The GT300/Fermi Thread - Part 2!

[kemo]

15 days means teasers numbering near 30-50 all with a resolution of 2560x1920 or higher. Where possible tessellation will be maxed out otherwise use of ultra high texture will be done to bottle neck the 1GB memory on the 5870/5970

Who cares if it is benched at 2560x1600 as long as it can provide playable frame rates with most of those games

[SKYMTL]

We're talking about clocks speed on the memory, not the amount. If I'm missing something here, you're being too subtle for me.

Both the maximum memory allotment and the speeds would be affected if the controllers were "weak" as was alluded to before. If there were issues with the controllers, you wound not see each one paired up with up to 1GB of GDDR5.

[saaya]

Same performance of course. Having the same bandwidth doesn't tell you anything. And there is no correct answer until we see hard numbers which is why I was asking why you think they need higher memory clocks. Maybe 160GB/s is more than enough. At least it's way higher than the previously rumoured 128GB/s.

well according to the numbers weve all seen, the 470 needs every little bit of additional perf it can get since it seems to be behind the 5870 in most scenarios...

Nothing is released to public yet, either the GPU or the box , some PR-hungry people leaks these unfinished products. They will fix it by the time of release.

Consider it as a beta-box.

which is what, in 2 weeks? they dont even have the box layout finished and plan to have thousands of them printed AND already shipped to distries and shops in 2 weeks? im finding it harder and harder to believe that the march 26 event will be a propper retail launch...

Considering the Tesla versions will have between 3GB and 6GB of GDDR5, I don't think the memory controller is anywhere near an issue.

capacity and speed are two different things...

Both the maximum memory allotment and the speeds would be affected if the controllers were "weak" as was alluded to before. If there were issues with the controllers, you wound not see each one paired up with up to 1GB of GDDR5.

right, there is some relation between speed and capacity, but just because there are 3 and 6gb cards you cant derive what speeds 1.5gb cards will clock at... for cpu/nb memory controllers this relation is common, but can you remember any differences in memory clocks in 1gb vs 512mb cards? and 512 vs 256, 128 vs 256 etc that was NOT based on how well the memory clocked? i cant!

[Iconyu]

Both the maximum memory allotment and the speeds would be affected if the controllers were "weak" as was alluded to before. If there were issues with the controllers, you wound not see each one paired up with up to 1GB of GDDR5.

I'm pretty sure if you design a memory controller to take 6gb in a certain configuration, you'd have to be pretty hard pressed to mess it up. That sounds like a outright logic error,

Much easier for it to fail at a rated speed than be unable to take larger memory IC's, it's a 384bit bus anything below 1gb and they wouldn't even release it.

Maybe I should go back to my original guess and just say nvidia are being cheap, if yeilds are sucky then theres really no point losing more money per card than they already are.

They could be sand-bagging, but I'm hearing this batch on release might be the only batch unless they can fix the issues. So once this batch of Fermi sells out on the 27th you won't get one until June-July.

[SKYMTL]

there is some relation between speed and capacity, but just because there are 3 and 6gb cards you cant derive what speeds 1.5gb cards will clock at...

I never said that you could derive speed from the amount of memory placed on higher end cards. However, both memory speeds and the amount of memory associated with each controller has an effect on stress levels.

If the card uses lower-end GDDR5, it is likely due to TDP and price issues rather than a design flaw IMO. Especially considering the GT 240 already uses higher binned modules in some cases.

[saaya]

I never said that you could derive speed from the amount of memory placed on higher end cards. However, both memory speeds and the amount of memory associated with each controller has an effect on stress levels.

If the card uses lower-end GDDR5, it is likely due to TDP and price issues rather than a design flaw IMO. Especially considering the GT 240 already uses higher binned modules in some cases.

tdp? i find it hard to believe that the memories power consumption is an issue

its possible that since its 320bit, they dont need high clocks, more bw barely scales... so they go for the cheapest low clocked stuff... makes sense, thats kinda what they did with gt200, very wide interface and then use cheaper gddr3.

[damha]

If the card uses lower-end GDDR5, it is likely due to TDP and price issues rather than a design flaw IMO. Especially considering the GT 240 already uses higher binned modules in some cases.

It most likely is a TDP issue, and that itself can be identified as a flaw. Not of the memory controller, rather the GPU design.

I think nVidia has already started a new radical GPU design. They know they cannot go on forever in this path. I don't know what too much is to them, but 3.2 billion transistors is 1 billion too many in my book. If ATI can get you the same performance out of 2.15 billion of the little guys then you have got a problem that cannot be ignored.

[eric66]

Who cares if it is benched at 2560x1600 as long as it can provide playable frame rates with most of those games

well cause when 2 gb version of 5870 will be out nvidia will be singing different song

[zalbard]

http://forum.donanimhaber.com/m_38364776/tm.htm

Absolutely hilarious box design!
And the card looks like a render. Not very impressive design anyway...

I never said that you could derive speed from the amount of memory placed on higher end cards. However, both memory speeds and the amount of memory associated with each controller has an effect on stress levels.

If the card uses lower-end GDDR5, it is likely due to TDP and price issues rather than a design flaw IMO. Especially considering the GT 240 already uses higher binned modules in some cases.

I don't think the price difference is huge. Look at 5770's memory. Can overclock even higher than 5870's GDDR5. And that's a budget card.
TDP? Yeah, maybe... But not very likely imo...

