FP16 and market support

[sonix666]

Nope, multiprecision will always exist. First of all, it already exists with respect to input/output (look at all the texture formats, and framebuffer formats) So storage is already multiprecision. Developers want the ability to deal with multi-precision data. We don't render everything at HDR resolution just because "it is easy and simple" to always use max-precision on everything.

Whats wrong with doing everything in high precision if speed is not affected.

Speed IS affected. That is the whole point.

1) Memory bandwidth.
Using lower precision, where you can, can lower memory bandwidth usage, which will result in better performance. The amount of performance difference is dependant on some factors, like was it already bandwidth limited.

2) Lower latency.
The lower the precision, the less cycles it takes to complete an operation. CPUs (and possibly GPUs shaders) nowadays are pipelined, so each cycle a new operation can be started. However, when another operation requires the results of another, the whole latency has to be waited. The GPU/CPU might try to reorder instructions to hide that latency.

As you can see I disagree with zeckensack that only bandwidth is the reason for going to lower precision. Another important factor is latency.

 

[Ailuros]

Nothing beyond what I posted.

Since it's bleedingly obvious where I was targeting at, there was actually no point at all in your former reply. Nothing beyond nothing then.

 

[Ailuros]

I don't see Microsoft adding fp16 in PS3.0.

I think an integer unit and an fp unit make sense, however having both fp16 and fp32 seems like a foolish waste of transistors.

I don't get your argument about data pecision formats because cpus have max precision fp units and deal with smaller data precisions all the time.

I think fp16 only exists so Nvidia can stuff two pixels down a 128-bit pipe instead on one. Works great for DX8 stuff.

Well most of it has been addressed already in other posts and the reasons why multiple precision formats or more specifically FP16 is and will continue to exist and have it's importance.

If you think that upcoming next generation hardware will not have any FP16 support, I'd say you should rethink it a tad more.

 

[Doomtrooper]

Speed IS affected. That is the whole point.

On what architecture would that be...

Developers seem to split on what control over precision they want, Sweeny thinks future hardware has no place for _pp.

http://www.beyond3d.com/interviews/ps_precision/

 

[Demirug]

2) Lower latency.
The lower the precision, the less cycles it takes to complete an operation. CPUs (and possibly GPUs shaders) nowadays are pipelined, so each cycle a new operation can be started. However, when another operation requires the results of another, the whole latency has to be waited. The GPU/CPU might try to reorder instructions to hide that latency.

Forget this point.

A GPU is more like a massive multithreaded VLIW-CPU. That is the reason why nVidia need such a large registerfile to store the state of all this running threads. But it solve all dependency problems.

 

[Demirug]

Speed IS affected. That is the whole point.

On what architecture would that be...

Developers seem to split on what control over precision they want, Sweeny thinks future hardware has no place for _pp.

http://www.beyond3d.com/interviews/ps_precision/

Sweeny thinks that there is no big future for special graphics hardware at all.

 

[Doomtrooper]

True

 

[KimB]

On what architecture would that be...

Given the same number of transistors, FP16 can be quite a bit faster than FP32. So, if a significant number of instructions only require the precision of FP16, then it makes a lot of sense to have FP16 support in hardware.

 

[arjan de lumens]

The only benefit of supporting smaller data types is in I/O bandwidth. That's the way it goes on CPUs, and that's the way it goes in proper GPU designs (yes, R300 can handle 16bit FP external data, such as textures and render targets).

I am disagree. A real clever GPU design use a FPU that runs FP32 at full speed and FP16 at a higher speed. It is possible to build such a FPU that use the same adder/halfadder array for both formats. You need only a little bit more control logic.

True. AFAIK AMD's SSE/3DNow implementations work this way.

False. Look here for the layout of the AMD K7 floating-point units. The 80-bit x87 adder is clearly separated from the 3dnow FP32 adders, whereas the 3dnow multipliers appear to be rolled into the x87 multiplier. This to me indicates that re-using parts of a wide FP adder in multiple narrower FP adders is a hopelessly/pointlessly complicated task (which is also evident if you try to write such a monstrosity in, say, VHDL; FP adders are *much* more complicated than integer adders), whereas multipliers are much nicer in this regard.

Also, if you take an FP16 adder and an FP32 adder and optimize/pipeline them, the FP16 adder should have only about 10-15% lower latency than an FP32 adder - something similar applies to multipliers as well.

 

[nelg]

Would it not be fair to say that, perhaps, Microsoft does not want to keep the pp-hint? With the increasing length time it takes to develop top notch games today would it not be a wise consideration to alleviate the step of having to look for precision issues?

 

[radar1200gs]

Nope, multiprecision will always exist. First of all, it already exists with respect to input/output (look at all the texture formats, and framebuffer formats) So storage is already multiprecision. Developers want the ability to deal with multi-precision data. We don't render everything at HDR resolution just because "it is easy and simple" to always use max-precision on everything.

