in MS?

I was contending with Chalnoth's statement that FP16 was not intended in the NV3x design and was just put in at the last minute after PP was added to DX9 as a late addition. I think Nvidia intended PP to be in the NV3x design long before DX9 was clearly defined, and it was PP that was levered into the API at the last minute because that's what Nvidia designed into their hardware.

In general, I don't think that moving back towards low precision is a good trend. After all, Nvidia is always telling us their FP32 is so important and gives better visuals than FP24, yet they promote FP16 to developers. PP gives more work for developers, and in cases where it is forced or used incorrectly we see artifacts in games like HL2 and Far Cry.

FP16 is just a stopgap measure, even more so than the FP24 that ATI offers at a faster speed. I expect Nvidia to drop it and use those transistors to help them get a faster, more straighforward "FP32 everything" chip as soon as they can. As everyone in the industry is moving towards FP32, with Nvidia offering it now on NV3x/NV4x, to promote going back to FP16 because Nvidia can't run FP32 quickly enough in this generation seems a disingenious quickfix to a performance problem they brought on themselves.

 

[DegustatoR]

FP32 is running quickly enough on NV4x hardware. It runs at exactly the same speed as ATI's FP24 if we use the same core clocks. And FP16 gives NV4x a nice performance lead over Radeon's in this case.

There is no 'FP16 transistors'. FP16 is just data which still goes through FP32 ALUs. NVIDIA isn't spending transistors on it. It's just a flexibility feature allowing more perfomance for everyone who wants to fine-tune their shaders. There is virtually no reason to drop it not now, not in the nearest future. Many effects are quite happy with FP16 so why use FP32 for everything? Maybe we should use FP128 for everything just b/c 128 is bigger than 32?

As everyone in the industry is moving towards FP32

ATI isn't 'everyone'. You're far too biased you know...

 

[KimB]

And yet Nvidia still tells everyone to use PP wherever possible to gain speed over their 32bit full precision, and has been doing so for the last few years. Why do you suppose Nvidia do that then? Why does NV3x gain performance when forcing PP, even though it's at the expense of IQ?

Register pressure. I was arguing that nVidia most likely believed until close to the first intended launch date that there would be no significant register pressure problems in the NV3x, that the compiler that compiles the assembly to the VLIW machine language would be able to move things around efficiently enough that it wouldn't be a problem.

This didn't turn out to be the case. nVidia admitted that a major mistake with the NV3x was waiting to develop the compiler until the end of development. I claim that this was the reason partial precision was introduced, nVidia didn't forsee the compilation problems that come with a VLIW architecture, and we saw the problem as register pressure.

 

[karlotta]

The only bit I never got with the NV30 launch was why nVidia insisted on bringing nV30 to market after R300 and the failed tapeout around that time. They should have gone straight to nV35 right then and there and saved a lot of embarrasment.

they had no choice, they had contracts to fill. And the n35 wasnt due till may but didnt hit any real amount untill july. Needed to be in the mix. Have websites like TH say the new king of graphix. and all the rest. Then they hide the flaws and lack of supply by cheating 3dmark. All to keep there name in the mix untill supply hits in august.

 

[KimB]

For the NV40 they doubled the amount of register space, and reduced the penalty for exceeding it. The result is that FP32 runs at full speed.

Where do you get that the NV40 has double the amount of register space? I remember there was speculation to the tune of this, but I don't remember any hard data or interviews stating it.