The Fermi Thread - Part 3

[Sn0wm@n]

Will there be a 2+ GB version of the GTX 480?

ive heard of a 3GB model .... but no pricing yet on those

[Solus Corvus]

40% faster than 5870 is exactly where the 5970 is, lol

Maybe at release. But not after the crossfire improvements in drivers released since then. I have been using a 5970 since shortly after release and it has improved substantially. In games that scale properly across 2 gpus and aren't cpu bound a 5970 is about 45-65% faster then a 5870. The average was dragged down by many games without proper 5970 support in the driver, but many were added in more recent drivers. Crossfire/SLI scaling isn't great, but it isn't 40% with proper profiles either.


That's why it's important to pay attention to driver version used in reviews. Some of the new fermi reviews will doubtlessly use old data for the rest of the cards. That doesn't really give you a picture of the reality of the marketplace now.

[zerazax]

That's why it's important to pay attention to driver version used in reviews. Some of the new fermi reviews will doubtlessly use old data for the rest of the cards. That doesn't really give you a picture of the reality of the marketplace now.

Yep especially since the latest Cat drivers have been excellent. It will be interesting to see if the 2GB 5870s get a boost too

[Vardant]

Maybe at release. But not after the crossfire improvements in drivers released since then. I have been using a 5970 since shortly after release and it has improved substantially. In games that scale properly across 2 gpus and aren't cpu bound a 5970 is about 45-65% faster then a 5870. The average was dragged down by many games without proper 5970 support in the driver, but many were added in more recent drivers. Crossfire/SLI scaling isn't great, but it isn't 40% with proper profiles either.

Last time HD 4870 and HD 4870 X2 were reviewed using Cat. 9.5, the difference was 46% on average. That should be around a year after the initial release, so plenty of time.
I don't doubt, that HD 5970 is doing better now, but across many games in different resolutions and settings, the average won't be much better. Definitely not good enough to justify going dual GPU.

[mindfury]

[Marios]

I made this using what hardware-infos say's is the final specs and those prices are launch ones not retail.

Are there any real measurements to backup the announced 2.7 TFLOPs "Computational Power" of HD5870?

I believe that these "computational power" figures have nothing to do with reality.
Did you use the formula [Computational Power=Shader clock*Shader units*2] to calculate the single precision floating point operations per second (FLOPS)?

Nvidia announced double precision floating point peak performance for it's Fermi based Tesla C2050 / C2070 GPUs to be between 520 - 630 GFlops.
Nothing else.

[Solus Corvus]

Last time HD 4870 and HD 4870 X2 were reviewed using Cat. 9.5, the difference was 46% on average. That should be around a year after the initial release, so plenty of time.
I don't doubt, that HD 5970 is doing better now, but across many games in different resolutions and settings, the average won't be much better. Definitely not good enough to justify going dual GPU.

How about we wait for some reviews.

I don't really agree with your method to average a range of resolutions and settings. I don't buy an expensive graphics card to run a range of resolutions - I buy it to run high resolution and high settings. When I'm looking at averages I'm looking at a certain resolution and quality or higher because I'm not going to be running lower. For the same reason I look at which games scale with which architecture (ati or nvidia) and compare to the games I play. The resolutions I use and games I play may not be what someone else uses. And it's ok, we can all make a rational choice for our individual needs.

[mindfury]

Are there any real measurements to backup the announced 2.7 TFLOPs "Computational Power" of HD5870?

I believe that these "computational power" figures have nothing to do with reality.
Did you use the formula [Computational Power=Shader clock*Shader units*2] to calculate the single precision floating point operations per second (FLOPS)?

http://forum.beyond3d.com/showthread.php?t=55830
5770--1Tflops
5850--1.4Tflops
5870@1G--2.2Tflops

GT200 never achieved announced 1Tflops either.Those announced numbers are just peak performance.

[Marios]

The average comparison of GPUs is not so bad.
There are many ways to express this percentage.
For example we may play the 10 most demanding games with all eye candy enabled.
Then compare the percentage of games that are fluently playable at certain resolutions with each card.

[takamishanoku]

Cheers mindfury for those terrific pics. The heat sink looks quality (as does the card in general)!

[ElSel10]

http://forum.beyond3d.com/showthread.php?t=55830
5770--1Tflops
5850--1.4Tflops
5870@1G--2.2Tflops

GT200 never achieved announced 1Tflops either.Those announced numbers are just peak performance.

I thought the 5870 was supposed to be 2.7 TFlops at stock speed?

[eric66]

thats probably the biggest heatpipe i have ever seen on non custom model. card must be really hot at 700mhz

[mindfury]

I thought the 5870 was supposed to be 2.7 TFlops at stock speed?

2.7 TFlops is peak performance in theory,performance in real world is lower.

[GoldenTiger]

Maybe at release. But not after the crossfire improvements in drivers released since then. I have been using a 5970 since shortly after release and it has improved substantially. In games that scale properly across 2 gpus and aren't cpu bound a 5970 is about 45-65% faster then a 5870. The average was dragged down by many games without proper 5970 support in the driver, but many were added in more recent drivers. Crossfire/SLI scaling isn't great, but it isn't 40% with proper profiles either.


