NVIDIA GF100 & Friends speculation

[Sontin]

In a paper launch, you have real reviews from independent (and hopefully, unbiased) 3rd parties.

No, this is a "soft launch". How can you call this a PAPER launch?
You get all information through independent reviews.

To avoid confusion, may be I should rename my definition of powerpoint launch to vapor/smoke launch. :smile:

You can launch whatever you want with Paper / Slides. Look at AMD's "Open Phyics" announcements. This is a good old paper launch. Or nVidia's launch of their 40nm lineup in last june. You don't need a real product for a paper launch.

 

[Creig]

The way I understand launches to be:

Hard launch - Product is available on the date set forth by the manufacturer and is available in quantity.

Soft launch - Product is available on the date set forth by the manufacturer but is available only in limited quantities.

Paper launch - Product is not available on the date set forth by the manufacturer but will be available in the near future.

Vaporware - Product is not available on the date set forth by the manufacturer and is not predicted to be available in the near future (or at all).

 

[trinibwoy]

Seems difficult to understand this design.

gtx480 / hd5870
14,500 / 17000 single tex fillrate
36,600 / 72000 multi tex fillrate
630 / 1100 pixel shader
512 / 430 vertex shader simple
340 / 420 vertex shader complex
308 / 310 particles test
10400 / 9700 sm 2.0 score
11800 / 10600 sm 3.0 score

Please excuse me if this is an irrelevant comparison to make, I am trying to gain a better understanding of these numbers.

I think it's relevant if only for the questions it raises. But 3dmark06 obviously isn't running the workloads that either Fermi or Cypress were meant for so in that respect we aren't going to get any answers by looking there. The half-rate blending is a mystery and lower performance in the pixel shader test than GT200 is a red flag.

 

[neliz]

Right now, as all evidence is given, it is a paper launch of the purest form.

Just ask an AIB when they will start shipping the boards, betcha it isn't before the 26th.

I think it's relevant if only for the questions it raises. But 3dmark06 obviously isn't running the workloads that either Fermi or Cypress were meant for so in that respect we aren't going to get any answers by looking there. The half-rate blending is a mystery and lower performance in the pixel shader test than GT200 is a red flag.

Couldn't the chip become a victim of it's own down-cutting? I've we take our modern Quasimodo, Cypress LE. There are so many factors that could affect it's performance, it could be a simple driver issue while the team tries to optimize for the current setups.

 

[DegustatoR]

Right now, as all evidence is given, it is a paper launch of the purest form.

Paper launch is when you say that you start selling something and no one can buy it. You will be able to buy GTX480 at PAX. Thus it's not "a paper launch of the purest form".

 

[neliz]

Paper launch is when you say that you start selling something and no one can buy it. You will be able to buy GTX480 at PAX. Thus it's not "a paper launch of the purest form".

Where do they actually say you can buy one? only thing I can read about is that you might win a system.

 

[PSU-failure]

BTW, just wondering ... are any of the people here going to Pax with the intention of buying one if it's available for sale? (As been hinted at.) If so, any chance of posting benchmarks on the 27th? (I don't buy into all the benchmarking guidelines noise, but internet news sites have a huge conflict of interest without explicit shenanigans ... so user reviews are always nice.)

There probably won't be any GTX400 board to sell, all points to one prize, perhaps some more...

So, if you want to buy one there, it will be from the winner(s)

 

[neliz]

There probably won't be any GTX400 board to sell, all points to one prize, perhaps some more...

So, if you want to buy one there, it will be from the winner(s)

The winners are not getting their system right away either (big surprise!)

 

[Mintmaster]

The half-rate blending is a mystery and lower performance in the pixel shader test than GT200 is a red flag.

Half rate blending isn't much of a mystery. In the real-world 192 GB/s won't give you much more than 15 GPix/s of alpha blending with a Z-buffer and realistic textures, and NVidia was expecting a higher clock.

But you're right: The low PS score is odd. It should be noted that GT200 scored unusually high vs. RV770 compared to other pixel shader tests, so whatever allowed it to do so may have been removed in Fermi. Could just be an immature compiler, too.

 

[CarstenS]

I like this "launch" because both companies are delivering equal performance. It's not as lop-sided as 9700Pro vs. Ti4600/GeforceFX or 8800GTX vs. 2900xt/hd3870. Now that the playing field has somewhat leveled out, it should be exciting to see who is the first one to release legendary next generation technology on the next go around. Regarding GF100 design, could anyone help out and make an assessment of those 3dmark06 Feature tests explaining the scores relating to the architecture? I've read this from Neeyik at futuremark thus far:

Seems difficult to understand this design.
edit: (numbers)
gtx480 / hd5870
14,500 / 17000 single tex fillrate
36,600 / 72000 multi tex fillrate
630 / 1100 pixel shader
512 / 430 vertex shader simple
340 / 420 vertex shader complex
308 / 310 particles test
10400 / 9700 sm 2.0 score
11800 / 10600 sm 3.0 score
Please excuse me if this is an irrelevant comparison to make, I am trying to gain a better understanding of these numbers.

Neeyik seems to go more by what the names of the tests suggests than what the actual results imply.

For example: Shader Particles is almost 50% higher on 2900 XT vs. 3870 and 20% higher than on 4770, which suggests a dependancy not only on shaders but also on internal AND external bandwidth (4670 is about 60% of 2900 XT here).

