The GT300/Fermi Thread

[ElSel10]

I've got a couple of customers that actually use DP (and yes, they are in the financial world) and they have used GeForces/CUDA all the time now (starting development on 8800GT's) And they were indeed quite expecting that all their CUDA efforts would get a nice boost come Q2.

Well, that is the kind of market that Tesla is targeting, not GeForce.

[W1zzard]

i'm not even sure if nvidia managed to produce 400 working gpus so far

[saaya]

if nvidia is smart (they probably are) they put some spares on their gpu which is basically extra pieces of hardware that can replace pieces where defects are in the silicon. for example you could imagine having a 5th GPC cluster that can replace one with defects. if you do the proper math you can compute spare designs that are statistically going to increase your per-die yield even though such measures increase the die area

well its not only about being smart... do you really think they put backup logic on fermi? its so huge that some of the blocks of a fully fledged GF100 probably ARE seen by nvidia as backup units... its just that nvidia can enabled all of them on some few very good cores, while ati has backup logic that can only be enabled if something is broken... know what i mean?

Worst NDA ever if you ask me!!!

hahah definately

Nope. They were showing a car racing game (can't remember which one) on three 24" Acer 120Hz monitors.

mhhhh in 3d? did you see it with glasses? what was your impression? i saw this at the display show here in taiwan and thought the bezel was terrible in 3d, cause it cuts the 3d objects in 2 pieces in a really weird way...

Just because the heatsink and PCB isn't finalized doesn't mean the chips can't be produced.

i didnt say the chips cant be produced... i just find it odd that the ref pcb for gt300 isnt finished yet... cause nvidia previously said they have the cards already and just wait for the chips... or am i mixing something up?

As for "mass production", if you ask five people at NVIDIA about actual production, two will say it is in mass production, two will say it isn't and the final person will just shrug their shoulders.

hehhehe it all depends on how you define mass production i guess

No, HawX was a simple slide which was supposed to illustrate how the new ROP layout and design decreases overhead when using higher instances of AA.

do you know if that was actual game perf then or just rop perf meassured somehow?

This is GeForce 5800 2.0 if you ask me.

naaahhh come on, the 5800 was a nightmare... it was slower in some games than its predecessor FFS! really cant compare that... and back then gpu power really mattered, nowadays you can play at max details even with mainstream cards

You need to pay attention...
They left A.I. ( Artificial Intelligence ) enabled, thus the benchmark becomes CPU limited and differences between differently clocked or even maker CPUs arise.
Considering that in a sum of games the 5870 needs an i7 clocked at 3.7GHz to be CPU-limit-free, and there you have it.

hmmmm what exactly does ai enabled do? does it reduce iq?

I know I'm going to have a hell of a time if what I saw and my other info is wrong, but here it comes:
The GTX 380 will be nearing the 5970 in most games, and will have a significant advantage over the 5870.
Regarding the price... well, if you're a realist and live under no illusions, then it's easy to say that it will be priced similarly to the 5970, normally 50 bucks lower or so.

well that perfectly matches the FC2 results...

[neliz]

Well, that is the kind of market that Tesla is targeting, not GeForce.

Yes, but why spend $2500 on the performance of a $100 card?

i'm not even sure if nvidia managed to produce 400 working gpus so far

Hence my Q2

I'm expecting 7800GTX 512 like availability.

[gamervivek]

Which part are we talking about?

We'd be talking about ~160 to 175 GFLOP for GF100 versus ~500 GFlop for HD5870 in DP theoretically.

or versus the 200GFlOPS of 4850 theoretically.

p.s.- and iirc hitting theoretical DP peak is much easier than single precision so those "better" efficiency claims won't work here.

[neliz]

or versus the 200GFlOPS of 4850 theoretically.

depending on the application, most people can come up with 80% efficiency.

[Sam_oslo]

Too early to be too sure about the GPGPU performance of GF100, but everything is indicating a new era, with unheard performance of personal supercomputers just around the corner.

[BenchZowner]

hmmmm what exactly does ai enabled do? does it reduce iq?

Leaving A.I. enabled in FarCry 2 benchmark does the following:

1) Allows the CPU performance ( not actual amount of cores, meaning a similarly clocked Core i7 920 with HT off will perform the same as a 920 with HT on ) to have a decent impact in the benchmark's performance results because it partially simulates normal gameplay ( as in the PC "calculates" the enemies & other "live" objects moves & artificial intelligence stuff ).

