The GT300/Fermi Thread - Part 2!

[RaZz!]

I[...]
I think the bad performance right now is bad drivers more than anything ...
[...]

in my opinion we shouldn't put much weight on current performance figures anyways. there are so many factors that could lead to bad performance, we don't even know the test setups etc...

we've been waiting so long for fermi to surface so that, now, it doesn't matter if we wait yet another week (or two, or three... )

[E30M3]

This. The ROP / Cache / Memory controllers scale together but separately from the CUDA cores.



However, at this point I won't be commenting in this thread other than to interject factual information simply because we keep going around the same circle:

Step 1: "I hear" or "rumor has it" or "supposedly" bandied around like people have a clue what is going on. They don't.

Step 2: Laughter by people who bought ATI cards and are desperately hoping they made the right decision and that prices won't go south if there is a strong GF100 launch

Step 3: Said ATI buyers or posters who want to start some controversy invariably quoting Charlie's work

Step 4: Someone actually posting some A+ info that has a firm foundation in reality.

Step 5: Repeat Step 2


Rinse and repeat.

+1

[MrMojoZ]

I find these early benchmarks to be disturbing, they aren't going to generate the price wars I was hoping to see at this rate.

[Andi64]

ROPs are related with the MC, and not with the SPs. In both arch. (ATi and nVidia) if you kill a set of ROPs you lose some of the Memory Controllers.

With HD5830, ATi killed half of the ROPs inside an entire ROPs cluster, so they can still use the MC related to that cluster.

nVidia and ATi can kill SPs without losing ROPs and MC.

Correct me if I'm wrong.

[saaya]

skyhtml, i dont think there are people here who hope that prices do NOT go down... i think everybody is hoping for lower prices, even people who bought a 5870 not too long ago...

in my opinion we shouldn't put much weight on current performance figures anyways. there are so many factors that could lead to bad performance, we don't even know the test setups etc...

we've been waiting so long for fermi to surface so that, now, it doesn't matter if we wait yet another week (or two, or three... )

yeah but all the benchmark numbers are in the same ballpark, and im sure they are on different setups, but they are all about the same... the picture i got is that except for crysis the 470 will realistically let you play games at the same res with the same settings the 5870 does, but it will cost more, run hotter and be in worse supply.

for enthusiast it should be interesting depending on how hot it runs and how high it clocks on ln2... then again 5870s clock to 1200+ on ln2 which is almost 50%... i think the only interesting card will be the 480, but the lack of benchmarks and even pictures of those cards is worrisome... there was only one pic of a 480 and that was a2... everything else as 470 and a3 right?

With HD5830, ATi killed half of the ROPs inside an entire ROPs cluster, so they can still use the MC related to that cluster.

oh interesting...
can nvidia do that as well?
man its a shame that there are no unlocked gpus like that huh?
id love to play around with different clock dividers on the gpu domains and enable and disable some blocks here and there and see how it scales...

[trinibwoy]

I think most of us expect Fermi to be a tesselation monster, but what makes anyone think that Fermi will also have an advantage in computer shaders? Even though Fermi should be able to reach a higher percentage of it's theoretical compute speed, it seems unlikely that Fermi will be able to make up for it's massive 1+ TFLOPs performance deficit that way. in doublr percision it will be even worse, where HD 5870 will be several times faster thanks to Nvidia crippling the DP performance of their Geforce line.

Read any paper on modern GPU architectures and you will see that raw flops is hardly ever the focal point. Getting data to and from the flops is as big or an even bigger issue. Stuff like the performance of global memory atomics is also a big deal that has nothing to do with flops. flop numbers are nice and easy for people to understand and while they're important there are a lot of other things that go into making an architecture fast at general computing. This is one reason the HD4870 never did that well in computing even with all those flops. It's shared memory architecture was still a bit behind Nvidia's.

@saaya, yes it takes time for software to catch up. This is exactly why Nvidia and AMD are sticking their noses in everyone's business and getting things added to PC games that take better advantage of the hardware.

[NaMcO]

Yeah that cooler does look superb and superior to the displayed GTX 470 version.

[LesGrossman]

However, at this point I won't be commenting in this thread other than to interject factual information simply because we keep going around the same circle:

Step 1: "I hear" or "rumor has it" or "supposedly" bandied around like people have a clue what is going on. They don't.

Step 2: Laughter by people who bought ATI cards and are desperately hoping they made the right decision and that prices won't go south if there is a strong GF100 launch

Step 3: Said ATI buyers or posters who want to start some controversy invariably quoting Charlie's work

Step 4: Someone actually posting some A+ info that has a firm foundation in reality.

