NVIDIA GF100 & Friends speculation

[Jawed]

The late- and crap-ness of GT21x GPUs doesn't stop them being over-priced.

Jawed

 

[FrameBuffer]

Oh I see what has happened now yes.

The only issue remaining is v1.0 vs v1.1 and the "30% boost".

I thought the 30% boost was in relation to the upcoming 2.0 Uengine update.. ?

 

[rpg.314]

The late- and crap-ness of GT21x GPUs doesn't stop them being over-priced.

Jawed

If they are overpriced AND crappy, then it'll show up in market numbers for Q210. Otherwise, we'll have to conclude that they aren't that crappy or overpriced.

 

[Bouncing Zabaglione Bros.]

With Fermi based GTX 400s, they are launching much later than the products they are trying to compete with. They can't charge more for it, especially when the performance isn't there (if these numbers end up being true).

All Nvidia have to do is find 5000 people worldwide who will pay $600+ because they are die-hard Nvidia fans, or want the latest/newest tech. A bit of hype on the exclusivity (ie spinning the lack of product). Some people will want SLI, some people just have a lot of money to spend on their hobbies, and some will buy to resell on Ebay at inflated cost for a quick profit.

If Nvidia only can supply the early adopters who will pay any price, then that's who they will get the most money out of. Mainstream pricing doesn't matter when you can't supply any product into the mainstream, and you're only servicing your niche/vanity customers where money is virtually no object.

 

[FrameBuffer]

Whilst I'm in no way expecting the GTX 470 to launch north of $600, if the rumours of only 5000 units being available at launch are true, why sell them all off at less than $400, when Nvidia know that they're going to sell out no matter how high they price them (well so long as they don't start charging $1000!)? That follows common sense as well, surely?

I keep seeing this 5000 GTX 470s worldwide however from what I was lead to believe, 5000 is absolute low end with the high end being near 9000 GTX 480s and that for every GTX480 there are about 9 GTX470s... which would mean at launch an avg of 7000 GTX80s and about 70,000 GTX470s. Again it seems from what I was told the numbers being reported of <10K units were for 480/512 parts but then again.. I've been burned before however I have no reason to doubt said acquaintance/friend.

 

[FrameBuffer]

Ah yes, nice observation of the grass scene. The GTX480 starts at 44fps and finishes at 72fps at the windmill. The HD5970 starts at 59fps and finishes at 112fps at the windmill.

But the Ati cards take a dive with the tessellation & wireframe. GF100 is just a beast with tessellation... funny considering all the crap that was talked about it having "dedicated tessellation hardware"...

eheh.

While all the fame and glory seems to go to the high end halo products I still question the mainstream/value segment as where things will be won/lost. After all given the hoopla over ATI's "dedicated" hardware for Tessellation it would seem things will only get worse for NV as derivative parts are brought forth.. and as such where full blown GF100 GTX 480/470s are capable of achieving (arguably) better performance than their ATI counterpart... how will NV compare against ATIs 5700/5600/5500/5400 as their "dedicated" hardware has little change but NV's could possibly see upwards of a 75% reduction. Granted on the extreme lower end the use of Tessellation will be limited due to actual performance anyway, or at least until it's use becomes highly refined and matures.. and that remains highly doubtful for the short term (<2-3 years).

 

[Razor1]

While all the fame and glory seems to go to the high end halo products I still question the mainstream/value segment as where things will be won/lost. After all given the hoopla over ATI's "dedicated" hardware for Tessellation it would seem things will only get worse for NV as derivative parts are brought forth.. and as such where full blown GF100 GTX 480/470s are capable of achieving (arguably) better performance than their ATI counterpart... how will NV compare against ATIs 5700/5600/5500/5400 as their "dedicated" hardware has little change but NV's could possibly see upwards of a 75% reduction. Granted on the extreme lower end the use of Tessellation will be limited due to actual performance anyway, or at least until it's use becomes highly refined and matures.. and that remains highly doubtful for the short term (<2-3 years).

It doesn't matter too much about the tessellation units, if the shader power isn't there too, tessellation performance doesn't only depend on the dedicated hardware for creating the traingles.

 

[DavidGraham]

we have optimized Heaven Benchmark, agressive culling gave 30% boost. We'll show updated version at GDC,

What do they mean by aggressive culling ? is it the same as back face culling ? will it result in decreasing the load on the geometry units and thus improve performance ? .. shouldn't that be applicable on both Nvidia and ATi hardware ?

 

[Sontin]

While all the fame and glory seems to go to the high end halo products I still question the mainstream/value segment as where things will be won/lost. After all given the hoopla over ATI's "dedicated" hardware for Tessellation it would seem things will only get worse for NV as derivative parts are brought forth.. and as such where full blown GF100 GTX 480/470s are capable of achieving (arguably) better performance than their ATI counterpart... how will NV compare against ATIs 5700/5600/5500/5400 as their "dedicated" hardware has little change but NV's could possibly see upwards of a 75% reduction. Granted on the extreme lower end the use of Tessellation will be limited due to actual performance anyway, or at least until it's use becomes highly refined and matures.. and that remains highly doubtful for the short term (<2-3 years).

nVidia has 16x FF-Tessellation units, AMD only one. But i don't see a 16x performance lead for nvidia. Maybe the performance is not limited by the number of the FF-Tessellation units...

 

[Jawed]

It doesn't matter too much about the tessellation units, if the shader power isn't there too, tessellation performance doesn't only depend on the dedicated hardware for creating the traingles.

It seems ALU capability isn't dominant in tessellation performance in Heaven. With ~13x more primitives performance is dropping by no more than 25% on NVidia, yet by ~70% (maybe less with version 1.1?) on ATI.

Though that doesn't actually rule out the possibility that the ATI compiler is entirely broken with HS and/or DS. But the chip has been up and running at AMD for ~1 year now.