2) In short-term benchmarks it can introduce a lil' bit more discrepancy between the results ( but that gets covered by the 3 runs of the bench )

It's like the Unreal Tournament "flyby" ( rendering ) and "botmatch" ( rendering + AI ) benchmark modes.

Yes, but why spend $2500 on the performance of a $100 card?

Why spend 500.000 on a Ferrari or even 1.000.000 on a Bugatti ?

[zalbard]

Why spend 500.000 on a Ferrari or even 1.000.000 on a Bugatti ?

Because it looks cool. Surely not the reason to get Tesla over GeForce.

[ajaidev]

I use milkyway@home and it uses DP and i am not so sure about Collatz Conjecture but i think it also uses DP not sure tough.

Our 4*5850 can crunch out more than anything nvidia has at the moment and what GF100 could do in the future. Also if i remember correctly my friends 4870 reached speeds of around 140-145 Gflops DP there....

[neliz]

Why spend 500.000 on a Ferrari or even 1.000.000 on a Bugatti ?

invalid comparison.

A Quadro has nearly identical stats and performance to a Geforce product. a Fiat 500 has neither of those compared to a ferrari. they only share the things that make them a "car" and not much else. I'm not comparing a S3 Trio to a Quadro here.

[Setix]

Anyone know when the next NDA lifts?

Hopefully we might see a street date, or perhaps some benchmarks of cards from vendors and not NVidia...as far as performance, the jurys out until Trubitar, Tom's Hardware, or HardOCP..etc..get a card to test

[neliz]

Anyone know when the next NDA lifts?

A tuesday after Cebit, that'd be March 9th at the earliest.

[kadozer]

Anyone know when the next NDA lifts?

Hopefully we might see a street date, or perhaps some benchmarks of cards from vendors and not NVidia...as far as performance, the jurys out until Trubitar, Tom's Hardware, or HardOCP..etc..get a card to test

There's probably an NDA about telling when the NDA lifts. Mid February perhaps?

edit nvm posted above right before me.

[Sam_oslo]

"Why spend 500.000 on a Ferrari or even 1.000.000 on a Bugatti ?" Is wrong question. Actually the opposite question applies to GF100.

Why spend 500.000 on a supercomputer or even 1.000.000 on a the dig blue, when you can get the same performance by spending $700 on a GF100?

[neliz]

Why spend 500.000 on a supercomputer or even 1.000.000 on a the dig blue, when you can get the same performance by spending $700 on a GF100?

$700 a supercomputer does not make.

It's application specific and the main reason you can only find one nvidia supercomputer in the supercomputer TOP-100. the "GPU supercomputer" is still a non-argument.

[BenchZowner]

"Why spend 500.000 on a Ferrari or even 1.000.000 on a Bugatti ?" Is wrong question. Actually the opposite question applies to GF100.

Why spend 500.000 on a supercomputer or even 1.000.000 on a the dig blue, when you can get the same performance by spending $700 on a GF100?

[kid with the why questions mode]And why should nVIDIA sell a "supercomputer card" for 700$ instead of 1500$?[/kid with the why questions mode]

[Sam_oslo]

$700 a supercomputer does not make.

It's application specific and the main reason you can only find one nvidia supercomputer in the supercomputer TOP-100. the "GPU supercomputer" is still a non-argument.

The new articulate of GF100 is aiming to dominate and change this landscape in 2010.
GPGPU is a very new technology, and has been in child-stage until now. It is really surprising to have 1 supercomputer in TOP-100 already based on the old technology. Stay tuned, everything indicating that you are going to get surprised soon.

[Olivon]

nVidia GF100 Fermi silicon cost analysis - BSN*

[saaya]

this whole thing sounds like its gonna be another 8800gtx vs 3870x2. except this time, the ati card was released first.

pretty much, yeah...

im surprised how confident ati is though...
they knew that fermi is more than double the hw logic, and they knew that nvidia always targets to reach the same perf with a single gpu card as their previous dual gpu card... and from what nvidia showed at the deep dive event, thats exactly where fermi perf is at...

I guess the drama goes on and the "deep dive revelations" only made it worse.

well yeah, nvidia didnt say if the card they showed was a 360 or 380, how many cores, what clocks etc... of course this created a wildfire of speculation...