PS scores almost the same on 5850 as on GTX 285, suggesting something more than just shader dependancy. Also, 4890 is 10% higher than 5770, despite same teraflops. 5450 scores a third compared to 5570 instead of a fourth (same clocks, 80 vs 320 shaders).

Complex Vertex Shaders, although twice as fast on 5870 vs. 5770 as it should be, are only 10% more compared to 4890. I suspect an artificial limitation in the drivers for vertex work.

 

[Tahir2]

We do not know how many cards NVIDIA have for its launch on March 26th.

We are expecting them to sell out almost immediately simply because NVIDIA will have less cards than there is demand.

If NVIDIA are not able to fulfill orders from AIB's and launch partners within a few short weeks then the launch on March 26th is a token launch even if NVIDIA have just one card or 5,000 for its entire customer base.

How many cards does NVIDIA need before we can say for sure that this is a hard launch?

A 100,000 would be a conservative estimate.

(I apologise for the language above - just wanted to be absolutely sure everyone understands that I am not spouting NVIDIA hate - just (un)common sense).

 

[trinibwoy]

Half rate blending isn't much of a mystery. In the real-world 192 GB/s won't give you much more than 15 GPix/s of alpha blending with a Z-buffer and realistic textures, and NVidia was expecting a higher clock.

GT200 achieved a measured blend rate of over 16GPix/s on 160GB/s. But maybe you're right and in real workloads it won't matter. Seems odd though.

 

[onethreehill]

AMD Is Getting Ready For GF100 - AMD's ATI 10.3a Catalyst Driver Update will be available to download on March 17th from undgerground.amd.com
http://www.hardocp.com/article/2010/03/16/amds_ati_103a_catalyst_driver_update

 

[GZ007]

GT200 achieved a measured blend rate of over 16GPix/s on 160GB/s. But maybe you're right and in real workloads it won't matter. Seems odd though.

As gpu-s have now biger on chip cache with speeds over 1 TB/s shouldnt the bandwith mean much less than before ?
The radeon 5870 doesnt show much much speed gain with increased bandwith.
GF100 has 768 KB very fast read/write cache. Couldnt that be used also as small on chip tile cache for less memory acess.

 

[XMAN26]

Press (EU at least) will get the cards in two days.

I can confirm 1 review site here in the states has both cards and is busily working on their review for the 26th.

 

[rjc]

The way I understand launches to be:

Hard launch - Product is available on the date set forth by the manufacturer and is available in quantity.

Soft launch - Product is available on the date set forth by the manufacturer but is available only in limited quantities.

Paper launch - Product is not available on the date set forth by the manufacturer but will be available in the near future.

Vaporware - Product is not available on the date set forth by the manufacturer and is not predicted to be available in the near future (or at all).

Why is the "Hard Launch" so deified?

The whole idea behind "Hard Launch" is to intimidate suppliers into charging less and customers into paying more, ie increasing long term profit. It is a subconscious demonstration to the counter party of who exactly is in charge in the transaction.

Soft launches and Paper launches should be celebrated, it means the suppliers and customers are much more in control and getting a good deal.

For example in this case it looks like the testing (ASL?) employee who leaked the pictures of the dead Fermi's is getting fired. In future Nvidia negotiations Nvidia will hammer the company about their lack of security and demand lower prices for the backend. Similarly AIBs all over know about the Fermi allocations barely making thousands for their initial worldwide shipment, if anyone breathes a word they personally risk termination and threats of discriminatory prices or being cutoff in future launches. Intimidated in future negotiations the AIBs will likely be willing to pay more for the same products and be forced to pass on higher prices to their end customers.

Must say though congratulations to Nvidia on this, they had a bad hand but looked to have played it quite well considering.

 

[trinibwoy]

As gpu-s have now biger on chip cache with speeds over 1 TB/s shouldnt the bandwith mean much less than before ?
The radeon 5870 doesnt show much much speed gain with increased bandwith.
GF100 has 768 KB very fast read/write cache. Couldnt that be used also as small on chip tile cache for less memory acess.

Depends on the workload. A cache only amplifies bandwidth if there's reuse. IIRC there's just too much turnover in framebuffer accesses for such a small cache to do much good there. The biggest gains will probably still be with texture/UAV reads or with anything that's read soon after it's written - tessellator output etc.

 

[Bouncing Zabaglione Bros.]

Why is the "Hard Launch" so deified?

Because a few years back, everyone got fed up with paper launches and spoiler launches where there was months of hype and spin before there was a product to buy or even read reviews of.

It doesn't do anyone any good when companies "sell" products, but don't have anything you can buy for months.

 

[nAo]

Depends on the workload. A cache only amplifies bandwidth if there's reuse.

Yes and no. Think about compressed textures that are stored uncompressed in texture cache.
Individual texels might be fetched only once, but the cache is effectively acting as a bandwidth amplifier.

 

[Tchock]

Paper launch is when you say that you start selling something and no one can buy it. You will be able to buy GTX480 at PAX. Thus it's not "a paper launch of the purest form".

You can buy leaked samples of Gulftown and Magny-Cours months before they really launch too... and Magny was actually proper OEM sampling.

Official or not does it matter? Even if you get a GTX400 at PAX, who's gonna take responsibility if it dies out?