Some hard Fermi performance # info

[XMAN245]

All indications are that the 20-30% average is for the 512SP part. It wouldn't make any sense for the 448SP part, as it would mean a 50% increase for the 512SP part. I very much doubt you will see the kind of improvement from Fermi to 5870 on average that the 5870 sees over the 4890 (+- 10%). In their benchmarks nVidia never specified that it was the GTX 360 that was giving that kind of improvement in Far Cry & Uniengine. Let's go ahead and assume that if they could talk those benches up more, they would have.

And it isn't very impressive considering that Fermi has 50% more transistors, although obviously if the price is correct it won't matter for anyone beside nVidia and the server-class market.

edit: It isn't right to take Razor1's comments on B3D and use them as fact, just like it isn't a great idea to take ChrisRay's comments on Fermi too seriously. The latter is the same person that said all of our questions would be answered come the Sunday reveal, and that Fermi is so great, that the people who doubt nVidia's focus on the consumer games market will be silenced. Then the reveal happens and he backpedals immediately. It's the same reason you shouldn't read Charlie's comments as absolute fact. The difference is guys like Chris Ray have some more credibility, but are actually significantly more biased towards one company. Not to pick on him, but he is close enough to nVidia that I wouldn't go by his notions. I'm not sure where Razor gets his information, but it doesn't seem to be right in this case.

Then you are in for one hell of a shock come march then as the part that has been shown to date is in fact the 448SP part. The 512SP parts are, it turns out, going to be coming from B1 steppings as the A3s haven't made them any more viable than A2s, EI 448SP working parts are what they are getting from TSMC in quanity.

 

[mafa]

If these benches are from the nvidia demo, it is pretty safe to assume that those are the best nvidia has now. That is quite sad.

 

[GoldenTiger]

If these benches are from the nvidia demo, it is pretty safe to assume that those are the best nvidia has now. That is quite sad.

Not if XMAN's post is accurate and they don't have enough 512SP parts that they're 100% certain they'll be launching with that at first... if that's the case, then they may have used the 448 as a demo, since they can't be sure with the yields that the 512 would be available commercially right away. Just speculation on my part... they may have the "B1"'s as he's saying coming out shortly but can't test much with them yet.

 

[XMAN245]

For those who dont know, a B1 stepping means a complete and utter total respin of the die package from the ground up. Not good news for the high end part, but maybe good news in the end if whatever is wrong gets fixed, improves yields and maybe even reduces heat/power useage.

 

[realitycheck]

Excellent points, and more or less my sentiments. Even if these performance results are accurate and indicative of overall performance (which is a long shot), 20% or even 30% isn't impressive - it's what I'd expect given the time frame. We also don't know much about power consumption, overclockability, or aesthetics (noise, etc.). Fermi so far isn't impressive except for the fact that market prices might be lowered due to their finally being some competition from NVIDIA. If AMD has a refresh ready to go, even better, something else to look at. Anyway, the completely lack of competition is disturbing, good word choice.

Another interesting note is that despite how people interpret them, these rumor sites post just that, rumors. I find it interesting how accurate many of these rumors were this time around.


No, you are exactly what realitiycheck was talking about, only on the NVIDIA side of things. Tom made some very thoughtful and logical statements. You're foaming at the mouth attacking anyone who doesn't love Fermi. Ironic, no?

No, you and Tom are what I was talking about.

You sit here and say 20 to 30% isn't that big of a deal? Please.

You sit there and say Fermi isn't that impressive. Well, to be completely candid, I sort of agree. I expect it to be 40 to 50% faster than the 5870.

But if you want to play that game you have to accept the fact that besides eyefinity, the 5870 isn't that impressive. New generation hardware that gets beat by hardware that is a generation old.....GTX 295 anybody?

You and Tom are making excuses for ATi (example: well....ATi has a refresh...so there!!!) and bashing Nvidia (example: 20 to 30% isn't that impressive), and then you tell another poster that he is "foaming at the mouth".

No one is foaming at the mouth here except people like you, who I WAS talking about. Instead of being happy for competition, you're making excuses for ATi and bashing Nvidia, and I gave examples to back up my statement.

Seriously, I'm not trying to call you out, but do you realize how ridiculous you sound when you say 20 to 30% increase isn't impressive?

So, according to you, if I could bench press 275 pounds, a guy who could bench 20 to 30% more weight isn't more impressive? According to you if I had a car that could do 10 seconds in the quarter mile, I shouldn't bother doing an engine upgrade that could give me times 20 to 30% better?

