The Fermi Thread - Part 3

[Chumbucket843]

No. Just agreeing with the other people questioning HWC's numbers. Their review favors fermi more than the other ten review sites. The reason is, as I said in the post above yours, because he didn't bench any games where fermi performed close or less than a 5870. He also benched the game where it excells twice.

not all sites do their reviews the same way. some leave AA and AF maxed at every res and game, some pick newer games or dx11 etc. there are some quirks on pretty much every review. i remember seeing WoW in one review which is completely useless.

[GoldenTiger]

Sorry, but I found their methodology to be the best yet. Up-to-date drivers, 4GHz i7 920, minimum framerate inclusion.

Are you accusing HardwareCanucks of falsifying data?

QFT, definitely the best-done review of the bunch.

AMD took 7 months to prepare a driver that finally started using the power of an architecture that is 3 years old.

Fermi has been running 4 months on a new architecture. Clearly if we compare the performance to fermi to AMD 5xxx when the cards were initially released, we would be seeing different conclusions, albeit, fermi still loses badly for performance per watt. The inconsistent performance in nvidia favoured games such as crysis and opengl games show the immaturity of the drivers.

In addition, fermi minimum frame rates are 19% better than 5xxx. Something that is substantial.

Well stated... drivers are clearly quite immature as yet.

No. Just agreeing with the other people questioning HWC's numbers. Their review favors fermi more than the other ten review sites. The reason is, as I said in the post above yours, because he didn't bench any games where fermi performed close or less than a 5870. He also benched the game where it excells twice.

Just because the data doesn't line where you expected, doesn't mean it's invalid . It's following the rough pattern of other reviews, just better numbers...

[570091D]

No. Just agreeing with the other people questioning HWC's numbers. Their review favors fermi more than the other ten review sites. The reason is, as I said in the post above yours, because he didn't bench any games where fermi performed close or less than a 5870. He also benched the game where it excells twice.

it seems to me that gtx4x0 cards get more performance out of overclocked processors than hd58x0 cards. also, i didn't see any titles on the list that were not modern or popular, he left out batman aa (heavily nvidia) and included L4D (heavily amd). skymtl also used more games than most reviewers, and i've also noticed that the more games the reviewer uses the better nvidia's average lead becomes. ultimately there is probably not one reviewer that replicates your gaming set up exactly so one needs to read many reviews using different games and hw combos to learn the advantages and disadvantages of each card.

[insurgent]

I think I get it now, if you want to feel good about Nvidia you go to HWC or Bjorn3D, if you want to feel good about ATI you go to Bit-Tech or HardOCP, right? For malware and everything else it's Anandtech

[zerazax]

BSN's review included a graph of password cracking ability:

[flippin_waffles]

Yeah I saw that. Brutal if you're nv. But I think it's maybe not just that 4xx are nowhere near what expectations were, but that ATi delivered an insurmountable blow with Evergreen. I wonder how much of an infuence AMD's prowess with manufacturing was. No doubt that played a big part in ATi being able to leapfrog nv in such a dramatic way (not taking anything away from ATi engineers, just that AMD's manufacturing experience must have played a role).

[annihilat0r]

I think I get it now, if you want to feel good about Nvidia you go to HWC or Bjorn3D, if you want to feel good about ATI you go to Bit-Tech or HardOCP, right? For malware and everything else it's Anandtech

That's just mean for dear Anand

I thought the best review was Anand's.

[leszy]

We expected a pretty good memory OC and not stunning for core. Unfortunately, it proved that the memory would not accelerate too much and did not want to work with a frequency higher than 1035 MHz - is still well below their specifications. As mentioned, the clock is connected to the core clock shader units, so crank up the card manipulating a slider. Core, we managed to turn up to 800 MHz. Above this value began to give a counter-security, which cuts off the power when it detects that the power flows from the system too much power. This is the first card that we tested, which when overclocking with air cooling and no voltage change is a problem. Radeon HD 5870 In response to the OCP (overcurrent Protection - Counter-security) is needed if the GPU voltage around 1.45 V and the core clock ~ 1300 MHz

http://translate.googleusercontent.c...lmVWflHq0owatA

[ElSel10]

BSN's review included a graph of password cracking ability:

It seems this the only case where AMD's theoretical 2.7TFlop comes to fruition. Unless, is that password cracking benchmark single precision or double precision?

