The GT300/Fermi Thread

[Origin_Unknown]

The 6GB's or ram costs a lot and i mean a lot, if this is a consumer card nvidia may have to sell it at a loss also i do believe 6 GB's of video ram for games is more than overkill its mega overkill.

I'm sure something like empire total war with one of the mod packs will look amazing with 6gb of ram for textures etc

or even the old classic - Oblivion with Quarls

[cowie]

i hope we never need 6g's for any games ever lol

[Lightman]

i hope we never need 6g's for any games ever lol

I hope you're joking

It's like Bill G. saying 640KB of RAM is all we ever will need

[orangekiwii]

except he might have poor internet


I know i'm stuck with 1 megabit for at least the next 7 years so I won't be able to download any of the new games without it taking a week (bandwidth caps to) and they won't sell games that large on 500 dvds cause its not "cool"

[trinibwoy]

32nm is a strawman. I said nothing of the sort.

Never said you did. I was referring to your comments on driver improvements.

That a chip with the same number of shaders, higher clocks, and higher bandwidth would consistently score lower then a 4870x2 (much less 4890CF) even though it doesn't have to interface through CF doesn't seem right. You keep mentioning a barely changed architecture - then where does the performance discrepancy come from?

The 5870 does not have more bandwidth than the 4870X2. It's 153GB/s vs 230GB/s on the latter.

Like I said, that's my optimistic (for NV) view. If you don't like it I could present my doom and gloom view.

Your optimistic view for Fermi is that its launch drivers would be crap and Cypress drivers would have improved tangibly by then? Hmmm, interesting take on optimism

i hope we never need 6g's for any games ever lol

I hope we do, textures are still pretty low resolution. But memory isn't the only limiting factor. Art is very expensive to create and it's not going to get any cheaper in the future.

[Hu1kamania]

The extra memory might be for more than just gaming. As nvidia (and ATi for that matter) try to take on extra responsibility in what they can render within the OS, that extra memory might come in handy. Future internet browsers are supposed to be GPU accelerated and we already have GPU accelerated adobe flash and shiny windows features that are bound to take up resources within the dedicated GPU memory and freeing up system dedicated memory. That doesn't account for 6g's of GPU memory, but it is a start.

[Marios]

Interview with Luciano Alibrandi, the Director of PR at NVIDIA
- Some people say maybe we shall see some working samples in CES
- I cannot answer that one. I wish we can actually... probably do some first claims of what actually the architecture does but nothing confirmed.

Nvidia CES 2010, Las Vegas
At CES 2010, NVIDIA® will be showcasing the latest NVIDIA® GeForce® GPUs powering the hottest PC games and CUDA™ apps...

LOL. Which are those "interesting latest NVIDIA® GeForce® GPUs"? Did I miss something interesting?

[Solus Corvus]

Never said you did. I was referring to your comments on driver improvements.

I don't see anything about driver improvements here. All I see is a thinly veiled accusation of being a fanboy and a snide remark about a 32nm refresh.

The 5870 does not have more bandwidth than the 4870X2. It's 153GB/s vs 230GB/s on the latter.

So it appears that there are some advantages with going for a multi-gpu highend. Doubled triangle setup and higher bandwidth. Are we expecting a GTX390 at the same time as the launch of the rest of the lineup?

Your optimistic view for Fermi is that its launch drivers would be crap and Cypress drivers would have improved tangibly by then? Hmmm, interesting take on optimism

Well if you think that drivers for a brand new arch are going to be more stable and more optimized then ones for an arch 3 generations old then I don't think I can really help you understand. And no, that's not part of my optimistic view. It's a separate issue.

[trinibwoy]

I don't see anything about driver improvements here. All I see is a thinly veiled accusation of being a fanboy and a snide remark about a 32nm refresh.

I must have missed where that comment was directed at you. It wasn't a reply to one of your posts was it?

So it appears that there are some advantages with going for a multi-gpu highend. Doubled triangle setup and higher bandwidth. Are we expecting a GTX390 at the same time as the launch of the rest of the lineup?