20 to 30% of an increase in monitor size isn't that impressive?
20 to 30% of an increase in a bank account isn't that impressive?
20 to 30% of an increase in penis size isn't that impressive?

c'mon dude, come off it. Man up and admit it's nice. Of course we all realize they had more time to do it in, and we all know ATi has a refresh....there is no need to try to force everyone to think Nvidia's card is no big deal.

You sound like the people that bash eyefinity because of bezels....ridiculous.

 

[akotlar]

Then you are in for one hell of a shock come march then as the part that has been shown to date is in fact the 448SP part. The 512SP parts are, it turns out, going to be coming from B1 steppings as the A3s haven't made them any more viable than A2s, EI 448SP working parts are what they are getting from TSMC in quanity.

Then the 448SP part will be the only production part. You'd be crazy to think that a stepping change will give them the ability to enable 2 extra blocks. If they can't make them now, even for demonstrations, they will not be ready for March. If you are right, and I'd like to make clear that XMAN is not going on any real information, then the 512 part will basically be the X800 XT PE, with such limited availability that it is essentially a PR part.

That being said, I doubt that the performance numbers were from a 448SP part. Especially since the geometry figures match up with their white paper figures, which all reference a 512SP "GF100". There is no distinction made in their white paper or in their benchmarks, so I think this is a very stupid assumption to make with such bravado. It is just wishful thinking.

Also a 448SP part would have 40ROPS, 56TMU's, and ~ 1.3-1.4 Gflops, I'd be very impressed if their optimizations bring them to 1.2-1.3x a 2.72 tflops, 80tmus part. That doesn't seem very likely.

 

[XMAN245]

Then the 448SP part will be the only production part. You'd be crazy to think that a stepping change will give them the ability to enable 2 extra blocks. If they can't make them now, even for demonstrations, they will not be ready for March. If you are right, and I'd like to make clear that XMAN is not going on any real information, then the 512 part will basically be the X800 XT PE, with such limited availability that it is essentially a PR part.

That being said, I doubt that the performance numbers were from a 448SP part. Especially since the geometry figures match up with their white paper figures, which all reference a 512SP "GF100". There is no distinction made in their white paper or in their benchmarks, so I think this is a very stupid assumption to make with such bravado. It is just wishful thinking.

Again sorry to disappoint, at launch, yes the 512 part will be basicly a PR part, but "IF" B1 fixes the issues, then 512 parts will start to land on store shelves by mid summer. It takes about 8 weeks min for a new stepping to be completed start to finish. Should it not, then the 512 part will be, as you say the X800XT PE(Press Edition). FYI, frm my understand, the problem stems from a miss layout and is contantly affecting the same block of SPs. Nvidia has no plans or desire to launch a 480SP part, so they disable 2 instead of just 1. It is that mislayout that they hope to correct with the B1 stepping.

And from what I've heard out of RTP, to date has been nothing more than the 448s, the number of 512SP parts they have gotten that work near target speeds you can count on 2 hands, the number of 448SP parts that work at targeted speeds has been about 1/2 the A2s to date and most of the A3s. Those that are not making targeted speeds for the Geforce line are being shifted the Tesla and Quadro lines.

 

[GoldenTiger]

Then the 448SP part will be the only production part. You'd be crazy to think that a stepping change will give them the ability to enable 2 extra blocks. If they can't make them now, even for demonstrations, they will not be ready for March. If you are right, and I'd like to make clear that XMAN is not going on any real information, then the 512 part will basically be the X800 XT PE, with such limited availability that it is essentially a PR part.

It sure can. By the way, I love the bravado which with you proclaim XMAN has no sources and/or is uneducated... A for gusto, F for logic! I'll be trying for a 512 myself but it does sound like it will be extremely hard to find initially.

 

[Mr. K6]

No, you and Tom are what I was talking about.

You sit here and say 20 to 30% isn't that big of a deal? Please.

You sit there and say Fermi isn't that impressive. Well, to be completely candid, I sort of agree. I expect it to be 40 to 50% faster than the 5870.

Then I gave you the benefit of the doubt and assumed you had some sense of logic when I shouldn't have; I don't know why I do that on the internet. Anyway, contradict yourself much?

But if you want to play that game you have to accept the fact that besides eyefinity, the 5870 isn't that impressive. New generation hardware that gets beat by hardware that is a generation old.....GTX 295 anybody?

The 5870 took the 4890 released only five months earlier, doubled it's architecture while keeping the same clockspeeds and REDUCING the power consumption of the part, and improved the cooling and aesthetics of the card, all the while adding new features (DX11) and technologies (Eyefinity, etc.). It beat out the highest performing single GPU of the time, the GTX285, by 40%, and traded blows with the fastest dual GPU solution (and I'm with GoldenTiger on this one, I don't even consider dual GPU solutions), but you weren't impressed? That brings fanboyism past retardation.

