GF100 (Fermi) previews and discussion

Note that these a previews and not proper reviews. Most of the ones I've seen contain no real benchmarks except the one from Hardware Canucks which has a small handful. I believe NVIDIA ran the GF100 benchmarks though, not Hardware Canucks (well, at least for the 1920x1200 results). But it was just the standard built in benchmark, nothing special.

http://www.tomshardware.com/review [...] ,2536.html
http://bit.ly/7uDKh0 (Hardware Canucks)
http://www.pcper.com/article.php?aid=858
http://www.techpowerup.com/reviews [...] hitecture/
http://www.anandtech.com/video/showdoc.aspx?i=3721

Link fixed. For some reason TinyURL no longer works to get around the word filter here.

i was fairly certain theyd regain the performance crown, but what seems odd to me is how linked the gpu is to memory bit rate or whatever. x number of cuda cores equals x memory inerface value. seems a bit cumbersome to plan for.

Hmm...I see what nVidia did there, but I well underwhelmed. In the youtube videos we see a Fermi card (GTX360 or GTX380) in Ranch Small, not the typical level most review site use to benchmark cards. The Fermi card scored 84 FPS. Now lets see some benchmarks:

http://www.pcper.com/article.php?a [...] pert&pid=6

The 5870 scores about 83 FPS at the same res and quality. We can have 2 conclusions for this depending on which Fermi card was tested:

1. The GTX360 tested (448 Cuda processors) =>
- 5870 ~ GTX360
- 5870 < GTX380 (512 CP) < 5970

2. The GTX380 tested... => nVidia screwed.

I bet Nvidia will have a 5850 beater at launch, price/performance.

All it takes is a price drop from AMD and that win can become a loss.

I think another BIG part of the story is going to be that Fermi does DX 11 faster,
much faster. So then it will be which one is really more future proof.

I wonder what will nVidia do regarding the 5970, because from the look of things, they don't have any competitive card for Hemlock.

From the rumor-mill the power required for the Fermi chips will be high so I don't know how feasible a dual chip can be.

Word.

I believe then they will have other problems with the 5000 series refresh products. 6 months later...the 6800 series will come, and I simply can't see what will nVidia do to compete against them.

It is rumored that the new series from ATI is a whole new architecture. And your comment about the 400 series from nVidia...hmm...they are already late for the 5000 series, where do you think they will be for the 6000 series?

I doubt it
this is the only thing nvidia does, meaning no stone left unturned to make a profit. I think theyll do a dual core card right out of the gate.

We have absolutely nothing on the 6000 series other than it will be a new architecture, will probably be on 28nm at global found, and will probably not be here until late Q410 at the earliest.

As for this:

We only have info on the highest end cards. Nvidia will be top down like always. Probably we will see something faster than the 5870 first (360), then at the same time but in limited quantity we will see the 380 slipping in likely under the 5970. Around late spring they might have a working dual 360 card. We won't see 5850 level cards until the spring summer time frame, starting higher and moving lower.

Most of the major sites have stated that the tessellation implementation is too different to really compare using that benchmark. On Fermi it is tied to the shader cores, so it will hit performance more than on ATI. Though that may not matter if it is sufficiently faster.

As for the "only" 448.. didn't most people claim that the 512 will come around just not in great numbers right away? Well except for the trolls. I don't think anyone rational was claiming otherwise.

Also, until I am installing my fermi it is blowing nothing out of the water...

I'm excited about the possibility.. but we have nothing yet.. I can't say I'm too keen on slides from nvidia either.. as the ATI slides for the 5870 claimed some ridiculous things as well.

Implemention != performance, true. But as others have pointed out, performance is what people buy, and the first comparison across a single benchmark gives a 20fps advantage to Fermi.

The only question I have now is actual gaming performance on non-DX11 SW and Price.

There is an ATX specified 300W limit (which is why the 5970 uses downclocked but higher binned 5870 Chips).

AMD circumvented the 300W limit by making their 5970 highly overclockable in the hands of the end user but the limit remains.

