GTX 460 Reviews

Wondering about the GTX 460.
Previously it was widely known that nVidia artificially limited the DP performances of its consumer level cards.

I also remember reading about "Cracked" versions of the drivers to take off this limitation.

Does it hold to the GTX 460?

 

I wasn't really talking about AA to begin with. The issue was simple: are GT215 and Juniper of the same ilk? And my answer is: about as much as Redwood and GT215 are. I'm not really saying that Juniper is a more appropriate comparison to GT215 than Redwood, just that they are both roughly equally valid.

Transistor count doesn't mean much more than, say, shader count. Obviously, NVIDIA gives up some density for speed in the hot clock domain. TSMC doesn't charge by the transistor anyway.

Why is it so crazy to think that manufacturing cost (because when I mention die size, that is what I'm really talking about, I just don't have definite figures for that) is a good criterion to group graphics cards together?

Hmm yeah, but there aren't as many GPGPU features in GF104 as there are in GF100. Namely, there's no ECC and DP is much slower. Still, GF104 does pretty well overall.

You mean something like 64 "CUDA Cores" per SM? In order to improve area efficiency, I assume… Seems possible, they don't need more than 16 polymorph engines anyway, and I don't really see any other reason to have many SMs. So maybe something like 16 64-wide SMs in 28nm would make a good replacement for GF100.

 

Yep, exactly. While there isn't any evidence so far to indicate good utilization on GF104's superscalar approach I think it's the only way they close the compute density gap with AMD. Finding two independent scalar instructions to issue per clock shouldn't be a huge deal compared to the four or five AMD needs but the register file bandwidth isn't there. That can be fixed next round.

I also think it's a fair assumption that they won't improve geometry throughput. It's already far ahead of the competition and probably overkill for current and upcoming titles. So there's a good chance the number of GPC's and polymorph engines won't change next generation. I'm expecting a boring update architecturally - 4 GPC's x 4 SM's/GPC x 64 ALUs/SM + 512-bit bus @ ~ 500mm^2, hopefully @ 700Mhz+.

Northern Islands will provide the excitement late next year.

 

Sounds about right. And yes, hopefully, NVIDIA will reverse the current trend in frequencies:

— G92 was released with shaders at 1500MHz with the 8800 GT, then 1625MHz with the 8800 GTS 512, then 1688MHz with the 9800 GTX and finally, peaking at 1836MHz with a shrink for the 9800 GTX+;
— GT200 was released with shaders at 1296MHz, peaking at 1476MHz with a shrink for the GTX 285;
— GF100 was released with shaders at 1401MHz and limited availability, and 1215MHz and somewhat higher availability for the GTX 470.

Basically, it went from 1836 to 1476 to 1401, with the TDP going through the roof… Let's hope GF200 (or whatever its name may be) won't have 1300MHz shaders.

 

We haven't seen the GF100 refresh yet so 1400Mhz isn't the correct comparison point for G92/b and GT200b. The initial chips went from G80@1350 -> GT200@1300 -> GF100@1400. Not much of a trend either way.

 

I'm not sure there will be a refresh, I have a feeling NVIDIA will just go straight to the next architecture, preferably on 28nm. And of course, if you look at MHz/W, there's definitely a downwards trend.

And just looking at initial clocks, neglecting power entirely, NVIDIA is sort of stagnating. I wonder why…

 

They aren't going to (can't) ride GF100 all the way to the end of next year when 28nm parts become available. Based on what we've seen from GF104 they still have lots of options on 40nm to change their fortunes in the high-end segment - new chip, GF104x2 etc.

You don't need to obsess about clocks if you improve the architecture itself or add functional units. Clocks didn't change at all from RV770->Cypress for example.

 

trini's right. Besides, I thought the MHz myth was behind us.

 

850MHz versus 750MHz.

 

Yep, but they eventually hit 850Mhz with essentially the same chip on the same process. Not to mention HD 3870 before that was already over 750Mhz. In any case, neither company is bringing spectacular clock speed increases to market with each generation.

 

I'm not interested in the argument, merely the factual error.

 

trinibwoy: And RV530 was clocked at 800MHz 3 years ago Yes. But RV770 was 30% bigger than RV670, RV870 was (again) 30% bigger than RV770. Such significant increase in die-size quite explains, why clocks changed more or less subtly.