You and Tom are making excuses for ATi (example: well....ATi has a refresh...so there!!!) and bashing Nvidia (example: 20 to 30% isn't that impressive), and then you tell another poster that he is "foaming at the mouth".

No one is foaming at the mouth here except people like you, who I WAS talking about. Instead of being happy for competition, you're making excuses for ATi and bashing Nvidia, and I gave examples to back up my statement.

Where? Because we aren't praising Fermi? It's not that impressive, sorry to rain on your parade. Tom made some very logical statements and I agree. 20-30% performance boosts are what we generally see between generations. It's been six months, NVIDIA competely missed the last generation, so performance is where it should be. I'm not impressed by mediocrity no matter what color it wears.

Seriously, I'm not trying to call you out, but do you realize how ridiculous you sound when you say 20 to 30% increase isn't impressive?

So, according to you, if I could bench press 275 pounds, a guy who could bench 20 to 30% more weight isn't more impressive? According to you if I had a car that could do 10 seconds in the quarter mile, I shouldn't bother doing an engine upgrade that could give me times 20 to 30% better?

20 to 30% of an increase in monitor size isn't that impressive?
20 to 30% of an increase in a bank account isn't that impressive?
20 to 30% of an increase in penis size isn't that impressive?

Who's foaming at the mouth and making ridiculous analogies because he can't face reality? Again, we see 20-30% boosts every generation, that's not impressive. You aren't adding to your argument's credibility by talking about bench pressing (lulz) and penis size (double lulz) in a thread about video cards.

c'mon dude, come off it. Man up and admit it's nice. Of course we all realize they had more time to do it in, and we all know ATi has a refresh....there is no need to try to force everyone to think Nvidia's card is no big deal.

You sound like the people that bash eyefinity because of bezels....ridiculous.

Like I said, fanboys foaming at the mouth. The minute people show a situation for actuality, you start foaming at the mouth due to their blasphemy. But thanks for replying, it's easier to make my points when you make my points for me.

If these benches are from the nvidia demo, it is pretty safe to assume that those are the best nvidia has now. That is quite sad.

Yes and no. Like I said, it's just average. Sad considering the hype? I suppose. But like I also said, there's no where near enough information available.

We know as much about the card as we did the 5870 pre-launch, which I also was excited about because the "rumors" generally are real info.

The bolded part is simply not true, and your inquisition is irrelevant to performance which is what this thread is about.

P.S. More personal insults/strawmen? Seriously, calm down dude.

At least a week before the 5870 launched, and mostly two weeks prior, I knew it's price, architecture, clock speeds, design, biggest features, power consumption, and how it looked. I know none of those for the Fermi part, and, as is reasonable, we're still 6 weeks out from a launch. I wasn't nearly excited about the 5870 6 weeks out either. Like I said, people are stroking themselves with no information, and then get pissy when not everyone wants to get in on the circle jerk. Settle down and wait for some more numbers and information to come in. And there's no strawman in there, learn your fallacies.

I, and many others, couldn't care less about dual GPU regardless of the pricing .

I'm with you on that one; the 5970 isn't even on my radar.

 

[nomu]

But if you want to play that game you have to accept the fact that besides eyefinity, the 5870 isn't that impressive. New generation hardware that gets beat by hardware that is a generation old.....GTX 295 anybody?\

Well, not to play in old mud, but didn't the 9800GX2 beat the 4870 512?

 

[shansoft]

No, you and Tom are what I was talking about.

You sit here and say 20 to 30% isn't that big of a deal? Please.

You sit there and say Fermi isn't that impressive. Well, to be completely candid, I sort of agree. I expect it to be 40 to 50% faster than the 5870.

But if you want to play that game you have to accept the fact that besides eyefinity, the 5870 isn't that impressive. New generation hardware that gets beat by hardware that is a generation old.....GTX 295 anybody?

You and Tom are making excuses for ATi (example: well....ATi has a refresh...so there!!!) and bashing Nvidia (example: 20 to 30% isn't that impressive), and then you tell another poster that he is "foaming at the mouth".

No one is foaming at the mouth here except people like you, who I WAS talking about. Instead of being happy for competition, you're making excuses for ATi and bashing Nvidia, and I gave examples to back up my statement.

Seriously, I'm not trying to call you out, but do you realize how ridiculous you sound when you say 20 to 30% increase isn't impressive?