Yes, as has always been the case. I wouldn't bet on any sort of multi-GPU Fermi at this point. It also remains to be seen if one will even be necessary (or feasible).

Well if you think that drivers for a brand new arch are going to be more stable and more optimized then ones for an arch 3 generations old then I don't think I can really help you understand. And no, that's not part of my optimistic view. It's a separate issue.

We're talking in circles here - one minute you're saying Cypress will benefit from driver improvements, now you're saying Cypress drivers are stable (which is what I said in the first place).

In terms of your optimism for Fermi I believe you're referring to it needing to beat the GTX295 or else it'll be a failure. That's not an optimistic view, it's simply setting the low watermark against which Fermi will be judged on release. Personally, I think GTX 295 is far too low of a benchmark.

[Solus Corvus]

I must have missed where that comment was directed at you. It wasn't a reply to one of your posts was it?

It's pretty clear in context whose ideas you were mocking.

Yes, as has always been the case. I wouldn't bet on any sort of multi-GPU Fermi at this point. It also remains to be seen if one will even be necessary (or feasible).

I wouldn't bank on a multi-gpu fermi until a refresh. But Nvidia has claimed there would be a multi-gpu version, without stating any sort of timeframe.

We're talking in circles here - one minute you're saying Cypress will benefit from driver improvements, now you're saying Cypress drivers are stable (which is what I said in the first place).

I'm saying that there will be more improvement possible in fermi drivers then evergreen drivers. I'm not saying that there will be no improvement in evergreen drivers at all. Software development is a process. There is rarely a case where a piece of software is bug free or as optimized as it could be. Also the hardware did change a little between generations, even if it's not as obvious as things like shader count. Latencies change, the memory controller and scheduler are tweaked, etc. Between now and when fermi is released, and beyond, the drivers will continue being developed to take advantage of those arch tweaks, fix bugs, etc. It would be unusual for for development to just stop and make no improvements/fixes from this point forward.

In terms of your optimism for Fermi I believe you're referring to it needing to beat the GTX295 or else it'll be a failure. That's not an optimistic view, it's simply setting the low watermark against which Fermi will be judged on release. Personally, I think GTX 295 is far too low of a benchmark.

I'm not the one that introduced the idea of being barely faster then a 295 is enough. I always assumed it would be significantly faster.

My optimistic view is 40% faster then evergreen with perfect hardware/driver execution. Pessimistic would be a little faster then 295 with some sort of problem that needs a respin, board revision, or several driver versions to fix (it's a complex new arch, don't say it's not possible). Realistically I'd say 30% faster then 5870 with some minor growing pains.

Since you like computerbase so much I figured I'd share these graphs from the same review with everybody:

[JohnZS]

I would suspect that the 6GB Quadro will remain a 6GB Quadro and will not filter down to consumer level, however (and this is a big however). Do you know that the FX5800 Ultra 4GB is infact known to consumers as the ASUS M.A.R.S 4GB GTX 295 So who knows maybe some board partner might make a consumer incarnation of the 6GB behemoth which is coming in Q3 2010 in Quadro form.
Now, from what I have read, RAM is useful not only for texture memory and all the fancy video effects and post processing effects, but also GPGPU wizardry too. Apparently stuff like OpenCL, CUDA, DirectCompute Physx etc all love the extra RAM and GPU processing power. So perhaps 3GB cards at the high end are not a "bad" thing.
Who would complain playing a DirectX11 title with high resolution textures (2GB's worth) and then a shedload of computational work which also uses up some of that VRAM for fast execution. (say a further 256MB's worth). Extra VRAM is good when you have it, and certainly a whole world better than hitching, pausing and stuttering if you do not have the VRAM.
Personally I cannot see nVidia releasing an incarnation of the Fermi with LESS than 1GB of VRAM (with the GTX 360 or greater nomenclature), if yeilds are bad and nVidia get bin happy and make the Fermi GPU go down throughout the range to the lowly numbers of GTS 350, 340 then we could see some 768MB models appearing.
IMHO 1.5GB is going to be the base model VRAM with the high end models having 3GB.
This isn't confirmed yet specifications for Quadro and Teslar almost make it safe to assume this will be the desired memory configuration for Fermi.
And yes, I think it is fair to say that we all want the expensive electricity bil, large carbon footprint FPS pushing Monster Fermi
John

[cowie]

Interview with Luciano Alibrandi, the Director of PR at NVIDIA
- Some people say maybe we shall see some working samples in CES
- I cannot answer that one. I wish we can actually... probably do some first claims of what actually the architecture does but nothing confirmed.