Whats wrong with doing everything in high precision if speed is not affected.

Nothing is wrong with doing everything in high precision if that is what you wish to do or is all your GPU is capable of doing.

The partial precision hint does not interfere with being able to render in high precision at all - if the hardware can't use a lower precision the hint gets ignored, otherwise the lower precision will be used.

Someone else in this thread complained that they as an ATi user should not have to suffer because of nVidia's architectural decisions - once again no-one is making you suffer, your GPU can't use the partial precision hint and therefore ignores it completely.

 

[RussSchultz]

Would it not be fair to say that, perhaps, Microsoft does not want to keep the pp-hint? With the increasing length time it takes to develop top notch games today would it not be a wise consideration to alleviate the step of having to look for precision issues?

If I want to avoid the time it takes to determine what _PP can be applied to and what can't I don't need microsoft telling me its ok to do that.

 

[YeuEmMaiMai]

Yes there is considering that

1. We have to put up with game delays because of their inferior design. For example

a. DoomIII
b. HL2

2. If game developers are forced to remove a benchmarking mode that makes their design look bad (I can't help it that nvidia's card performs inferior to ATi's) Tomb Raider AOD for example

3. developers code specifically to vendor's hardware (STRYKER is being recoded to take advantage of Nvidia's so called special features all the while not implimenting the same features on ATi's hardware even though it can be done)

so now we have a divided community over what? a hardware vendor that decided to take on the industry and is loosing. If Nvidia repeats the same crap with nV4X then what will their excuse be then?

Nvidia choose to impliment FP16 against what MS stated was the spec. MS really had little choice but to allow it since there is a large based of nvidia users installed....

Someone else in this thread complained that they as an ATi user should not have to suffer because of nVidia's architectural decisions - once again no-one is making you suffer, your GPU can't use the partial precision hint and therefore ignores it completely.

 

[radar1200gs]

Yes there is considering that

1. We have to put up with game delays because of their inferior design. For example

a. DoomIII
b. HL2

2. If game developers are forced to remove a benchmarking mode that makes their design look bad (I can't help it that nvidia's card performs inferior to ATi's) Tomb Raider AOD for example

3. developers code specifically to vendor's hardware (STRYKER is being recoded to take advantage of Nvidia's so called special features all the while not implimenting the same features on ATi's hardware even though it can be done)

so now we have a divided community over what? a hardware vendor that decided to take on the industry and is loosing. If Nvidia repeats the same crap with nV4X then what will their excuse be then?

Nvidia choose to impliment FP16 against what MS stated was the spec. MS really had little choice but to allow it since there is a large based of nvidia users installed....

Someone else in this thread complained that they as an ATi user should not have to suffer because of nVidia's architectural decisions - once again no-one is making you suffer, your GPU can't use the partial precision hint and therefore ignores it completely.

Developers will almost always code for what is the most popular hardware, not what is the technically superior hardware. There are more nVidia owners than ATi owners, even in the DX9 class. Specifications are meaninless if they don't apply to the hardware the app is to be run on. The specs for an airplane engine won't do a car owner much good, no matter how much a mechanic might like to use them.

FP16 IS part of Microsofts specifications - it is the minimun precision allowed under directx9 and is invoked with the partial precision hint.

 

[AlphaWolf]

There are more nVidia owners than ATi owners, even in the DX9 class.

Only if you count the 5200 cards, which pretty much will be running dx8 or lower paths on games, because they cannot perform many dx9 tasks even _pp at acceptable speeds.

I would say that atm ATI dominates the truly capable dx9 market.

 

[radar1200gs]

There are more nVidia owners than ATi owners, even in the DX9 class.

Only if you count the 5200 cards, which pretty much will be running dx8 or lower paths on games, because they cannot perform many dx9 tasks even _pp at acceptable speeds.

I would say that atm ATI dominates the truly capable dx9 market.

The expectations of the value market re: performance is very different to the high end market. So long as you avoid the 64 bit variants (which are only suited to 'net pc or office pc usage IMO) the 5200 is fine for its target market segment.

If you are inclined towards being a high end user and you saddle yourself with a value card, you deserve all the pain headed your way. Use your brain.

 

[AlphaWolf]

I am glad you agree with my assessment.

 

[BRiT]

There are more nVidia owners than ATi owners, even in the DX9 class.

You have failed to show us the numbers... For all we know, you're talking out your anus, spreading even more FUD.

 

[jvd]

the 5200 is fine for its target market segment

Not really . Considering it gets sub 5fps in half life 2. It is marketed as a dx 9 card. If it can not even play a dx 9 game then it is not fine for its target market. If its target market was the land fill then it would be fine for its target market segment.

 

[radar1200gs]

The only fud around here is coming from the ATi fanboys - "FP16 is not supported by DX9", "Dawn is a DX9 showcase"...

The 5200 runs half-life2 - it may run it slowly, thats beside the point - if you want performance, buy a more expensive card, not a value/entry level card.