So, according to you, if I could bench press 275 pounds, a guy who could bench 20 to 30% more weight isn't more impressive? According to you if I had a car that could do 10 seconds in the quarter mile, I shouldn't bother doing an engine upgrade that could give me times 20 to 30% better?

20 to 30% of an increase in monitor size isn't that impressive?
20 to 30% of an increase in a bank account isn't that impressive?
20 to 30% of an increase in penis size isn't that impressive?

c'mon dude, come off it. Man up and admit it's nice. Of course we all realize they had more time to do it in, and we all know ATi has a refresh....there is no need to try to force everyone to think Nvidia's card is no big deal.

You sound like the people that bash eyefinity because of bezels....ridiculous.

it not impressive when it took nVidia half year to catch up and shows only 20-30% performance increase...

if its 50%+, then I will be impress and grab the card right away, by then, no its not impress in any aspect, maybe to some people...

and by the time when Fermi finally gets release, I believe ATI already have something ready to answer or even with HD 6000 series coming up.....

 

[akotlar]

It sure can. By the way, I love the bravado which with you proclaim XMAN has no sources and/or is uneducated... A for gusto, F for logic! I'll be trying for a 512 myself but it does sound like it will be extremely hard to find initially.

Sorry if you followed XMAN posts you'd know he has no solid info. There is speculation that the benchmarks released prior to Sunday were from a 448 SP part, but that is not solid info. And the white paper clearly refers to the full part. It would not make sense if the performance figures in the whitepaper referenced a 448SP part.

Again, if the Sunday performance figures they released, called "GF100" vs "5870" were referencing a 448SP part, then there is no 512SP part. Unless you can give me a good reason for hiding your performance figures when you are 6 months late?

Nice try Golden, but A for gusto, F for logic!

 

[akotlar]

Then you are in for one hell of a shock come march then as the part that has been shown to date is in fact the 448SP part. The 512SP parts are, it turns out, going to be coming from B1 steppings as the A3s haven't made them any more viable than A2s, EI 448SP working parts are what they are getting from TSMC in quanity.

If Fermi is, in its fastest config, faster than the 5870 by 50% on average (or more), 1920*1200 and up, I'll take a month off these boards and paypal you $5 conciliatory bucks. If you're wrong you'll put me in your sig with the following tagline: "akotlar is a god among men...I pray to a mini akotlar idol"

Deal?

 

[LeninGHOLA]

Until these cards are in the hands of reviewers, these cards do not perform at all. They don't exist.

I would be very happy if they did perform as speculated, but since none of us are running them or have seen real benches, who knows? No, the Hardware Canucks article cannot be considered a real benching by any stretch.

The leaked 'benches" even for the 5870 were overblown, why do we always accept these speculative BS leaks at face value?

 

[akotlar]

Until these cards are in the hands of reviewers, these cards do not perform at all. They don't exist.

I would be very happy if they did perform as speculated, but since none of us are running them or have seen real benches, who knows? No, the Hardware Canucks article cannot be considered a real benching by any stretch.

The leaked 'benches" even for the 5870 were overblown, why do we always accept these speculative BS leaks at face value?

I don't see why Hardware Canucks can't be used. Their GTX 280 numbers were within 2% of Nvidia's, so the comparison should be valid.

 

[GoldenTiger]

If Fermi is, in its fastest config, faster than the 5870 by 50% on average, 1920*1200 and up, I'll take a month off these boards and paypal you $5 conciliatory bucks. If you're wrong you'll put me in your sig with the following tagline: "akotlar is a god among men...I pray to a mini akotlar idol"

Deal?

Until these cards are in the hands of reviewers, these cards do not perform at all. They don't exist.

I would be very happy if they did perform as speculated, but since none of us are running them or have seen real benches, who knows? No, the Hardware Canucks article cannot be considered a real benching by any stretch.

The leaked 'benches" even for the 5870 were overblown, why do we always accept these speculative BS leaks at face value?

Sounds like you're both ostriches near sand .

 

[akotlar]

Sounds like you're both ostriches near sand .

Sounds like you can't see the forest from the trees.

 

[GoldenTiger]

Sounds like you can't see the forest from the trees.

Sounds like I'm going to need to agree to disagree... I think we've both talked about what we can, and will just have to see what happens at this point .

 

[LeninGHOLA]

I don't see why Hardware Canucks can't be used. Their GTX 280 numbers were within 2% of Nvidia's, so the comparison should be valid.

They can't be used because these tests were not run by Hardware Canucks themselves, Nvidia ran the Far Cry 2 bench and the Dark Void bench in front of them, in a closed box.

