NVIDIA Fermi: Architecture discussion

[Ailuros]

I thought he is joking

...and I wan't obviously sure about it.

 

[seahawk]

I wonder why the would neeed to do an A3 "to kill time". If A2 would have been good and just the process at TSMC not ready, they could have waited with the mass production and go with A2. NV must have gotten something from A3, be it higher clocks or better yields or both or whatever.

Time is actually working against them.

 

[silent_guy]

Let's see what silent_guy's answer to the above question will be. Unless of course it was some form of sarcasm, which I of course with my equal to 0 knowledge on matters like that cannot detect.

Sarcasm. My apologies for not being able to come up with more ridiculous sounding words to make it more obvious.

I don't understand how people can bring themselves to making these pointless statements. It's a pure crapshoot, not hindered by any amount of relevant experience.

The best you can do is speculate about the general likelihood about the nature of a metal spin. My personal take on this is clear: At least 95% of all fixes are pure functional bugs. A very generous 4% are timing related. 1% are yield. (On all the chips I've worked on, I know of only 1 metal ECO, out of thousands, that was yield related.) And that's really the best you can do. Yet reading the comments here, it's all about getting clocks up by some fantastic amount or doing magic tricks with yield. Some even persist in claiming that you can reduce leakage this way. All with a metal spin... Go figure.

 

[silent_guy]

... or whatever.

Give this man a cookie for reluctantly including the most probable option!

 

[Tamlin]

Oh dear, well I certainly hope developers target something less than a 5760x1200 resolution for their games because all we'll end up with is a lot of ugly. Do you really look forward to running the best that the Xbox and PS3 can do at the end of their lifetimes blown up in all its unrefined glory? There's something to be said for playing less demanding games at a higher resolution but there are much better uses of the available horsepower IMO. In this respect I think Nvidia's strategy is more potent in the longer term because at the end of it all Eyefinity is just upping the resolution.

I've been following this thread from the beginning and are content of just being a reader. However, I wish to clearify something here.

Eyefinity (and TH2G) is not about upping the resolution. Thats not what attracts the majority of the users. Its the change of aspect ratio, which gives more game content that is important. This is what we want to see when using Eyefinity or TH2G:
http://www.widescreengamingforum.com/screenshots/hl2-lc-th2go.php

Going from 1920X1200 in 16:10 to 2560X1600 in 16:10 is "just upping the resolution". It doesn't add to game other then giving a higher resolution image. Going to 48:10 in Eyefinity or TH2G gives you 3 times more game content. You can like it or not, thats subjective, but objectively I don't think you can disagree that seeing 3 times more game content is a major change of the gameplay.

 

[Ailuros]

I wonder why the would neeed to do an A3 "to kill time". If A2 would have been good and just the process at TSMC not ready, they could have waited with the mass production and go with A2. NV must have gotten something from A3, be it higher clocks or better yields or both or whatever.

Time is actually working against them.

Look at silent_guy's last answer. NV is obviously waiting for 40G yields to improve and no they're IMO not in the same position as AMD as yields are obviously better on a by X% less complex chip as Cypress. NV has fixed or added "whatever" with the minor A3 respin, but that whatever doesn't make the "killing time" notion any exaggeration in the end.

 

[aaronspink]

The best you can do is speculate about the general likelihood about the nature of a metal spin. My personal take on this is clear: At least 95% of all fixes are pure functional bugs. A very generous 4% are timing related. 1% are yield. (On all the chips I've worked on, I know of only 1 metal ECO, out of thousands, that was yield related.) And that's really the best you can do. Yet reading the comments here, it's all about getting clocks up by some fantastic amount or doing magic tricks with yield. Some even persist in claiming that you can reduce leakage this way. All with a metal spin... Go figure.

Yep, metals spins are almost ALWAYS functional. Either you can do slight rewires to existing logic or add small additions to the logic with available bonus gates. The later certainly is not positive as far as timing goes. If you are fixing anything physical related you're effectively trying to shut the barn doors after the horses have left, that SHOULD be taken care of in pre-silicon. Leakage is pretty much a transistor level issue as well.

The only other thing you really fix with a metal fix are things like ROMs though any sane design with ROMs has limited patch capability anyways.

You aren't going to get better timing, you aren't going to to reduce leakage, and unless you are totally fubar in your DRC (which means you have basically failed), you aren't going to improve yield.

 

[DemoCoder]

but objectively I don't think you can disagree that seeing 3 times more game content is a major change of the gameplay.

It's such a major change of gameplay that it's banned in many online multiplayer games (Valve games for instance). FOV is locked at 90 degrees max. It may be great for sims and car games, but I find it annoying otherwise.

 

[A.L.M.]

No one has already heard about this post?