[Arxagelos]

Dual Geforce 400 to come later

http://www.fudzilla.com/content/view/18038/1/

[-Sweeper_]

Dual Geforce 400 to come later

http://www.fudzilla.com/content/view/18038/1/

fudzilla makes news out of obvious things

[Boissez]

Dual Geforce 400 to come later

http://www.fudzilla.com/content/view/18038/1/

Misleading title really: Nothing in that article suggests that a dual fermi is about to come. It's all if's and buts... not even rumor worthy.

[Sam_oslo]

If I'm not mistaking, Fudz is trying to say the double Fermi is "postponed" to some unknown date in late Q2. Isn't he trying to say, forget it for this round?.

[Boissez]

tdp? i find it hard to believe that the memories power consumption is an issue

its possible that since its 320bit, they dont need high clocks, more bw barely scales... so they go for the cheapest low clocked stuff... makes sense, thats kinda what they did with gt200, very wide interface and then use cheaper gddr3.

High speed GDDR5 consumes power like a mofo - it's the reason the 4870 had such a horrendous idle power consumption. And also the reason the 5970 only has 1000 Mhz GDDR.

[onethreehill]

Zotac planning GeForce GTX 480 AMP! version
http://www.nordichardware.com/en/com...p-version.html

[saaya]

High speed GDDR5 consumes power like a mofo - it's the reason the 4870 had such a horrendous idle power consumption. And also the reason the 5970 only has 1000 Mhz GDDR.

that doesnt make sense, then why does the 5870 have such low idle numbers with even faster memory? i find it hard to believe that clocking gddr5 memory higher increases the power consumption notably... a 5870 consumes less under load than a gtx285 iirc, and dont tell me the gpu is running that cool that it can even make up for the gddr5 consuming more power than the gddr3 on the 285...

[Boogerlad]

isn't it because the 5XXX series have better idle clock changing mechanism to reduce power consumption when idle?

[Chumbucket843]

that doesnt make sense, then why does the 5870 have such low idle numbers with even faster memory? i find it hard to believe that clocking gddr5 memory higher increases the power consumption notably... a 5870 consumes less under load than a gtx285 iirc, and dont tell me the gpu is running that cool that it can even make up for the gddr5 consuming more power than the gddr3 on the 285...

its downclocked quite a bit. 4870 idle was from that and terrible power gating. its ~3 watts per IC with gddr5 and there are 12 on the 480. thats 35 watts. tesla was delayed because they are waiting for hynix 2Gb IC's. even then with 6GB you are still consuming a lot of power.

[saaya]

its downclocked quite a bit. 4870 idle was from that and terrible power gating. its ~3 watts per IC with gddr5 and there are 12 on the 480. thats 35 watts. tesla was delayed because they are waiting for hynix 2Gb IC's. even then with 6GB you are still consuming a lot of power.

hmmmmm
http://www.pcper.com/article.php?aid...=expert&pid=10

37W higher power consumption of 1gb gddr3 vs 1gb gddr5?
that CANT be caused by the memory chips alone... impossible...
there are only 4 memory chips on those cards, it would mean almost 10W per memory chip... MORE than gddr3... which would mean above 10W per memory chip! they would burn up in no time without a heatsink on them... normal mem chips have 1-2W tdp...

and lets say the imc runs much hotter if its driving gddr5 and the memory chips ONLY consume 5W each, then the memory chips would still burn up without a heatsink and the imc had to run over 15W hotter just by driving gddr5 instead of gddr3? and thats only a 128bit imc? that CANT be right... that gddr5 card must have a terribly unefficient design and they somehow fcked it up or used super cheap very unefficient components in the pwm...

i couldnt find any useful information on wikipedia or samsung or hynix sites regarding gddr5 power consumption... anybody?

[trinibwoy]

well according to the numbers weve all seen, the 470 needs every little bit of additional perf it can get since it seems to be behind the 5870 in most scenarios...

You've seen numbers from launch clocks?

[onethreehill]

NVIDIA Explains Fermi GTX 470 and GTX 480 Delay
We redesigned GF100 from the ground up to deliver the best performance on DX11. This meant adding dedicated h/w engines in our GPU to accelerate key features like tessellation. We also made changes on the compute side that specifically benefit gamers like interactive ray-tracing and faster physics performance through things like support for concurrent kernels. Unfortunately all of these changes took longer than we originally anticipated and that’s why we are delayed. Do we wish we had GF100 today? Yes. However based on all the changes we made will GF100 be the best gaming GPU ever built. Absolutely.

http://news.softpedia.com/news/NVIDI...y-137157.shtml

[SKYMTL]

Zotac planning GeForce GTX 480 AMP! version
http://www.nordichardware.com/en/com...p-version.html

Just like EVGA will have Superclocked versions, ASUS will likely have a TOP edition, BFG will have an OC.....how is this news?

[Chickenfeed]

I don't buy it lol... If we knew when and why this so called " redesign " started, then perhaps it would be more believeable. Otherwise I call shananagins!

[-Sweeper_]

Just like EVGA will have Superclocked versions, ASUS will likely have a TOP edition, BFG will have an OC.....how is this news?

they're desperate for anything about fermi

[LordEC911]

High speed GDDR5 consumes power like a mofo - it's the reason the 4870 had such a horrendous idle power consumption. And also the reason the 5970 only has 1000 Mhz GDDR.

4870 had borked MCs that couldn't dynamically downclock the GDDR5.