Not only that, they are using drivers we know nothing about, which could easily be tweaked to show large performance gains in those measly 2 benches that were run. The 285 was using release drivers, and the Fermi was using unknown drivers.

There wasn't a control in the tests, meaning we don't have exact specs for the test system.

Dark Void hasn't even shipped, and there are obviously no benches for the 5870 on that game.

Personally, I have a feeling Fermi will be a faster card than the 58xx series. I'm definitely not biased one way or the other, the 5850's are the only ATI cards I've had since the 9800 pro. I almost always buy Nvidia because it is quite common for them to be ahead.

...But, until these things are out, what exact reassurance do I have that any of this speculation is fact?

 

[LeninGHOLA]

Sounds like you're both ostriches near sand .

If there was something factual for me to see, I'd be looking.

What exactly would I lose if Fermi was as good or better than is being speculated? I hope Fermi is faster than what ATI has out, it would be a shitty year for gaming if these cards failed.

 

[kllrnohj]

Well, not to play in old mud, but didn't the 9800GX2 beat the 4870 512?

The 9800GX2 also beat the GTX 280 in benchmarks. There were tons of threads when the GTX 280 launched about 9800GX2 owners being disappointed and seeing no reason to upgrade.

 

[Matthew Kane]

I'm just in for whatever offers best bang for buck. If its green or red, my nvidia fanboy days are over.

I think its been mentioned by one of the guys here, but if that benchmark was a incomplete gf100, then it'd be pretty impressive when the full product comes out let alone, if that gf100 was a 448sp one and equiv to GTX350 or 60 then the later amped up versions (GTX360, 80 90 or even 95) would be really good performers, yet again, I'd be eagerly waiting to see what both the green and red team are going to release.

Also from the nvidia site, it mentioned Ray Tracing technology, I've seen some tech demos on that and its pretty impressive requiring alot of processing power though, it'd be a first to see it implemented into the card when released.

And also it'd be pretty cool, if nvidia releases the fermi cards within the range of $380 - 450 due to all this delay we not seeing them in the market since ATi's first DX11 release.

 

[Matthew Kane]

The 9800GX2 also beat the GTX 280 in benchmarks. There were tons of threads when the GTX 280 launched about 9800GX2 owners being disappointed and seeing no reason to upgrade.

Weren't those beta drivers being used with the GTX280 during those benches?

 

[heatlesssun]

And also it'd be pretty cool, if nvidia releases the fermi cards within the range of $380 - 450 due to all this delay we not seeing them in the market since ATi's first DX11 release.

nVidia will sell their cards at what the market will bear. If GF100 is about 25% overall than the 5870, just my hunch, could be wrong, then expect about a $520 GF100.

nVidia isn't stupid. They are well aware of their competitive disadvantage versus AMD, the press release on GF 100 tells you right there they know, at the moment and have to be sensitive to pricing, and the fact that AMD has the whole price spectrum covered with the 5000s. It makes sense that the GF100s would go around $520 simply because GF100 will be between the 5870 and 5970 in performance and likewise in price. The higher the performance, the higher the price.

But I GUARANTEE that GF100/1xx prices will be in line with what AMD has at the time. I wouldn't be surprised of AMD lowered its prices a bit before the GF100 launch just to keep the pressure up but nVidia isn't just going to pull a number out of a hat.

 

[akotlar]

You're missing the point. HardwareCanucks asked nVidia about the specs of the system, built one just like it, and using a GTX 285 (which was used in nVidia's testing) confirmed that performance was within 2% of nVidia's. Then they ran the same benchmark with a 5870 in that system. Of course it could be slightly different, but it is doubtful that the difference is much greater than 2%. No, it doesn't cut mustard for Kyle, but contrary to common Texan legend, Kyle is not God's communicator. It is as scientific as we need, since an error margin of 2% is just fine, and is a fairly normal margin of error for benchmarks. Btw, for all those complaining, Kyle & Brent's "scientificWOW(tm)" "real-world" testing has error margins far larger than that.

And there was a control. It was the 285. That was the calibration reference point. And the calibration was found to be +/- 2%. While you may be uncomfortable with the concept of matching systems, (and no indication of drivers used), the way HardwareCanucks did it was right. Again we're only talking about the fixed Far Cry bench, which showed a 2% variation for their (Canucks) GTX 285 results, versus the Nvidia reference. That's probably due to driver variation, but it doesn't really matter, since the latest 5870 drivers were used. The only variation from the 285 reference point was accountable by different drivers used, and that doesn't affect 5870 results. Kyle didn't like it because he didn't do it. He has a vested interest in knocking down his competition. It's a defensive move, don't take it as gospel.