G80, GT200 and GF100 are much closer in die-size (+19%, -9% or +9% compared to GT200b). Despite the die-size stays at the same level, frequency doesn't change or even declines slightly. MHz myth isn't quite an argument, I remember some slides (at the time of G80 launch) speaking about 2GHz+ clock speed for next product based on this architecture. And that was 90nm process. 80nm, 65nm, 55nm, 40nm... and they still didn't hit this 4 years old target. It seems that things aren't as easy as they expected.

 

I sort of expect them to release a full, 1.4~1.6 GHz (shaders) GF104 and just drop GF100 for the gaming market.

Sure, but AMD (and indeed NVIDIA, prior to the last 3 generations) has been steadily increasing clocks at the same time…

Well it's not a myth: any given architecture will be faster as you increase its operating frequency. When you improve both architecture and clocks, obviously, the speedup is more significant.

With the exception of the HD 3870, sure, but that was when going from a 320-SP GPU to a 800-SP one, on the same process.
And RV790 was on a different process, a 55nm one, but higher-end. Not to mention that it featured several improvements at the physical implementation level. Cypress could have hit 950MHz pretty easily, had NVIDIA put more competitive pressure on it. Actually, I think Southern Islands will be clocked in that area.



The bottom-line is that GPUs had been getting more transistors and more MHz for a long time, and that's still the case for AMD, but strangely, not for NVIDIA. And I wonder what the underlying cause may be.

After all, even Rys and Damien originally expected GF100 to be somewhere around 1.6~1.7GHz (shaders).

 

True, the clock speeds have been stagnating for a while now. R600 hit 740 mhz on 80nm, and we had 800 mhz from RV530/510 on 65 nm. Since then we've had RV670 at 775 mhz, RV 770 at 750 mhz and RV 850 at 850 mhz. But as no-X explained in his post, the die sizes have gone up considerably and speeds are kept lower due to power constraints. But In general it seems that semiconductor scaling has hit a wall and has been showing diminishing returns with each new process.

AMD had a hard time going from 90nm to 65nm as well. the 90nm FX-62 was clocked at 3.2 ghz which afaik no 65nm chip was able to match. Heck the 65nm Athlon's didn't break 3 ghz until a year after the process was introduced. Intel had a much better transition from 90-65nm(Looking at the 65nm P4's and dual core Pentium D's compared to the 90nm versions)

Edit: Had a quick afterthought. If merely adding some decoupling capacitors allowed AMD a free speed increase of 100 mhz for RV790, why haven't they implemented the same strategy for all their chips till now? At least the high end chips where the die size increase is not as significant(Afaik RV790 was 20-30mm2 bigger than Cypress)

 

I think Cypress features all the improvements brought by RV790. Don't know about Juniper et al.

 

The decap ring increased die-size (I think it made the die 0.75x0x75mm bigger... it was more or less an afterthought). I don't believe they would use it for RV870, which was ripped of all parts, which increased die-size and weren't essential, but I can be mistaken, of course

 

Aye, I'd expect GF100 not to get much more resources devoted to it. Rather something along the lines of a GF104+ or b or whatever. Or a GF101/102. GF100 I feel is going to share the same fate as R600. A short lived product which has influence on many chips to come.

Regards,
SB

 

Sure it is because the basic premise seems to be that clockspeeds are a primary indicator of progress. We all know that is silly considering how much Intel scaled back clocks in the last few years. The Phenom/Core i7 comparison is another nail in the coffin of the Mhz myth.

A full 1.6Ghz GF104 is at best 35% faster than the GTX460. Don't think that's gonna be enough to replace GF100, especially with AMD's refresh looming.

Nvidia has also had more significant architectural changes in the same period and a different feature focus. Looking at clockspeeds alone doesn't tell you much of anything.

We've had extraordinary increases in performance and new features with each new generation. What diminishing returns are you seeing? I think your definition of scaling might be a bit too narrow, especially when it comes to GPUs.

 

Well RV670 wasn't always a clear improvement over R600 (in terms of raw performance) but that didn't stop it from replacing the latter.

GF100 just doesn't look viable for the gaming market without a very substantial performance advantage over its smaller and cheaper competitors, kind of like R600, actually. So I think focusing on GF104 is a better option for NVIDIA at this time.

 

Same with G92 replacing G80, but GF100 is significantly quicker vs GF104 than G80 was vs G92 or R600 was vs RV670. A full GF100 should still have a place in the market even against a full GF104.

But this is assuming NIVIDIA can fix GF100. In its current state it won't make too much sense if a full and higher clocked GF104 comes along