Jawed

 

[Razor1]

It seems ALU capability isn't dominant in tessellation performance in Heaven. With ~13x more primitives performance is dropping by no more than 25% on NVidia, yet by ~70% (maybe less with version 1.1?) on ATI.

Though that doesn't actually rule out the possibility that the ATI compiler is entirely broken with HS and/or DS. But the chip has been up and running at AMD for ~1 year now.

Jawed

That is true, but for lower end ATi cards, they still have the same tessellation units but less ALU's, so the performance factor won't be the tessellation/or setup its going to swing towards ALU's, now don't know how much shader power drop will cause the ALU's to be the primary bottleneck, but just saying you will still need all three, setting up the triangles, tessellation, and shader power for the HS, DS. It might have been a good move NV's Fermi has variable degree's of all these three, will have to see in real world tests.

 

[rpg.314]

What do they mean by aggressive culling ? is it the same as back face culling ? will it result in decreasing the load on the geometry units and thus improve performance ? .. shouldn't that be applicable on both Nvidia and ATi hardware ?

There are many types of geometry culling. One possibility is frustum culling.

 

[Jawed]

That is true, but for lower end ATi cards, they still have the same tessellation units but less ALU's, so the performance factor won't be the tessellation/or setup its going to swing towards ALU's, now don't know how much shader power drop will cause the ALU's to be the primary bottleneck, but just saying you will still need all three, setting up the triangles, tessellation, and shader power for the HS, DS.

HD5770 is considerably more than half as fast as HD5870 on the benchmark.

I'm not sure what performance is like with HD5770 on the worst-case close-up of the dragon though.

Jawed

 

[GZ007]

nVidia has 16x FF-Tessellation units, AMD only one. But i don't see a 16x performance lead for nvidia. Maybe the performance is not limited by the number of the FF-Tessellation units...

The hull shader and domain shader is another unified shader stage like the pixel shader or vertex shader. They ALL use the same SP-s. So its not something like 16 vs 1.
ATis teselation unit was surely done with all 1600 SP-s in mind and nvidias with one SM(x16) so i wouldnt compare them 1 on 1.

 

[DavidGraham]

There are many types of geometry culling. One possibility is frustum culling.

What I meant is , shouldn't the "aggressive" culling Nvidia talked about result in performance improvements for both Nvidia and ATi hardware ? or just solely on Nvidia hardware ?

And if ATi hardware is involved too .. why didn't they do it long time ago ?

 

[trinibwoy]

nVidia has 16x FF-Tessellation units, AMD only one. But i don't see a 16x performance lead for nvidia. Maybe the performance is not limited by the number of the FF-Tessellation units...

It's too early to tell for sure but the number of tessellation units is probably of no real consequence since as you said it seems way over-engineered in Fermi. The more likely limitation is 4 triangles / 32 pixels per clock. In any case, even rasterization might not be too worrisome. 2GPC/16ROP and 1GPC/8ROP combinations work just as well. It's the more realistic stuff like 1GPC/16ROP combos that might require changes somewhere.

ATis teselation unit was surely done with all 1600 SP-s in mind

That's not obvious at all at this point. If that was the case we shouldn't expect Fermi to have any sort of advantage in heavily tessellated scenes.

 

[rpg.314]

What I meant is , shouldn't the "aggressive" culling Nvidia talked about result in performance improvements for both Nvidia and ATi hardware ? or just solely on Nvidia hardware ?

It should help both, AMD a bit more I guess as it has less setup/raster/tess throughput (on paper atleast).

And if ATi hardware is involved too .. why didn't they do it long time ago ?

Devs are still learning I guess.

 

[mczak]

While all the fame and glory seems to go to the high end halo products I still question the mainstream/value segment as where things will be won/lost. After all given the hoopla over ATI's "dedicated" hardware for Tessellation it would seem things will only get worse for NV as derivative parts are brought forth.. and as such where full blown GF100 GTX 480/470s are capable of achieving (arguably) better performance than their ATI counterpart... how will NV compare against ATIs 5700/5600/5500/5400 as their "dedicated" hardware has little change but NV's could possibly see upwards of a 75% reduction. Granted on the extreme lower end the use of Tessellation will be limited due to actual performance anyway, or at least until it's use becomes highly refined and matures.. and that remains highly doubtful for the short term (<2-3 years).

FWIW, this article here http://ht4u.net/reviews/2010/nvidia_geforce_gtx_470_480_cebit/index5.php says nvidia is NOT planning on a low-end fermi derivative, because it would be pointless (tell that to AMD...). That doesn't exactly inspire confidence on perf/mm^2 for lower end parts...
Though this really only affects the bottom of the barrel, even gt220 isn't low end according to that anymore. But it should mean we won't see a fermi part with 32SP (since gt220 was already 48SP), if they only scale SMs across GPCs and GPCs it could mean the lowest end fermi part planned is something along 1 GPC / 2 SMs (with 32 "cores" each), which would smoke Cedar but look awful compared to Redwood, so maybe it will be something completely different after all...
BTW I'm looking forward to see clocks for GTX480 as they were supposed to be leaked today (or rather, nvidia is supposed to give this information to partners today ).

 

[DavidGraham]

It should help both, AMD a bit more I guess as it has less setup/raster/tess throughput (on paper atleast).

So , I guess the comparison between version 1.0 and 1.1 is completely invalid , right ?

 

[trinibwoy]

BTW I'm looking forward to see clocks for GTX480 as they were supposed to be leaked today (or rather, nvidia is supposed to give this information to partners today ).

Yeah hopefully we get something solid this week. There are people claiming the samples in AIB hands aren't final in terms of clocks and cooling apparatus. What was the point of sending out the samples in the first place then?