Nvidia CES 2010, Las Vegas
At CES 2010, NVIDIA® will be showcasing the latest NVIDIA® GeForce® GPUs powering the hottest PC games and CUDA™ apps...

LOL. Which are those "interesting latest NVIDIA® GeForce® GPUs"? Did I miss something interesting?

were getting there were getting there

well maybe at ces .....oh jeez i sure hope so the natives are getting restless

[trinibwoy]

My optimistic view is 40% faster then evergreen with perfect hardware/driver execution. Pessimistic would be a little faster then 295 with some sort of problem that needs a respin, board revision, or several driver versions to fix (it's a complex new arch, don't say it's not possible). Realistically I'd say 30% faster then 5870 with some minor growing pains.

That sounds about right. But keep in mind 40% is Nvidia's number and that puts it right on top of the HD5970 which would pretty much render that part irrelevant (unless you have a hankering for Eyefinity).

[eric66]

http://bbs.chiphell.com/viewthread.php?tid=64351

Post #66 cfcnc:
顺路发个消息，TSMC的圣诞礼物- Fermi A3已经顺利出样
Conveniently to issue this news, TSMC to give Christmas present -Fermi A3 already smoothly to come out shape(?)

Post #69 tomsmith in reply:
有个不太好的消息，符合2070 频率目标的比例还不理想，做成2050 的稍微多一点
To have this not so good news, in accordance with the 2070 frequency goal proportion still not ideal, to turn into 2050 a little bit too many.

Quick analysis: A3 out in time for christmas, plenty chips of C2050 ability(~1200Ghz shaders), not enough of C2070 quality available(~1400Mhz shaders).

more rumours partially good

[Solus Corvus]

That sounds about right. But keep in mind 40% is Nvidia's number and that puts it right on top of the HD5970 which would pretty much render that part irrelevant (unless you have a hankering for Eyefinity).

That's their number for full target clocks and shaders. Seeing as how they seem to be having some sort of difficulty, I wouldn't be surprised if the numbers are somewhat lower.

As for the 5970, whether or not you believe evergreen drivers will get better, crossfire will and already has gotten better in the newer drivers. Also, as I have mentioned, I don't think AMD will wait till the very end of the generation to release a refresh like last time. They have been executing well lately and I doubt they will sit around and lose the initiative.

If Fermi is 40% faster then evergreen and nothing changes on the AMD side then GTX380 would be well placed in the market. But if that is going to be the case remains to be seen. There are some potential issues, as I have explained.

[Milos-stancene]

Nvidia Is Happy With Performance of GeForce GF100 “Fermi” Graphics Card.

http://www.xbitlabs.com/news/video/d...hics_Card.html

[JohnZS]

Nvidia Is Happy With Performance of GeForce GF100 “Fermi” Graphics Card.

http://www.xbitlabs.com/news/video/d...hics_Card.html

Interesting read

The flagship Fermi graphics processor will feature 512 stream processing engines (which are organized as 16 streaming multi-processors with 32 cores in each) that support a type of multithreading technology to maximize utilization of cores

Hmmm so any takers on this.

360 = 448
380 GTX = 480
380 Ultra (limited quantity, very expensive, loud, hot, and a large carbon footprint and a high electricity bill. yet Flagship Fermi) = 512

I really do hope the 448 pipes are a "Tesla only" bin. And that we will see 360 with 480 and 380 GTX with 512!!

Who knows if nVidia do what they did last round, we COULD see the GTX 360 with 448 pipes and the GTX 380 with 512 pipes and then later on nVidia do a rebrand of the 360 say 360-480 with 480 pipes

Who knows?!?

