AnandTech NV30 article

[Bigus Dickus]

The NV30 won't show up this year in anything but a few review specimens for websites. Just not gonna happen.

[nutball]

Originally posted by StealthHawk

the R9700 with 1-4 non-functional pipelines become R9500s, so that helps solve poor yields or R9700 by turning otherwise useless chips into sellable ones.

Are you sure about that? It would suprise me if the design was that fault tolerant.

[Fotis]

Originally posted by netviper13
With the HUGE increase in bandwidth we're talking about for NV30, it is plausible for lossless AA to occur. With current games at least, because they won't even come close to using up all that bandwidth and fillrate.
Again the 48gb/sec is effective bandwidth!!The radeon 9700 has about 60 when using fsaa with hyperz3!!David Kirk said that 256 bit bus is overkill and that they won't be using tile rendering.Don't get me wrong I want nv30 to be a greater product than radeon 9700 but we don't have any solid info in our hands!!If we take rumors then radeon 9700 pulls 325 million triangles/sec and nv30 just 200million/sec.See where it's getting nv30 cood do 500million triangles/sec we just don't know anything about it except for the cinefx!!
I really want to see what those nvidia&3dfx guys have cooked up!!This is will be an historic product in the game industry!!(at least it seems that way)

[StealthHawk]

Fotis,

where did you get the 60GB number for R9700 from? could you please show a link

[StealthHawk]

Originally posted by nutball


Are you sure about that? It would suprise me if the design was that fault tolerant.

some fanATIc at Beyond3d said that was the plan with the R9500. which is one of the reasons why it is rumored to have only 4 texture pipelines. personally i haven't heard any this on any news sites or anything, but i believe it

[Uttar]

Originally posted by Fotis

Again the 48gb/sec is effective bandwidth!!The radeon 9700 has about 60 when using fsaa with hyperz3!!David Kirk said that 256 bit bus is overkill and that they won't be using tile rendering.Don't get me wrong I want nv30 to be a greater product than radeon 9700 but we don't have any solid info in our hands!!If we take rumors then radeon 9700 pulls 325 million triangles/sec and nv30 just 200million/sec.See where it's getting nv30 cood do 500million triangles/sec we just don't know anything about it except for the cinefx!!
I really want to see what those nvidia&3dfx guys have cooked up!!This is will be an historic product in the game industry!!(at least it seems that way)

That's ridiculous! Last guess i've seen on effective R300 bandwidth is *30GB/Sec* - so PLEASE! Do not invent numbers!

As for Radeon 9700 pulls 325 million triangles/sec, that's simply marketing stupid numbers.

From http://service.madonion.com/compare?2k1=4202539

High Polygon Count (1 light) 79.1 MTriangels/s
High Polygon Count (8 lights) 16.9 MTriangels/s

And that's on an overclocked CPU beast! So it's less on a normal CPU.

I know, i know, there are lights and special effects - but what's the use of use theorical numbers if it drops to nearly nothing with lights? 350 million would be a 200% increase compared to Ti4600. Here, all i see with 8 lights is a 50% increase - not as impressive.


Uttar

[Fotis]

Originally posted by StealthHawk
Fotis,

where did you get the 60GB number for R9700 from? could you please show a link
I read it at rage3d a week ago and I'm a little bored to search for it right now.I know it's not the most reliable source but neither is nvnews!!

[Fotis]

Originally posted by Uttar


That's ridiculous! Last guess i've seen on effective R300 bandwidth is *30GB/Sec* - so PLEASE! Do not invent numbers!

As for Radeon 9700 pulls 325 million triangles/sec, that's simply marketing stupid numbers.

From http://service.madonion.com/compare?2k1=4202539

High Polygon Count (1 light) 79.1 MTriangels/s
High Polygon Count (8 lights) 16.9 MTriangels/s

And that's on an overclocked CPU beast! So it's less on a normal CPU.

I know, i know, there are lights and special effects - but what's the use of use theorical numbers if it drops to nearly nothing with lights? 350 million would be a 200% increase compared to Ti4600. Here, all i see with 8 lights is a 50% increase - not as impressive.


Uttar
Could you please show me the nv30 specs with raw bandwidth or even effective?That's right these are only rumors and pr!!:rolleyes:

My point is that the rumor and pr game can be played both ways!!