Originally Posted by Rahja the Thief
At midnight (EST, -5 GMT), I will fill in the blanks.... MUHAHAHAA.

GF100 outperforms ATi's 5870 by 46% on average
GF100 outperforms ATi's 5970 by 8% on average

The GF100 gets 148 fps in DiRT2
The GF100 gets 73 fps in Crysis2
The GF100 gets 82 fps in AvP3

*GTX misnomers removed due to Business NDA*

GF100's maximum load temperature is 55 C.

The release week of GF100 is Mar 02nd

Blackberry ordered about a million Tegra2 units for their 2011 smartphone.
Apple ordered a few million Tegra2 units for the 2011 iPhone.
Nintendo ordered several million Tegra2 units for their next gen handheld (the DS2!)

*Removed: Under business NDA*

That's all for now kiddies! See if you can guess for the time being, each - represents a letter or number

Extra spoilers!

• GF100 and GF104 will feature a new 32x Anti Aliasing mode for enthusiast setups.
• GF100 and GF104 can do 100% hardware based decoding for 1080p BluRay and H264 playback.
• GF100 and GF104 feature full SLi capability and a new scaling method for rendering!
• GF100 and GF104 will provide full on chip native C++ operation for Windows and Linux environments. This will be further augmented with CUDA and OpenCL.
• GF104 will feature new technology designed for UHD OLED monitors!
• GF100 promises to deliver at least 40% more performance than the GTX295 for less money. GF104 promises double that.

http://www.guildwarsguru.com/forum/rahjas-freeeeeeee-t10420384.html

Any comments? Fake or not in your opinion?
I'm more for the "fake" option...

 

[Pressure]

No one has already heard about this post?



http://www.guildwarsguru.com/forum/rahjas-freeeeeeee-t10420384.html

Any comments? Fake or not in your opinion?
I'm more for the "fake" option...

Well, it would be more apparent the closer we get to the official launch of this much delayed chip whether or not there is any truth to these numbers. However, one thing that sticks out as a sore thumb is the Tegra2 "news" concerning Apple. They would much rather make their own solutions, so they are in control of every aspect of the chip. Especially considering they bought P.A. Semi.

 

[Ailuros]

I'm more for the "fake" option...

Given that the claim gets even more colourful in a post below that claiming:

Those tests were run using a Corei7 920, 6GBs of DDR3 1333, and an Intel 64GB SSD paired with a single GF100 card. The tests were run at 1920x1200 with 4x SSAA and 16xAF.

I'm actually wondering why you're even asking. I was wondering when this nonsense will appear here and I wish the performance differences would be the only laughable claim in that post, but the rest doesn't have to do anything with the topic here.

 

[LonelyMind]

It's such a major change of gameplay that it's banned in many online multiplayer games (Valve games for instance). FOV is locked at 90 degrees max. It may be great for sims and car games, but I find it annoying otherwise.

If it is seen as an unfair advantage in gaming when not everyone has it, is quite understandable.

What baffles me though is that you find a >90 degree field of vision annoying. You must have a hard time being an ordinary human being then, or do you use blinders in daily life to limit your vision?

Please explain how a wider field of vision is a bad thing, besides that it's rather new in gaming and as such takes some getting used to.

 

[CarstenS]

I doubt that a bit as some beefed up operand collector could feed SP and DP subblocks simultaneously, even if the needed register bandwidth cannot be sustained in all cases. Real code doesn't exist exclusively out of FMAs with 3 source operands.

Seems like a tradeoff Nvidia wasn't willing to make - (obviously I) dunno about the reasons though. Whatever the die space cost associated with this keep in mind that it's x16 for the whole chip.

From the implementation efficiency I would favor two 16 ALU subblocks, which can be chained together with some additional circuitry to enable 16 DP results. Basically similar to what ATI does with the 4 ALUs in a VLIW, just that the nv units have beefier ALUs (most important the multiplier) to start with and one can get away by coupling only two of them.

For some reasons, Nvidia doesn't seem to want or to not be able to go the AMD way. Their chips tend to be quite big anyway and so making the basic units even bigger for coupling is maybe a worse choice wrt to their basic architecture than had they gone also for a 5-way VLIW.

Plus, it would go completely against their traditional strategy of first implemention for experiments, then making the feature usefull and only after that to go fully along that route.

 

[Mize]

No one has already heard about this post?



http://www.guildwarsguru.com/forum/rahjas-freeeeeeee-t10420384.html

Any comments? Fake or not in your opinion?
I'm more for the "fake" option...

As I predicted! Right on time. Leaks or fake leaks...

 

[XMAN26]

I've been following this thread from the beginning and are content of just being a reader. However, I wish to clearify something here.