But what we do know is that the Fermi is sounding like it is going to be the FreakingFastFPSPushingMonsterMi

John

[LinkinParkBoy]

the company is not only working hard on the new chip itself, but is also developing “perfect” drivers in an attempt to provide ultimate experience for the end-user.

Fap fap

[RPGWiZaRD]

Who knows if nVidia do what they did last round, we COULD see the GTX 360 with 448 pipes and the GTX 380 with 512 pipes and then later on nVidia do a rebrand of the 360 say 360-480 with 480 pipes

I'm also thinking this will be the case that GTX 360 will have 448. Also if you concider the price NVIDIA needs a card that should target $399 to not be far too expensier than the HD 5870 while still beating it noticably if topend card GTX 380 would cost $599 that's roughly the same or say 10% from a 5970 in performance. With 480 pipes nvidia couldn't target it at $399, more like ~$499 and be too close to GTX 380 and this will hurt its sales and it really needs a card that can at least come close to 58xx series cards in price as not that many people buys $400+ cards. Better to save 480 pipes for a future GTX 375 or whatever that will counter AMDs higher clocked 5890 card.

[cowie]

i'm sure alot of people would ebay there 5970 IF nv comes out with a card around 599 that had the same or similar performance.
as for nv making sure the drivers are "just right"(thats yet to be seen ofc)....thats good the other company did not with there last batch of cards.

that seems to be the biggest issue most consumers of the 5xxx complain about,and first hand i can tell you no matter how great and fast the hardware is,its means nothing if you have to workaround many driver issues.

[JohnZS]

Not meaning to be controversial here, but the Forceware 195 drivers are not exactly what I would call "perfect".

1) No PCI-E Gen 2.0 support on X38 and X48 chipsets
2) Stutters in 3dmark Vantage

I have reported these issues to nVidia and ASUS, but still there is no fix other than downgrading my drivers. If I want PCI-E Gen 2.0 on my X38 I have to use Forceware 186.18WHQL.

Hopefully nVidia will resolve this issue in the next release.

RPGWiZaRD
I can see exactly what you mean, from a competative pricing point of view 448 pipes on the 360 makes sense, I am guessing if any of the hype is to believed Fermi Pipes are much more efficient and powerful than G200 pipes, so we may see GTX295 performance on the GTX360?

John

[RPGWiZaRD]

My predictions if we say GTX 295 is in avg 10~20% faster than HD 5870 the GTX 360 is maybe 10% slower than GTX 295 and like 10% faster than HD 5870 and then I think HD 5890 could be made like 10~15% faster than HD 5870 why this GTX with 480 pipes would come handy later. However that Fermi is actually performing this well is still a bit difficult to believe and I wouldn't wanna become disappointed either but this is really my most optimistic view about Fermi because nvidia really needs to meet these targets at minimum or else they'll get a hard time and I'm sure nvidia knows this.

[cowie]

@ john
the new nv drivers did borke more then fix for me but nothing major.

i think it was a ion 210/220 thing to put them out ....what else lol

[Piotrsama]

Nvidia Is Happy With Performance of GeForce GF100 “Fermi” Graphics Card.

http://www.xbitlabs.com/news/video/d...hics_Card.html

Another chapter of Nvidia's "Wait!, performance will be great, don't buy ATIs, really guys, great performance coming soon".

The words come from Luciano Alibrandi (the head of Nvidia public relations department). Very reassuring.

[RPGWiZaRD]

As far as the drivers go, nvidia made huge improvements starting from 195.62 so it seems like a good start and that they're serious about driver improvements now. All drivers since 19x.xx until 195.62 has been a huge disappointment, 195.62 is the first driver I see as an improvement since 186.xx/187.xx series. 195.62 is more similar to 186 or 187 series than 19x.xx. I'm guessing they jumped back to the older drivers and started working from there instead. NVIDIA has different driver teams working on different versions simlarly so that's one reason to the back and forth jumping number revisions and why you can get such a different results depending on driver used.