3dmark 2001 se is DX8 test so we can't have any clear results as 3dmark 2000 won't give the difference between gf2 ang gf3!!!
As for the triangle number noone seems to chalenge it and it's official number from ATI.Unless someone from a respectable site prooves otherwise I'll believe it.
Nv30 could be 128bit bus with 16Gb/s but with a drastic overdraw technic.I simly don't know.
I really don't care about number because they sometimes proove wrong so the best performance/feature/price card will get my money this december.
As I've mentioned before I'm looking forward to the first joined nvidia-3dfx engineers product.It should be exciting!!

[Uttar]

Originally posted by Fotis

Could you please show me the nv30 specs with raw bandwidth or even effective?That's right these are only rumors and pr!!:rolleyes:

My point is that the rumor and pr game can be played both ways!!



I was refusing the numbers you gave for the R300 to compare them with the NV30 - This number of triangles per second is SO theorical that it's practically useless.

And i searched for bandwidth on rage3D and i couldn't find a single news post saying the R300 had 60GB in effective bandwidth. Maybe i didn't search well, but for now, i still don't trust that number.

Uttar

[Bigus Dickus]

Originally posted by Uttar



I was refusing the numbers you gave for the R300 to compare them with the NV30 - This number of triangles per second is SO theorical that it's practically useless.

And i searched for bandwidth on rage3D and i couldn't find a single news post saying the R300 had 60GB in effective bandwidth. Maybe i didn't search well, but for now, i still don't trust that number.

Uttar

And you trust NVIDIA's "rumors" that 32GB/s real bandwidth will somehow be 48GB/s (or 60GB/s if you believe alexsok, who deleted his posts at Beyond3d claiming that 60 was the correct number... depends on which rumor you believe I guess)? If you're going to make up nonsense PR numbers, what's the difference between 48, 60, and 3849273?

[StealthHawk]

let's just say it's easier to believe 32 can be expanded to 48 or 60 than 20 can be expanded to 60.

but you're right, those numbers are essentially meaningless and i for one will always reserve judgment for where it counts - benchmarks.

[jbirney]

20 can be expanded to 60.

its been said before that HyperZIII and other bandwidth saving tech can reduce the required data by 1/3 so you get effectivly 3 times the usable bandwidht or 3*20 = 60.

But I think we will see pigs fly before that happens. Its the same none sense that RC posted....

[imtim83]

Take with a grain of salt.

[legion88]

Originally posted by jbirney


its been said before that HyperZIII and other bandwidth saving tech can reduce the required data by 1/3 so you get effectivly 3 times the usable bandwidht or 3*20 = 60.

But I think we will see pigs fly before that happens. Its the same none sense that RC posted....

Let us pretend for the moment that the "1/3" figure is true, which as you claim gives an effective bandwidth that's 3X normal. (In other words, you claim that it cuts the requirement TO 1/3, not BY 1/3). Now let's apply it.

Card A has XYZ fillrate but that fillrate requires 200 bytes/sec bandwidth to reach theoretical max. The card only has 100 bytes/sec. So at fillrate limiting condition, the performance hit is 50% of theoretical max since the card does not have enough bandwidth.

Okay. Enter the super duper HyperZ. The bandwidth requirement is now cut down to 200/3 = 66.667. Since the card has 100, it now has enough bandwidth to sustain the XYZ fillrate.

So what is the performance increase of HyperZIIII? 2X is the answer since the card will not perform any faster than its theoretical maximum fillrate.

These are all theoretical maximums here so the actual increase would be less than 2X, based on assuming large 50% performance hit without hyperZ, and, as noted earlier, the "1/3" number is allegedly accurate.

So even cutting the requirement to 1/3 does not provide the perceived 3X the performance increase.

Now is it really cutting the requirement to 1/3 or by 1/3?

It's been a long while but lets calculate the bandwidth requirement for a fillrate of 2.6 billion. 32-bit color, 1 color read/write, z-buffer--> 4 *3 = 12 bytes. Anything else that eats up bandwidth is ignored.

2.6 billion * 12 = 31.2 billion bytes/sec.

Now what is the bandwidth of 256-bit DDR 300MHz?
256/8 * 2 * 300 million = approximately 20 billion bytes/sec.

So if HyperZ can cut the requirement by 1/3, then the 300MHz 256-bit DDR would be sufficient to sustain a fillrate of 2.6 billion.

So, I guess that's where the "1/3" number comes from. Has someone been speculating again?

[Bigus Dickus]

Originally posted by legion88
So at fillrate limiting condition...

2X is the answer since the card will not perform any faster than its theoretical maximum fillrate.


All well and good, but not all situations are fillrate limited. ;)

BTW, I don't like bogobandwidth either. Just state the raw bandwidth, and then put up the benches to show how well you use it.