Discrepancies in Anandtech's tech preview?

[Bigus Dickus]

Once again he has posted information that is in conflict with what was provided in the Beyond3D review of the 9700 Pro.

His tables for both Vertex and Pixel shaders simply does not match: number of instructions/loops, number or registers in both pixel and vertex shaders, swizzling, advanced functions/macros such as sin/cos, etc.

Who is right? If Beyond3D is correct, shouldn't someone inform Anand to stop using numbers nVidia provides him, and instead use the correct numbers?

Oh, and maybe someone should let him know that 16 x 4 != 48.

 

[KnightBreed]

I believe Reverend addressed some of this in his CinFX article for ve3d.com a while ago. A shame I can't find the article now that they've removed the Search function.

Edit:I found it here, but that Domain has been ditched, and using ve3d.com doesn't work either.

 

[Dave Baumann]

I'm just reading his article. Spot the error here:

The lossless compression algorithm can obtain up to a 4:1 compression ratio which explains where NVIDIA gets their 48GB/s of memory bandwidth from. Remember that the 500MHz DDR2 memory on the GeForce FX provides 16GB/s of bandwidth on the 128-bit memory bus, but multiply that by 4 and you'll get the 48GB/s of memory bandwidth NVIDIA is claiming

Math doesn't seem a string point.

However, on the subject of colour compression under AA Anand says:

It is this technology that truly sets the GeForce FX apart from the Radeon 9700 Pro.

How? R300 also has colour compression with AA.[/quote]

 

[Maverick]

Math doesn't seem a string point.

Just like spellong isn't yours?

 

[Bigus Dickus]

Dave,

Could you clarify the issue of number of loops, number of registers, swizzling, etc... in just a few words?

Like, "yes, or B3D 9700Pro review had the correct information, Anand is incorrect" or otherwise.

I keep seeing the numbers presented in the CineFX papers reprinted everywhere, and I'd like to know if they are in fact correct, of it B3D's numbers are correct, before I stick my foot in my mouth anywhere.

 

[Dave Baumann]

Well, here's the thing: B3D's numbers are from 'sireric', someone who is known to have worked on the R300 architecture. Anand's numbers are consistent with the numbers that NVIDIA is going around quoting in relation to R300.

Take your pick...

 

[T2k]

Well, here's the thing: B3D's numbers are from 'sireric', someone who is known to have worked on the R300 architecture. Anand's numbers are consistent with the numbers that NVIDIA is going around quoting in relation to R300.

Take your pick...

Ridiculous, frankly.

 

[KnightBreed]

Well, here's the thing: B3D's numbers are from 'sireric', someone who is known to have worked on the R300 architecture. Anand's numbers are consistent with the numbers that NVIDIA is going around quoting in relation to R300.

Take your pick...

Ridiculous, frankly.

The information from sireric -> here

 

[Hyp-X]

On the topic of texture filtering, NVIDIA finally offers a non-purists anisotropic (and trilinear) filtering algorithm. This driver selectable option will enable users to choose from an accurate filtering algorithm (similar to what the GeForce4 currently has) or a slightly less accurate but very high performance algorithm (similar to what the Radeon 9700 Pro uses). The end result is that the performance hit incurred when enabling anisotropic filtering or even trilinear filtering when in this "performance" mode is considerably less. NVIDIA claims that their anisotropic filtering algorithm is more precise than ATI's, so the GeForce FX's anisotropic filtering should look just as good if not better than the Radeon 9700 Pro's.

What he is talking about? (Apart from spreading FUD...)

The Radeon 9700's anisotropic filtering quality is (finally) correct, and actually exceeding GF4Ti's as it can do 16x whereas GF4Ti only does 4x or 8x (depending on which driver used).

 

[Bambers]

Nah the 9700 still has weak angles, it just added a few more filter shapes.

It only does about 2xaniso at 22.5 + n*90 and 67.5 + n*90 degrees.

 

[T2k]

Well, here's the thing: B3D's numbers are from 'sireric', someone who is known to have worked on the R300 architecture. Anand's numbers are consistent with the numbers that NVIDIA is going around quoting in relation to R300.