[insurgent]

That's just mean for dear Anand

I thought the best review was Anand's.

I was just keeping it light, my favorite tech site is Anandtech

[SKYMTL]

You know that the HardwareCanucks site review claims gtx480 is ~23% faster than hd5870 at 1920x1200 4xAA (in the actual benchmarks), while Anand, TPU, Guru3d, PCGH, Computerbase, Hexus and many others claim ~15%. Anand and Hexus say ~11%. Most of HWC benchmarks are going to show fermi in the best light possible.

Where the heck do you see that?!?! Seems you have to stop reading half-completed comparison charts.

[***Deimos***]

Quote:
Originally Posted by ***Deimos***
Assuming by April 12th, nVidia comes through with 10K+ availability, GTX480 is HUGE WIN for nVidia.
In some benchmarks its 30, 40 even 60% faster than 5870. I never expected 700Mhz, let alone stock HS OC to 820Mhz... WOW.
And do a double take on those DX11 and tesselation benchmarks. Certainly nobody would prefer a GTX295.
Unlike R600, FX or whatever you want to compare it to, IT IS the PERFORMANCE KING, the best GPU in history.

For the launch MSRP, nobody would deny its a WIN if it was <200W. Nobody.

But, thats the problem. Even the lowly GTX470 is a bit faster than 5870 half the time and between it and the 5850 the other half. But, its almost 100W more than 5850 and nearly 5970 numbers. Water cooling may lower temps, but its still the same ~300W going into your house.

Although 5870 is $420, and GTX470 is $350, it is VERY VERY EASY for AMD to lower price $50. On the other hand, much more difficult for nVidia to lower power by 50W (25%).
+$50 or +$50Watts, which would you choose?

Winners and losers:
Fastest GPU - GTX480
Fastest card - 5970
Best performance for 200W - 5870
Best performance for 300W - 5970
Best performance for $350 - GTX470*
*until inevidable price cuts.

My point is that nVidia's only advantage is PRICE. And their costs are much higher. So AMD can easily win here.

What reviews did you read and what are you on? I'd like some

Can't argue your points on price though, though I disagree with your "best performance for $350". I'd replace that with the 5850 even if it's priced $50-100 lower than that. The 5850 OC's to 1GHZ core easily. That's a massive ~40% OC. I'm not sure the 470 can hit speeds similar to match 1ghz in benchmarks. The 5850 can do it while running quieter, cooler, and less power too.

If you start comparing products including their overclockability you end up with stupid things like Core i7-920 better than Core i7-965 (just as an example).

If you could "group" them, would you put it:
A: G2xx, 4870/90,
B: 5850
C: 5870, GTX470*
D: GTX480, GTX295
E: 5970
F: GTX480 SLI (keep flamable objects and cats far away!)

*I placed GTX470 so high because it get lower (5850) performance in older less relevant games.

====================

Yes totally agree that preliminary overclocking tests show Fermi benefits more than 58xx.

====================

Some people MUST HAVE the BEST PERFORMANCE possible irregardless of money or whatever. GTX480 SLI is for them.

The early adopters or nVidia fans will naturally pick up some shiny GTX. Question is which one - budget decision probably.

Others have 50" HDTV which *only* does 1920x1280. They want something that plays games A, B and C smoothly.. for them a 5870 or even 5850 will more than suffice.

Those looking for best "value", should probably wait 1-2 months for price gouging and MSRP to fall.

[Chumbucket843]

password cracking is just one very simple gpgpu app. it doesnt determine gpgpu performance. here is som gpgpu benches straight from anand. i guess recursion helps.

It seems this the only case where AMD's theoretical 2.7TFlop comes to fruition. Unless, is that password cracking benchmark single precision or double precision?

the max performance achievable is 2.2tflops because that is peak register file bandwidth. it because evergreen can do bitops on 192bit integers which is amazingly fast on password cracking.