Step 5: Repeat Step 2


Rinse and repeat.

I guess there aren't nvidia trolls in this thread, or maybe you're being incredible bias once again. I wonder how can you assess a product when you can't even assess a thread.

[Olivon]

I guess there aren't nvidia trolls in this thread, or maybe you're being incredible bias once again. I wonder how can you assess a product when you can't even assess a thread.

Please, it's possible to stop this kind of things ... ?

XS deserves way better ...

[Andrea deluxe]

ROP's amount is not related to SP's amount.

5850 only has 1440SP,but full 32 rops.

GTX260-216 increased 24SP,but stayed with the same ROP's amount as GTX260-192.

is true.....

i have mistaken....

[WeeMaan]

oh interesting...
can nvidia do that as well?
man its a shame that there are no unlocked gpus like that huh?
id love to play around with different clock dividers on the gpu domains and enable and disable some blocks here and there and see how it scales...

That would be totally awesome, 480 / 5x70 extreme edition.

[zalbard]

They are not in direct contact with the GPU though, and that might make a difference...

Yeah, DHT's proven to be worse so far.
Maybe video cards will change that.

[mao5]

calm down mate you're forgetting this is the GTX470, it was never meant to wipe the floor with all AMD's offerings. There is still the 480 which should offer 15% (at least) more performance than the 5870. If priced right, I'm sure you won't be complaining about buying one

actually it's 480, you can type 470 in your photoshop, then blur it, to see if it get matched with the 470's blur image in the warhead bench picture.

[johnieboy]

12V@1,8A? with a 70mm fan???

this means the fan alone will pull 21,6W @100% and will be very loud.
And if Nvidia use something like this at the "small" <225W GTX470 part i wonder what they will use at the GTX480

am i the only one having a fx 5800 deja vue?

[NaMcO]

am i the only one having a fx 5800 deja vue?

I smell NV30 all the way since a few days ago. I don't think they're gonna make it if the 480 cuda cores thing is true

To me it's not all about performance, but to most people is and those buy video cards.

[SKYMTL]

Once again, do research before assuming.

The BFB0712H series is rated for 2.64W @ 2900RPM and a noise output of 38 dB-A.

You can't safely push more than 10W through a standard GPU fan header. Not to mention that this is only a 75mm fan which means 21W could translate into 23,000+ RPMs. I don't know of a single commercially available bearing design that can put up with that for an extended amount of time

I don't know where that 1.8A comes from but it is either a misprint or some random number thrown in.

[johnieboy]

I don't know where that 1.8A comes from but it is either a misprint or some random number thrown in.

The only useful datasheet i found is on this website:

http://directories.csa-international...sl/certrec.xsl



I dont know how important the last 4 letters are, but it seems to be the same model.

The 1,8 should have been 1800 i guess, cause 1,8mA is clearly not enough to cool down a high-end card like the GTX470

[Solus Corvus]

Once again, do research before assuming.

The BFB0712H series is rated for 2.64W @ 2900RPM and a noise output of 38 dB-A.

You can't safely push more than 10W through a standard GPU fan header. Not to mention that this is only a 75mm fan which means 21W could translate into 23,000+ RPMs. I don't know of a single commercially available bearing design that can put up with that for an extended amount of time

I don't know where that 1.8A comes from but it is either a misprint or some random number thrown in.

Why are you assuming that BFB0712H is the same as BFB0712HF? Not all the fans in a series are going to have the same specs.

[kemo]

I dont know how important the last 4 letters are, but it seems to be the same model.

The 1,8 should have been 1800 i guess, cause 1,8mA is clearly not enough to cool down a high-end card like the GTX470

That is the one SKYMTL talking about ...


Edit : the BFB0712H is rated at 360mA in the link your posted , anyway the HF usually have 20-50 more mA than the regular H

[Solus Corvus]

http://i45.tinypic.com/333yfsx.jpg

That is the one SKYMTL talking about ...

That's not BFB0712HF...

Edit: Seriously, do you really think 11CFM would cut it?

[SKYMTL]

Why are you assuming that BFB0712H is the same as BFB0712HF? Not all the fans in a series are going to have the same specs.

Check out other Delta fans. 90% of the ones I have seen have an "F" at the end. It's just part of their usual product listing it seems.

If you guys think a 75mm fan will consume 21W when installed onto a GPU, you are sadly mistaken. C'mon, this isn't rocket science but rather pretty straightforward.