Take your pick...

Ridiculous, frankly.

The information from sireric -> here

I mean the article - but thanks for the link.

 

[demalion]

Nah the 9700 still has weak angles, it just added a few more filter shapes.

It only does about 2xaniso at 22.5 + n*90 and 67.5 + n*90 degrees.

? OK, I'm not trying to rehash this discussion, but about "it only does about 2x aniso at...22.5.."?

Here are the pictures from the thread discussing this issue at 22.5 degreees, and I'd offer you open the links in their own windows and flip between them and evaluate whether that description ("only does 2xaniso") fits.

I'm assuming the pictures represent what they actually say they do, and are still up when you view this (they are as I post).

No need to even have a discussion about it here, it has been done and you can read it in the rest of the thread, but here are the links for convenience:

Radeon 9700 16x

GF 4200 8x

My opinion: "Only does 2xaniso at n*90+/-22.5" is a wrong statement. Please, keep opinions concerning this based on these pictures brief to avoid thread hijacking, the thread for this discussion already exists.

 

[WaltC]

Well, here's the thing: B3D's numbers are from 'sireric', someone who is known to have worked on the R300 architecture. Anand's numbers are consistent with the numbers that NVIDIA is going around quoting in relation to R300.

Take your pick...

This kind of &*(^ tees me off. I remember us giving Anand the devil back on the old 3dfx forums for a few things like this. Why do people bother with doing so-called "product reviews" which are actually "product comparisons" which then make no attempt to be fair or accurate with the product being compared? Anand should have, as a matter of journalistic integrity, cross-checked his nVidia-supplied R300 information with ATI (Somebody like SirEric, for instance.) OTOH, if you have no intention of being fair but instead you are trying to create a false impression--making such a cross-check would be exactly what you avoid doing.

I for one was delighted to see the little blurbs ATI put up about SmartGart. I'd like to see companies do more of this relative to the technologies they are selling, so that this info comes from "the horse's mouth" so to speak. Otherwise, it's a case of the fox guarding the henhouse...

 

[WaltC]

SirEric Says...

Actually, this was so interesting I thought I'd transplant it in this thread, with SirEric's graces I hope....

SirEric said:

Simple answers:

1) The total number of possible vertex shader instructions executed is 65,026 on the 9700. The computation is as follows:
a) The 9700 supports any number in the range [0,255] as a loop count
b) There are 256 possible "static" instructions
c) Loops can exist in any of the first 255 instruction slots (not in the last, since you need another instruction, at least, to represent the looping "code")
So, total instruction execution = 255*255 + 1 = 65,026
I believe that most DX9 parts will execute something similar. Saying 64k is reasonable.

2) The R300 has 32 temporary registers in the vertex and pixel shaders (64 "total"). We currently "reveal" 12 in the pixel shader (not sure about vertex shader), following DX9 recommendations. We will raise that as caps bits allow or DX9 specs change.

Heh-Heh...I'd love to know what else is yet to be "revealed"...Not at all surprised, though, as it certainly makes sense for a lot of reasons that the actual design of the hardware far exceed DX9 requirements in as many respects as possible, where it makes sense to do so. This really helps with the design's longevity.

 

[Reverend]

I think Anand used what NVIDIA gave all the rest of the sites. And how NVIDIA arrived at those R300 specs are what they (i.e. NVIDIA) believe them to be. Obviously it is best to say something like "The following chart is what NVIDIA gave us and we cannot vouch for its accuracy"

When you don't know what to ask, it is difficult to give correct info.

Regarding Anand's apparent lack of math skills re bandwidth calculations, remember that we are talking about 4:1 ratio ... you can't forget the base 16GB/s bandwidth (hence 4 x 16GB/s = 64GB/s but this is a 4:1 ratio and you have to minus out 16GB/s, which will give you 48GB/s).

 

[T2k]

I think Anand used what NVIDIA gave all the rest of the sites. And how NVIDIA arrived at those R300 specs are what they (i.e. NVIDIA) believe them to be. Obviously it is best to say something like "The following chart is what NVIDIA gave us and we cannot vouch for its accuracy"

When you don't know what to ask, it is difficult to give correct info.