We know that a Fermi based card with 448SPs (and a 600-750MHz core clock) has a power envelope of >225W. How much of a cut down would nVIDIA need to get two Fermi's on a single card? I would think too much of a cut down, so much so that a Dual GPU card is no longer feasible as it cannot compete with the 5970.

I think nVIDIA will wait for a die shrink before they release a Dual GPU product.

really , tell us how you truly feel.

I am almost certain they will release a beast of a dual GPU card if only to ensure that the fastest thing on the planet is Nvidia for some unknown amount of time.

While the radeons are certainly more powerful per area of die.. I'm not sure how it will stack up performance/power. We might see a dual 8pin 'special' edition 5980 that is clocked slightly higher with better cooling (to overclock the 590 well requires special VRM cooling, they are not dealt with well on the stock cooler) but that is as far as ATI would bother going with this gen. While I would not at all be surprised to see a GTX390 that was effectively two 448 parts together with no regard for staying in spec. I'm sure there is some stupid way around the spec they will exploit.. They would only be after claiming the single card crown, if one fermi does not do it they will figure out a way to smash exactly 300W of computing power together as maintaining the crown seems important to their business model.

It is still impressive to watch it chug it out..

Mind you, I'm more interested in how everything else light that benchmark is, as opposed to tessellation heavy. It doesn't look nice without tessellation giving me the impression that other than tessellation it is very easy on a GPU... I'd like to see how well fermi handles tessellation if it is also computing a texture fill heavy scene with AA. At what point does the dedicated hardware of ATI's tessellation engine overtake fermi, if ever, or is the computing power of fermi such that it won't make a difference.

Speaking of Eyefinity, Nvidia will get it in their new cards through 3D Vision Surround. Ive allways thought that Nvidia would have to come out with there own version sooner or later, and well here it is!. The only realy incredibly stupid thing I see with Nvidias implementation is that they force you to have TWO cards in SLI. Youd think that they would learn form ATIs mistake, being that the 3rd monitor had to be Displayport, and just put 3 outputs on the back of the card...

Wait..lol?

Can we see half a g100 running tesselation before you claim the 5770 and below does dx11 'badly'?

I really, really, really, really, really hope you are being sarcastic.

repeat of the RV600? vs G80, 800 bucks for just a single gpu, but in that situation nvidia wins by far in terms of performance.

The 3D nVidia Implementation is similar to ATIs 6 monitor. If its that much better, you shouldnt need SLI.
My take is this. Fermi is the first in a 3 gen (most likely) rehash down the road, and has the baby steps of whats going to happen down the road, much like the R600 was thus the 5xxx series, and next new gen, where we'll most likely see similar implementations, as early DX11 games wont be that hard, since most wont be ground up games using it, and top cards should handle it, its the second gen as usual thatll do better, as the DX model and games mature in its usage.
I would say AOHD is fairly demanding on DX11, and if used, its not nearly as scalable as tesselation is.
Theres other conversations going elsewheres saying some things, most arent forgetting where these "benches" are coming from.
The nVidia slide showing the 8xAA abilities at 35x16 is a great example of how these "benches" may be applied.
If the Fermi marks are indeed 2.33 times better than G200 at 25x16 using 8xAA, all this proves is what we do know.
We do know G200 8xAA sucks, and often fails at 25x16, so saying 2.33 is very subjective, and is what nVidia shows as triumph.

As for highend? I think we need to look at the partners here, as itll be them doing an end around on the PCI specs, and creating sone kind of frankensteinian "top" card, using more than 300 watts.
If this is so, not sure how nVidia will be able to claim victory here, as such a card will be outside their specs, even tho they can unofficially help in its desgn

A few more things we know
http://www.xtremesystems.org/forum [...] count=1484
Read the bottom of the post, and this is coming from Rys, whos had the best and closests understanding/ability of Fermi thus far.
I quote:
As for my clock estimates, I doubt 1700 MHz hot clock at launch (:sad, but the base clock should be usefully higher, up past 700 MHz. They still haven't talked about GeForce productisation or clocks, but at this point it looks unlikely the fastest launch GeForce will texture faster than a GTX 285.

Since its tied to the hot clock, going smaller will require large jumps in clocks to overcome in scaling, so, the lil Fermis may not show much promise at all, or worse.