Seems like fermi will not be as good as it was supposed to be and not as bad as it was hoped to be (by some)

whoever thought that fermi would be notably faster than a gtx295 wasnt anticipating but dreaming... :P

This is why you gotta applaud Nvidia for at least trying to do something different.

they are creating a massive expensive power hungry super performance part... hows that trying something different?

i think atis strategy to bring the same perf level down to much lower prices makes a lot more sense, especially if you consider how the pc industry has developed in recent years...

thx for all the interesting infos a couple of pages back LordEC911!

AMD/ATi gets ~60% more dies per wafer, Cypress vs GF100.
Then again, current yields put AMD/ATi ~3x higher.

40% yields for fermi? thats not going to be fully functional chips though... i can believe that its 40% 484core... but 40% 512core? nah...

I think the most impressive part about the Hair demo is that it was done TWO YEARS ago.

well.. i wasnt impressed by it tbh... it looks a lot like the mairmade nalu demo nvidia used a long time ago, and perf wasnt orders of magnitude higher... so i doubt we will see games simulate hair like that and it will remain a tech demo just like the thing nvidia showed before...

i hope im wrong though!

A fact? Really?
Two weeks ago I received my monthly investor update from TSMC which noted mass production has begun on nVidia's high end 40-nm GPU after months of delays caused by wafer production problems.

hmmm neliz probably meant card MP then... it takes a couple of weeks from wafers to chips... but im sure you know that
thx for the headsup!

nVidia's CEO also clearly stated at CES 2010 that the GF100 is now in mass production.

well, how would you prove he was wrong? im not saying hes lying, but he very well could be... hes famous for twisiting words around...

Leaving A.I. enabled in FarCry 2 benchmark does the following:

1) Allows the CPU performance ( not actual amount of cores, meaning a similarly clocked Core i7 920 with HT off will perform the same as a 920 with HT on ) to have a decent impact in the benchmark's performance results because it partially simulates normal gameplay ( as in the PC "calculates" the enemies & other "live" objects moves & artificial intelligence stuff ).

2) In short-term benchmarks it can introduce a lil' bit more discrepancy between the results ( but that gets covered by the 3 runs of the bench )

It's like the Unreal Tournament "flyby" ( rendering ) and "botmatch" ( rendering + AI ) benchmark modes.

so AI enabled basically means a FC2 run becomes a flyby, while AI disabled the bench tool will pretend somebody is playing the game and do all the ai calculations?
didnt know that, thx!

[neliz]

The new articulate of GF100 is aiming to dominate and change this landscape in 2010.

Hitler aimed to rule the world.. just saying...

GPGPU is a very new technology, and has been in child-stage until now. It is really surprising to have 1 supercomputer in TOP-100 already based on the old technology.

That's just as much as "ATI" has.. and their system is considered beyond poor

Stay tuned, everything indicating that you are going to get surprised soon.

You cannot surprise someone that knew more about Fermi back in July then you do now. You sound awfully optimistic.. too optimistic.

hmmm neliz probably meant card MP then..

No I didn't.. the investors "mass production" refers to production RAMP, they are still quite a few weeks off from real mass production.

[Sam_oslo]

..
You cannot surprise someone that knew more about Fermi back in July then you do now.

...

You didn't know nothing back then, and you still don't know anything. Nobody knows for sure yet, because the retail product (or any of details) is not out yet.

A couple of other kids have been speculating with this kind of childish arguments since July.

[saaya]

nVidia GF100 Fermi silicon cost analysis - BSN*

cypress=100$? that would mean its @ 33% yields? 0_o
and first he says gt300 is @ 40% yields right now, then he says quadro and tesla will subsidize geforce until nvidia can reach 40% yields with gt300... huh? :P

interesting read, but a bit weird... and i dont trust the numbers :P

[saaya]

No I didn't.. the investors "mass production" refers to production RAMP, they are still quite a few weeks off from real mass production.

ahhh so he WAS twisting words again
i thought there wasnt much you could interpret when saying MP has started... but i guess you can always claim your doing MP in a special way...
so MP starts in your company means you set up a row of meetings for the next weeks at which you will discuss how and when to start MP

You didn't know nothing back then, and you still don't know anything. Nobody knows for sure yet, because the retail product (or any of details) is not out yet.

A couple of other kids have been speculating with this kind of childish arguments since July.

take it easy, no need to get rancid

[Olivon]

interesting read, but a bit weird... and i dont trust the numbers :P

Right, difficult to know ... B1 rev. will be the retail version ?