That's why it's important to pay attention to driver version used in reviews. Some of the new fermi reviews will doubtlessly use old data for the rest of the cards. That doesn't really give you a picture of the reality of the marketplace now.

It's 40% faster on average at 2560x1600 with some AA (2x-4x) in general still for a stock 5970 compared to a 5870, from what I've seen with newer drivers (I haven't seen a good article on 10.3a's yet). Regardless, 40% on a slower 5970's GPU's (core clocks/etc.) isn't 40% over a 5870, anyway: the 5870 has higher stock speeds and can clock higher when oc'ing as well. Crossfire tends to spike your max FPS making an awesome-looking average, while not really helping the minimums a whole lot for the most part. I don't need 20min, 75avg. 140 max, I need 50min, 80avg, 110 max!

Nice finds on those pics Mindfury, thanks!

Yep especially since the latest Cat drivers have been excellent. It will be interesting to see if the 2GB 5870s get a boost too

Excellent?

[pentium777]

Can anyone remember a time in Nvidia history when a big release on final info and benches has been this secretive? 5 days left now and it had better not be a paper launch or some of 100 cards available. They would embarass themselves linking it to PAX with all those gamers there and word of mouth.

Next, how close are we going to get to launch day when we see 100% specs with lots of benches? I'm guessing they have the NDA fear imbedded but people in countries like China etc why would they care? Not like Nvidia is going to sue them across the world.

Thanks to everyone contributing to the info in this thread. To whoever commented about Quadfire 5970s the performance is POOR, as is the Quad GTX 295, I'm thinking Tri-GTX 480s will be highest end shortly. Not for me though GTX 480 SLI is my limit only because of 2560x1600, hopefully will be happy with them as long as I have been with the former GTX 280s that lasted from Nov 08 till now.

[badboy18187]

@GoldenTiger: i don`t understand why are you so biased against Ati?have you even tried the new drivers to see how they stack up?but as I have read your past posts it seems your are an Nvidia fanboy.
Ontopic:I can sense the tension Nvidia has started until now with Fermi, although as I am not a fanboy of either the green team or the red team, i just hope both gtx470 and 480 are very good.Especially the 480 because in the pics that mindfury posted, those heatpipes and that serious cooling indicated that it might actually be a serious overclocker.I don`t know, in the end we`ll see on Friday and be thoroughly amazed or simply dissappointed.

[trinibwoy]

Pixel Fillrate : 33.6 GPixel/s (700*48)

Pixel Fillrate is only 700*32=22.4 GPix/s.

There are 48 ROPs but the chip can only rasterize 32 pixels per clock. Can't output more than you input. The extra ROPs help with AA and/or in cases where you don't get nice compression on the writes to the framebuffer.

[saaya]

thats probably the biggest heatpipe i have ever seen on non custom model. card must be really hot at 700mhz

yeah, and i count 5 of them too...

the card does look nice though... stylish design... way better than atis cheap cr4ppy red n black plastic batmobile nonsense...

[Vardant]

Or maybe this cooler was designed for something hotter and they just used it for this one, who knows

[Sam_oslo]

You kids have been behaving while I've been busy with OCIng my new CPU?

That cooler looks really good. There is not such a thing as hot component, there is only bad cooling salutations.

[ajaidev]

yeah, and i count 5 of them too...

the card does look nice though... stylish design... way better than atis cheap cr4ppy red n black plastic batmobile nonsense...

Much stylish but i would have liked a back plate also.

Or maybe this cooler was designed for something hotter and they just used it for this one, who knows

Unlikely even if its 250 W tdp its a huge number the 2900 XT tdp is 215W

[saaya]

You kids

look whos talking

so let me get this straight, the card doesnt have a beefy heatsink cause it needs it, it doesnt... not at all... it probably runs super cool, nvidia only put an expensive beefy heatsink on there because... well...?

[Vardant]

Because they designed it for a card with higher TDP and changed it at the last minute, so they didn't have a replacement?

Unlikely, but still possible.

[Macadamia]

Because they designed it for a card with higher TDP and changed it at the last minute, so they didn't have a replacement?

Unlikely, but still possible.

Err they changed the GTX 480's cooling 2 months before the launch. Nearly painfully obvious that they wanted to push for more ambitious/less embarrassing performance figures and this had to be the way to go.

250W TDP on a single board with a single big hotspot isn't easy to deal with, in fact as heat and power go up more it seems to get harder to deal with them. And really, nVidia of all companies wouldn't hesitate a minute to cut down GTX480 cooling if it was "cool-running" (look at how they even butchered the GTX470 power segments)

The 5870 Eyefinity6 has the same core just more RAM, and in a sense is much easier to cool, so you don't see super buffed cooling despite the extra TDP.

[Sam_oslo]

look whos talking

so let me get this straight, the card doesnt have a beefy heatsink cause it needs it, it doesnt... not at all... it probably runs super cool, nvidia only put an expensive beefy heatsink on there because... well...?

LOL, you knew who I was talking to.

Because all high end GPU should come with a high end cooler. Those old crappy coolers have been limiting the OC headroom. These cards are for enthusiasts and they want to OC, so a good cooler makes scenes.

Maybe it has other reasons too, but everybody should be happy with a good cooler, not matter why it is there.