When benchmarks are released, the performance delta between 5870 and Fermi will be no more than 2% off (probably less since most of the variation is probably down to drivers, not system config), unless some major driver changes occur (for both companies) by release time.

Dark Void wasn't tested on the 5870.

This is a reply to LeninGHOLA

 

[Matthew Kane]

nVidia will sell their cards at what the market will bear. If GF100 is about 25% overall than the 5870, just my hunch, could be wrong, then expect about a $520 GF100.

nVidia isn't stupid. They are well aware of their competitive disadvantage versus AMD, the press release on GF 100 tells you right there they know, at the moment and have to be sensitive to pricing, and the fact that AMD has the whole price spectrum covered with the 5000s. If makes sense that the GF100s would go around $520 simply because GF100 will be between the 5870 and 5980 in performance and likewise in price. The higher the performance, the higher the price.

But I GUARANTEE that GF100/1xx prices will be in line with what AMD has at the time. I wouldn't be surprised of AMD lowered its prices a bit before the GF100 launch just to keep the pressure up but nVidia isn't just going to pull a number out of a hat.

I know that, I was just imagining how cool it would be if they did that kind of pricing on their soon-to-be-released cards

 

[GoldenTiger]

You're missing the point. HardwareCanucks asked nVidia about the specs of the system, built one just like it, and using a GTX 285 (which was used in nVidia's testing) confirmed that performance was within 2% of nVidia's. Then they ran the same benchmark with a 5870 in that system. Of course it could be slightly different, but it is doubtful that the difference is much greater than 2%. No, it doesn't cut mustard for Kyle, but contrary to common Texan legend, Kyle is not God's communicator. It is as scientific as we need, since an error margin of 2% is just fine, and is a fairly normal margin of error for benchmarks. Btw, for all those complaining, Kyle & Brent's "scientificWOW(tm)" "real-world" testing has error margins far larger than that.

This is a reply to LeninGHOLA

Absolutely agree, and yes, the "real-world" testing is just a less precise way of doing things in a different manner, not "better" than benchmarks with minimum, average, and max FPS listed that you then interpret yourself from the numbers at different resolutions/etc. A computer-controlled run-through will always be more precise than a human running through "gameplay" in an attempt to duplicate a previous run-through while other systems unrelated to graphics such as AI do random things causing variation. I basically read HardOCP's reviews as "impressions" nowadays, and enjoy the forums. I never agreed with the switchover in methodology as it's got a lot more room for inaccuracy.

@ heatlesssun: Yep, they'd be foolish to sell it for less than they could get... they're in business to make money .

 

[akotlar]

Absolutely agree, and yes, the "real-world" testing is just a less precise way of doing things in a different manner, not "better" than benchmarks with minimum, average, and max FPS listed that you then interpret yourself from the numbers at different resolutions/etc. A computer-controlled run-through will always be more precise than a human running through "gameplay" in an attempt to duplicate a previous run-through while other systems unrelated to graphics such as AI do random things causing variation. I basically read HardOCP's reviews as "impressions" nowadays, and enjoy the forums.

@ heatlesssun: Yep, they'd be foolish to sell it for less than they could get... they're in business to make money .

Yeah I agree with what you're saying. I'm not saying that HardOCP's method doesn't have some added value over regular testing - it does - but we agree that those accepting HardOCP's methods should have no problem accepting the error margin on Canuck's Fermi vs 5870 benchmark.

I also go with Hard for a little extra info. I think they need to revamp their articles to better serve the more intelligent readers (too heavy handed and filled with unjustified claims to scientific method), but that's off topic here.

Of course Fermi will be target-priced just like every other product on the market today. It seems to be a good product, its just not the right product for the consumer space. I'm investing in Nvidia not because I think Fermi is a great consumer GPU, but because I think it will probably end up seeding a differentiated product line serving fairly new markets better than the competition, while maintaining a good hold of their consumer GPU space.

 

[heatlesssun]

I also go with Hard for a little extra info. I think they need to revamp their articles to better serve the more intelligent readers, but that's off topic here.

What exactly more could they provide, or that one for any matter. Pretty much 90% of what all the sites are reporting comes straight out of the documents nVidia has posted on its site.

This was just a glorified press release, nothing more. I do think that nVidia will be keeping people fairly well updated along this roughly two months to what everything is thinking is a late March hard launch.

I don't know what more could be added, even for the more "intelligent" reader.

 

[GoldenTiger]

What exactly more could they provide, or that one for any matter. Pretty much 90% of what all the sites are reporting comes straight out of the documents nVidia has posted on its site.