You also have to realize that since the fan is within an enclosed, directed environment, it doesn't need high CFMs to move a sufficient amount of air.

[***Deimos***]

Read any paper on modern GPU architectures and you will see that raw flops is hardly ever the focal point. Getting data to and from the flops is as big or an even bigger issue. Stuff like the performance of global memory atomics is also a big deal that has nothing to do with flops. flop numbers are nice and easy for people to understand and while they're important there are a lot of other things that go into making an architecture fast at general computing. This is one reason the HD4870 never did that well in computing even with all those flops. It's shared memory architecture was still a bit behind Nvidia's.

@saaya, yes it takes time for software to catch up. This is exactly why Nvidia and AMD are sticking their noses in everyone's business and getting things added to PC games that take better advantage of the hardware.

(made up guesses)
90% vs 96%. Difference cant be much, right? Afterall its just 6%, isn't it?
L2 cache hit ratio. 2MB vs 6MB. Nobody would buy an E5400 (2MB wolfdale) if it was only 6% cheaper than some 6MB model like E8200.

My 1st point - what seems like small architecture changes can have huge effects.

My 2nd point.
Compare application performance simply recompiled with newer compiler.. sometimes HUGE difference. Recompiled with some version of SSE - gains of hundreds of percent not uncommon. And yet, these are just for single threaded performance.

In huge wide parallel designs like Fermi (512) and 5870 (320 x 5), drivers and resource management are very very important. A couple lines of (equivalent) shader code replaced could be difference between 0.1fps and 100fs (ie 1000x !!)

That's insane! Impossible right? Well if you're into programming imagine this nonsense example:

float getResult(float a){
return a / 2.0;
}

float n = 0;
for (i=0; i<1000; i++){
n = n + getResult(i);
}

On typical Athlon/Pentium, if compiler doesnt inline getResult, thats 16-20 cycles (and bad cache coherency) x 1000.

Unnecessary use of floats (expensive). Float divide by 2, also dozens of cycles (as opposed to perhaps 1 cycle shift right)

And ofcourse, there is a simple formula for summing integers which would be more than 1000x faster

The problem is 99% of game developers are being wipped to quickly get product with shiny effects into market. How often do you hear - "we only had 1 month to add DX10 support".

Conclusion:
5 years of R&D, 3B trans, huge heat and power << efficient optimized algorithm.

I smell NV30 all the way since a few days ago. I don't think they're gonna make it if the 480 cuda cores thing is true

To me it's not all about performance, but to most people is and those buy video cards.

So you're saying it will be much much slower than ATI in DX11 (despite DX11 Heaven benchmark), questionable partial precision optimizations due to resource contention, it will be crippled by 128bit memory bus, and manufactured on old 130nm process.. (I guess 65nm is today's equivalent), and be big and hot. 4 "no"s and 1 yes, still means no.

[Solus Corvus]

Check out other Delta fans. 90% of the ones I have seen have an "F" at the end. It's just part of their usual product listing it seems.

If you guys think a 75mm fan will consume 21W when installed onto a GPU, you are sadly mistaken. C'mon, this isn't rocket science but rather pretty straightforward.


You also have to realize that since the fan is within an enclosed, directed environment, it doesn't need high CFMs to move a sufficient amount of air.

I realize that with squirrel cage coolers the goal is pressure, not max CFM. But IMO, 11.3CFM isn't going to be sufficient to cool a GPU.

Are you seriously arguing that ATI needs a ~10W cooler to cool a 5870 but that Nvidia found a way to get by with ~3W on a hotter GTX480?

[johnieboy]

[IMG]

Edit : the BFB0712H is rated at 360mA in the link your posted , anyway the HF usually have 20-50 more mA than the regular H

Well, then either you are wrong or the list i´ve linked and the picture of the fan are wrong/fake.

Edit: @Solus Corvus: its 470 btw, but still thats a very good point

[***Deimos***]

Plz correct if I'm wrong..

Amdahl's law states that if P is the proportion of a program that can be made parallel (i.e. benefit from parallelization), and (1 − P) is the proportion that cannot be parallelized (remains serial), then the maximum speedup that can be achieved by using N processors is

( (1-P) + P/N )^-1

Therefore, doesn't matter if you got Fermi, G200, or G92, if "only" 90% of code can be parallelized, anything beyond 10 SP or CPU wont be faster.
Thus with 512 "cores", maximum efficiency requires 99.8% parallelization.

Seeing how vast majority of programs, and even games are barely even optimized for 2 cores, you can imagine how much ingenuity this requires on GPU side.