Eyefinity (and TH2G) is not about upping the resolution. Thats not what attracts the majority of the users. Its the change of aspect ratio, which gives more game content that is important. This is what we want to see when using Eyefinity or TH2G:
http://www.widescreengamingforum.com/screenshots/hl2-lc-th2go.php

Going from 1920X1200 in 16:10 to 2560X1600 in 16:10 is "just upping the resolution". It doesn't add to game other then giving a higher resolution image. Going to 48:10 in Eyefinity or TH2G gives you 3 times more game content. You can like it or not, thats subjective, but objectively I don't think you can disagree that seeing 3 times more game content is a major change of the gameplay.

Do you honestly believe you dont get more visually on the screen going from say 640x480 to 1280x960 or 800x600 to 1600x1200? I can tell you, having played at 1024x768 and then 1600x1200, there is a good deal more visually from games on the screen from the higher resoution, and whether or not you wish to believe it, Eyefinity and TH2Go does nothing more than the same thing. Ups the resolution to give you more visually on the screen(s).

 

[Ninjaprime]

No one has already heard about this post?



http://www.guildwarsguru.com/forum/rahjas-freeeeeeee-t10420384.html

Any comments? Fake or not in your opinion?
I'm more for the "fake" option...

So... the 5970 is only 38% faster than the 5870?

Edit: Honestly though, I wouldn't be surprised with the way NV PR has been acting if they came out with some random situation where crossfire scaling didnt work and with some old driver revision that they "think" or "expect" it was that way, even when they really know its not true.

 

[PSU-failure]

Why do you think that a 15 sq. mm cluster is "too low" for a 60 sq. mm part? GT218 is below 60 sq. mm (also on 40nm), and I very highly doubt its single (apparently 16-way only instead of 24 as the other GT2xx parts) cluster is approaching that size...

I wrote that based on a 10 sq. mm estimation and ninja edited after a check, which gave about 15 sq. mm.

It's too low considering computing (ridiculous throughput), but if we take the whole GPU into consideration it's too high at the same time as that would imply a ~100 sq. mm die, which I strongly doubt would be better than GT216 as that would require a quite high frequency.


Btw, I came across Golubev's IGHASHGPU home recently, and HD5870 has been given ~70% efficiency on md5 hash (down from ~78% on RV740 and RV770).

GF100 won't be able to to any better with md5 (expected ~1900M max at 1.5GHz for 512SP, ~2400M for HD5870) and SHA-1 (expected ~600M max, ~700M for HD5870). A that's Int, it will probably be better for DP, but even worse for SP.

 

[Silus]

So... the 5970 is only 38% faster than the 5870?

Tech-Report's review doesn't show much more than that on average. There's only a case or two where it actually goes beyond 50% faster.

Edit: Honestly though, I wouldn't be surprised with the way NV PR has been acting if they came out with some random situation where crossfire scaling didnt work and with some old driver revision that they "think" or "expect" it was that way, even when they really know its not true.

Every company does that. Going back to the HD 3870 X2, when AMD had nothing to compete with NVIDIA in the high-end, they used Call of Duty 4 to show how the X2 beat the 8800 Ultra, when there were many other games, much more demanding, where the X2 failed miserably against the year old 8800 GTX.

 

[FrameBuffer]

So... the 5970 is only 38% faster than the 5870?

Edit: Honestly though, I wouldn't be surprised with the way NV PR has been acting if they came out with some random situation where crossfire scaling didnt work and with some old driver revision that they "think" or "expect" it was that way, even when they really know its not true.

seems like source of supposed performance is none other than nV (as OP later comments):

"No, of course I didn't witness them firsthand. That is why I said, this is in house testing, and it should be treated as if tainted with bias.

However, the person who emailed them to me (and only a small section was emailed) is a Sr. Hardware Engineer at the Santa Clara facility. He shall remain unnamed."

then later ..

I know that GF104 outperforms the HD5970 hands down, but as for GF100... I just don't know tbh.

Kind of weird that he hasn't seen (witness) them firsthand, admits supposed results were emailed to him from nV then claims to know that the GF104 which by all accounts isn't even on the table yet "outperforms the HD5970 hands down"..

All while running at a max load temp of 55C.. umm ok..

 

[trinibwoy]

Eyefinity (and TH2G) is not about upping the resolution. Thats not what attracts the majority of the users. Its the change of aspect ratio, which gives more game content that is important

Yes, and exactly the same thing happens when you go from 4:3 to 16:10/16:9. Revolutionary stuff, really. It's really nice to have a wider vista in flight or driving sims and certain other titles but for other stuff it just doesn't work, do I really need to see "more wall" in a corridor FPS?

No one has already heard about this post?

Yeah but it's so full of BS it's not funny. 55c temps? Apple buying Tegra2 for a 2011 smartphone? He got the relative perf of HD5870 and HD5970 right but he could pull that off of any review.