Why?
I don't understand something: can't he just ask ATI for confirmation, spend some time to collect information everywhere, etc? Did he request? If so, why it's not mentioned ("we got no answer" or something similar) anywhere and why numbers have been taken as hard facts?

It's simple laziness... shows only one thing: he didn't care about his own product (article), his own quality = he doesn't give a flying *** about his readers, has no respect at all.

The worst from Anand, ever.

It's disinformation, nothing else.

PS: I've spent several years from my life as a journalist (later news editor, etc) at TV, news papers.

 

[demalion]

Regarding Anand's apparent lack of math skills re bandwidth calculations, remember that we are talking about 4:1 ratio ... you can't forget the base 16GB/s bandwidth (hence 4 x 16GB/s = 64GB/s but this is a 4:1 ratio and you have to minus out 16GB/s, which will give you 48GB/s).

OK, now you've confused me.

4:1 is the ratio. If you have that ratio, your final bandwidth would have that ratio to the base bandwidth.

The ratio of 64 GB/s to 16 GB/s is 4 to 1.

The ratio of 48 GB/s to 16 GB/s is not (it is 3 to 1).

I don't follow the "subtract out" idea at all...if you could represent 3 bytes with one byte, you'd get 48 GB/s "effective" bandwidth (when that representation occurred) out of a 16 GB/s bus, and would call that a 3:1 compression ratio. Am I missing something?

 

[mech]

Yeah, exactly. If you have 64Gig/s of traffic, and it's compressed to a 4:1 ratio, it will be 16Gig/s, "effectively" 64Gig/s.

So what's this "subtracting out" bullshit?

 

[WaltC]

I think Anand used what NVIDIA gave all the rest of the sites. And how NVIDIA arrived at those R300 specs are what they (i.e. NVIDIA) believe them to be. Obviously it is best to say something like "The following chart is what NVIDIA gave us and we cannot vouch for its accuracy"

Seriously, though, how can it not occur to virtually any of those sites that the best place to derive information about ATI hardware is not nVidia, but rather ATI?.... Why would they simply use a chart nVidia supplied them which purported concrete information about an ATI product without verifying the information beforehand? I mean, they wouldn't have to verify the nv30 information since nVidia gave it to them. But to accept data about a competitor's product which casts that product in an unflattering light--without verifying it with ATI? How can such a thing be? Are they really that lazy? Or is there more to it...?

(Rev, mostly these are rhetorical questions which I know you can't answer, and I don't expect you to. I'm just amazed that this kind of thing goes on in web sites. Your suggestion is a good one, but I think an even better approach is to say, "Here is what nVidia told us about ATI's R300 hardware. We contacted ATI and received some different information, and so we present both accounts for your pleasure. You will find what nVidia told us about R300 below left. To the right is what ATI told us about its R300 graphics processor." 'Least that's the way I'd do it. Makes for a lot more interesting copy, too... )

 

[T2k]

Seriously, though, how can it not occur to virtually any of those sites that the best place to derive information about ATI hardware is not nVidia, but rather ATI?.... Why would they simply use a chart nVidia supplied them which purported concrete information about an ATI product without verifying the information beforehand? I mean, they wouldn't have to verify the nv30 information since nVidia gave it to them. But to accept data about a competitor's product which casts that product in an unflattering light--without verifying it with ATI? How can such a thing be? Are they really that lazy? Or is there more to it...?

(Rev, mostly these are rhetorical questions which I know you can't answer, and I don't expect you to. I'm just amazed that this kind of thing goes on in web sites. Your suggestion is a good one, but I think an even better approach is to say, "Here is what nVidia told us about ATI's R300 hardware. We contacted ATI and received some different information, and so we present both accounts for your pleasure. You will find what nVidia told us about R300 below left. To the right is what ATI told us about its R300 graphics processor." 'Least that's the way I'd do it. Makes for a lot more interesting copy, too... )

Agreed. This is exactly what I miss...