This was just a glorified press release, nothing more. I do think that nVidia will be keeping people fairly well updated along this roughly two months to what everything is thinking is a late March hard launch.

I don't know what more could be added, even for the more "intelligent" reader.

I think he means in general, not on this specific article. Though, even on this one, they could have easily added some explanation of what each architecture decision talked about means, possible ways it could be implemented and what the result might be, as well as their concerns in more detail. Kinda off topic though, really, to this thread.

 

[akotlar]

]I think he means in general, not on this specific article. [/B]Though, even on this one, they could have easily added some explanation of what each architecture decision talked about means, possible ways it could be implemented and what the result might be, as well as their concerns in more detail. Kinda off topic though, really, to this thread.

Bingo. With regards to this article, I thought it was good, and Kyle's technical articles usually are. He's a smart guy, as is Brent. My only dissapointment stemmed from how he hyped it a few hours before release. He said that it would be so good that it would "win awards" and was a dramatic departure for the consumer GPU scene (that it would dispel our doubts about nvidia in the consumer GPU scene). It did none of that. In fact, the only improvements I see are pretty much derivative of their GPGPU feature focus.

Namely the large caches, the presence of a read/write L1 that can be used for any data storage. Apparently people with GPU programming experience are saying that these large caches are really good for improving performance and decreasing local memory grabs, so that's not a major complaint. The DP focus is obviously of no benefit to us. And the focus on geometry processing seems really strange. It's obviously important, but I know from playing around with Crysis configs that running something like 5-6M triangles a frame (after culling), didn't impact framerate very much. Coupled with Baumann's claim that ATI's internal testing showed a 2% performance increasing from removing the setup bottleneck, and I'm left wondering if the geometry processing increase isn't a byproduct of some GPGPU related function. I believe it was Jawed who alluded to such an explanation, but I can't seem to find it.

 

[XMAN245]

Bingo. With regards to this article, I thought it was good, and Kyle's technical articles usually are. He's a smart guy, as is Brent. My only dissapointment stemmed from how he hyped it a few hours before release. He said that it would be so good that it would "win awards" and was a dramatic departure for the consumer GPU scene (that it would dispel our doubts about nvidia in the consumer GPU scene). It did none of that. In fact, the only improvements I see are pretty much derivative of their GPGPU feature focus.

Namely the large caches, the presence of a read/write L1 that can be used for any data storage. Apparently people with GPU programming experience are saying that these large caches are really good for improving performance and decreasing local memory grabs, so that's not a major complaint. The DP focus is obviously of no benefit to us. And the focus on geometry processing seems really strange. It's obviously important, but I know from playing around with Crysis configs that running something like 5-6M triangles a frame (after culling), didn't impact framerate very much. Coupled with Baumann's claim that ATI's internal testing showed a 2% performance increasing from removing the setup bottleneck, and I'm left wondering if the geometry processing increase isn't a byproduct of some GPGPU related function. I believe it was Jawed who alluded to such an explanation, but I can't seem to find it.

Thats 2% on ATIs hardware, Last I knew and checked, Dave B. worked for AMD/ATI not Nvidia. So how the hell can he comment on how it will affect Nvidias hardware? He can't, so use AMD/ATIs findings on what they have done with their hardware has no baring on how removing the triangle/geometry setup rate on Nvidias soon to be released hardware will affect it.

 

[pasta4u]

Reports are that nvidia will only release the lower end gts 360 boards in late march with the gtx 380 coming in June.

I don't see why ati will care about this. Even if the gts 360 is faster than the 5870 its over 6 months late and will cost much more than a 5870. Whats more while they both offer the ful lrange of dx 11 specs. The 5870 can do 3d on 60hz monitors and eyeinfinty. Things the gts 360 can not do without a 120 hz monitor and expensive shutter glasses and in the case of eyeninfinty they need two cards.

There will still be value in the 5870 esp with a price advantage it will enjoy. A refresh will enjoy the same things but be faster. Do you really believe ati has done nothing in the last 6 months since it released its cards ? THats silly . Current 5870s will go to a 1.2ghz core with a small increase in vcore and stay realtively cool. A better heatsink fan on them will allow current 5870s to easily stay at 1.2ghz. That is over a 300mhz increase in speed couple that with a respin to fix any bugs , faster memory and it can easily close any gap that exists between it and the gts360.

The gtx380 might be out of cypress's reach. But i doubt amd cares. It will be much much more expensive than the 5870 or gts 360 and as i said rumors point to June for that point. Rumors point to october for ati's new ground up gpu. Till them hemlock will just have to compete against the gtx 380

I don't see a problem for ati nor do I see a problem for the consumer who will happy swoop up price dropped 5870 or a gts360 or whatever part you have. The amount of people actually buying cards over $300 are much less than those buying cards under $300.