[mAJORD]

Where the heck do you see that?!?! Seems you have to stop reading half-completed comparison charts.

Sent you a PM

Hopefully we can work out why there's a difference..

As I mentioned to some, if anything I should remove METRO 2033 results from the average, because technically it's not AA/AF only AF. Taking it out results in 21.5% , which is still more than your 18%

I certainly haven't missed any games that I can see in the review. and everything (bar metro) i strickly 1920x1200 4xAA/16xAF results.

Leaving out the syntheric UNigne demo only makes my figures better for Ati since Fermi chews up that bench.

[SKYMTL]

Sent you a PM

Hopefully we can work out why there's a difference..

As I mentioned to some, if anything I should remove METRO 2033 results from the average, because technically it's not AA/AF only AF. Taking it out results in 21.5% , which is still more than your 18%

I certainly haven't missed any games that I can see in the review. and everything (bar metro) i strickly 1920x1200 4xAA/16xAF results.

Leaving out the syntheric UNigne demo only makes my figures better for Ati since Fermi chews up that bench.

We'll get it sorted. I sent you the source file for all our charts. Should help A LOT.

[jaredpace]

Where the heck do you see that?!?! Seems you have to stop reading half-completed comparison charts.

majorD's chart http://img39.imageshack.us/img39/211...v2smallpng.jpg

heres the numbers on your site & in majord's chart:
game -480 -5870 -xx%
avp 56.21 44.74 25.64%
bc2 58.41 50.42 15.84%
dt2 87.41 71.30 22.59%
dao 103.17 98.11 5.15%
fc2 79.78 65.74 21.35%
fc2 101.01 74.12 36.27%
ld2 130.31 105.11 23.97
total: 21.55%

metro (0xAA) & heaven (synthetic) which only increase the percentage...
met 31.84 25.02 27.25%
hev 48.10 39.50 21.77%
total: 24.51%

SKYMTL: You've made an error in your summary chart (it doesn't match your benchmark graphs), or either mAJORD & myself have made the same error getting numbers off your site.

[LordEC911]

AMD took 7 months to prepare a driver that finally started using the power of an architecture that is 3 years old.

Ummm... architecture is ~6months old...
Just because it isn't a complete architectural overhaul doesn't mean a lot of things didn't change under the hood that takes time to optimize. Look at all the new code that is being used thanks to a new API.

To think that AMD/ATi could have been optimizing and coding their drivers better for a future (unreleased) product is completely inane.

I don't think its Nvidia's fault though. All of GTX480's issues arise from poor yields. Now, if they had done their homework like AMD they might not have these issues, but the point stands Nvidia really as little control over the yield - thats all tsmcs fault. In addition to this, there wouldn't be as many cut down parts. If TSMC had gotten better yields you would have full core chip and higher clocks and likely lower power consumption.

Yep, because TSMC problems have absolutely no work arounds and no one else is able to produce a chip on their 40nm PP... Nvidia works with TSMC's engineers just as much as AMD does.
The fact is that AMD was able to find the problems and iron them out early in the process' life because they spent the time and money to do so.
It is TSMC's fault that Nvidia has broken MCs? That is news to me.

Also, you might want to wait for Bx silicon before saying that everything is TSMC's fault.

[randomizer]

Well I'm certainly replacing my grill. It's got George's approval!

[hurleybird]

epic pic

[Blacky]

Fermi reminds me of something, the reborn of FX 5800

[LiquidReactor]

Any ideas on how much drivers might improve Fermi arch performance?

[OutSider]

Time will tell.

[Blacky]

Any ideas on how much drivers might improve Fermi arch performance?

they had 6 months to improve the drivers, I don't expect miracles, though SLI is were Fermi really shines but thats about it.

[bill_d]

Any ideas on how much drivers might improve Fermi arch performance?

with the reviews over it may get slower to drop power and temp

[annihilat0r]

more than 5870 will go up from now on, prolly