 

[Kowan]

What report shows the GTX380 launching in June?

 

[akotlar]

Thats 2% on ATIs hardware, Last I knew and checked, Dave B. worked for AMD/ATI not Nvidia. So how the hell can he comment on how it will affect Nvidias hardware? He can't, so use AMD/ATIs findings on what they have done with their hardware has no baring on how removing the triangle/geometry setup rate on Nvidias soon to be released hardware will affect it.

Why would it matter if it was on ATI hardware or Nvidia hardware? Both have a 1triangle/clock setup rate. The bottleneck would be smaller on a 48ROP Fermi 1triangle/clock setup product, although if I'm reading this correctly Fermi can only accept 32 inputs for its ROP's at a time, unless you are dealing with overlapping triangles, but maybe I'm confusing something. Anyways even if Fermi would otherwise have set up 32 triangles/clock that would make a 2% bottleneck a 2.6% bottleneck. Again, on my GTX 285 I was seeing very little performance drop from increasing to nearly 6M triangles/frame in Crysis. That could be down to something with Crytek's engine, but it is telling how little framerate is impacted by the increase in triangle count.

And yes Baumann works for ATI, but Nvidia works for itself. I fail to see how having two sides to the "triangle setup bottleneck" story hurts the discussion.

Also, it is very relevant to the discussion of purported benefits of Fermi's 4triangles/clock setup vs. the 5870. If we're looking at a 2% improvement on ATI products, that means that enormous OOO-like geometry assembly is really doing very little to outperform its competition, it means its only giving 2% more performance than it would have gained over the 5870 had it left the geometry engine alone (unless you seriously want to claim that GTX 285's bottleneck vs 4890 was the setup rate).

 

[SicKlown42012]

Current 5870s will go to a 1.2ghz core with a small increase in vcore and stay realtively cool. A better heatsink fan on them will allow current 5870s to easily stay at 1.2ghz. .

I'm not going to respond to the other part of your post because I do agree with it for the most part. But 5870s will not easily hit 1200 on the core. Pushing the voltage to 1.35(which is more than a slight boost), most cards will top out at anywhere from 1050-1075. The only way I cans see a 5870 hitting 1.2Ghz is if you're using sub-ambient cooling. Like the Phenom IIs, Cypress seems to respond better to lower temperatures than voltage.

 

[XMAN245]

Why would it matter if it was on ATI hardware or Nvidia hardware? Both have a 1triangle/clock setup rate. The bottleneck would be smaller on a 48ROP Fermi 1triangle/clock setup product, although if I'm reading this correctly Fermi can only accept 32 inputs for its ROP's at a time, unless you are dealing with overlapping triangles, but maybe I'm confusing something. Anyways even if Fermi would otherwise have set up 32 triangles/clock that would make a 2% bottleneck a 2.6% bottleneck. Again, on my GTX 285 I was seeing very little performance drop from increasing to nearly 6M triangles/frame in Crysis. That could be down to something with Crytek's engine, but it is telling how little framerate is impacted by the increase in triangle count.

And yes Baumann works for ATI, but Nvidia works for itself. I fail to see how having two sides to the "triangle setup bottleneck" story hurts the discussion.

Thats the thing, Nvidia has improved the tri per clock rate beyond 1/clock.Again, you can't use Dave B' for a basis for Nvidia hardware.

And if you look at triangle setup rates for R600 to R770 vs G80 thru GT200, yes, ATI has a big advantage in setup rates, but Nvidia had the fillrate advantage which made up the difference, but it still suffered. As it stands now, the 5870 can do 800M+ triangles a sec, fermi is estimated at 2800M triangles/sec, propably around 2400 for the 448 part. And now they have 48 ROPs to feed that monsterous setup rate. There will be some games where it will affect very little, but others where you think there was no setup rate problem, will show good gains from fermi.

 

[heatlesssun]

Reports are that nvidia will only release the lower end gts 360 boards in late march with the gtx 380 coming in June.

Links to info please.

 

[MrGuvernment]

didnt read it all, but even if it kicks ati's ass i am sure the price will kick anyone's ass who buys it...

 

[RamonGTP]

Links to info please.

I read something similar. I don't remember the actual release timeline but I do remember reading that the lower end model will be available first with the higher end shortly after but in very limited quantities. Sorry, don't have a link, there's been way too much Fermi info floating around for me to keep track of it all.