Facebook/Twitter launch more like it.
GF100 evaluation thread
- Thread starter rpg.314
- Start date
Could deferred shading mitigate the 4x fragment shading penalty with micropolygons?
The int32 mul, which is essential to half rate dp in fermi, almost assuredly takes as much area as the fp32 mul unit.
The number of fp32 and int32 multipliers is matched.
Without knowing how much area the FP32 mul takes up as a fraction of die size, we don't know how much area the FP64 functionality takes up on top of that.
Bear in mind, by the way, that the cost to nVidia for going half-rate FP64 was almost certainly much less than ATI, for one simple reason: ATI has more math units in the same area, while nVidia relies upon more efficient execution on fewer units. The smaller number of units means that nVidia is spending more die area on getting the info to the math units themselves than ATI, which means increasing the size of those math units doesn't do as much.
aaronspink
Veteran
Yes but it is likely that each nvidia ALU requires more area due to upsizing of transistors and/or increased pipe stages in order to run at ~2x the frequency. 1/4 rate like ATI does it should be able to be implemented with almost zero area increase while 1/2 rate requires 2x the multiplier area.
Well, if I'm reading the architecture specs right, ATI also has more than 3x the math units on a 33% smaller die (512 vs. 1600). Per die area, then, ATI has somewhere around 5x the the math units. I seriously doubt that the optimizations for higher clock rate take up 5x the die area.
"Not enough" is not accurate. Anandtech's 5870 crossfire test, showed that the 16X+16X vs 8X+8X configurations, have only 2-7% performance difference.
There are motherboards like the MSI Big Bang Trinergy (not to be confused with the Fusion) which sports a NF200 chip that gives a full complement of 32 PCIe 2.0 Lanes which can be divided at 16X+16X for the graphics cards (there's also the option for 16X+8X+8X). It works for both SLI and Crossfire and I should know since I own it as well as two 5850s.
The funny thing is that its cheaper than quite a few vanilla high end P55 mobos (vanilla=No NF200).
fp alu's certainly form a substantial portion of SMs prior to fermi, though I don't have a hard number.
:|
OTOH, I think nvidia is paying substantial penalty for half rate dp. It needs int32 mul that nobody else cares for, and certainly not in matched ratios. ATI's dp is as good as free as they reuse the xyzw lanes to make one dp mul for 4 sp mul's. ATI does quarter rate dp and that is the highest ratio you can do without paying dp cost somewhere.
Just wanted to throw this out so I was being called out. I picked a random review (Firingsquad) and after running the numbers throw, results were:
GTX480 54% faster than GTX285
5870 60% faster than 4890
By the same yardstick that Cypress was concluded to be "meh", GF100 is less than even a "meh", "WTF" territory since its less
Unless the obvious double standards show up again. 
GTX480 54% faster than GTX285
5870 60% faster than 4890
By the same yardstick that Cypress was concluded to be "meh", GF100 is less than even a "meh", "WTF" territory since its less
Unless the obvious double standards show up again. That's the problem. Hard numbers are important here! If it only cost them, say, 5% of die area vs. doing it ATI's way, then it really isn't a big deal, is it?
It might be substantial as a fraction of the die area allocated to the math units themselves, but if nVidia's math units take up less than half the die area of ATI's, then it's a much lower total cost than if ATI had tried the same thing. That's what I'm trying to say.
Ninjaprime
Regular
Just wanted to throw this out so I was being called out. I picked a random review (Firingsquad) and after running the numbers throw, results were:
GTX480 54% faster than GTX285
5870 60% faster than 4890
By the same yardstick that Cypress was concluded to be "meh", GF100 is less than even a "meh", "WTF" territory since its less
... How much time after the 4890 launched did the 5870 launch? Now do the same for the GTX 285 vs the GTX 480. 4890 ---> 5870, ~6 months. GTX 285 ---> GTX 480 ~15 months. Maybe thats the problem.
Didnt you know, time is irrelevant. (along with money, power, heat & noise)
Shame on ATI for launching so soon. Investing in AMD gpus is a bad idea since they get outdated pretty quickly.
Which proves my point ? They wanted to release it in October (before the HD 5970 was launched - released in November). If they had launched it in October, they would have the fastest graphics card. Since they didn't, they can only claim fastest GPU. Isn't that surprising
You seem to be forgetting (as usual) what ATI also said about RV870 and didn't deliver (performance wise). But obviously you only have problems with NVIDIA's PR
Do you really believe that back in September they didn't know they wouldn't make it for October launch? Yet all those statements quoted before were made after Cypress' launch.
Last edited by a moderator:
AFAIK, Cypress != Hemlock, so fastest graphics card after Cypress is correct. After Hemlock, only fastest GPU makes sense...
Anyway, whatever...enough with this. It's quite pointless to discuss anything when the people involved have their minds set on something completely different.
I'm sure it is about conditioning in relation to "meeting expectations". For the 58xx series it met expectations on performance in the most and also exceeded them in the power and temp case. Therefore people were conditioned for lower temp and power on the smaller process, hence why they now feel disagreeable to 4XX's extra temps and power. If 58XX had still been as hot as 4XXX series it wouldn't be so much of a big deal.
That's the problem when you come out of the blocks 2nd, you are a lot more likely to disappoint than to blow people away.
That's the problem when you come out of the blocks 2nd, you are a lot more likely to disappoint than to blow people away.
Indeed, though I am also beginning to suspect that nVidia missed their performance target somewhat with this part. Here's hoping that later iterations of the chip have significantly higher performance per die area (and per watt...).
I think it's obvious that they did missed their targets. Clocks are certainly not high enough (especially compared to GT200b) and they certainly didn't want to come out with their flagship with units disabled. Both would without a doubt, increase performance, had they been possible.
Still, power draw is the biggest problem I see that they need to address in future iterations. If the GTX 480 had less power consumption under load (let's say 50w), the results would be far more appealing.
http://www.fudzilla.com/content/view/18277/34/
If it's this limited, it makes me wonder how much truth is there in Charlie's yields claims, despite the the ridiculously low yields.
If it's this limited, it makes me wonder how much truth is there in Charlie's yields claims, despite the the ridiculously low yields.
Yield numbers in the 70% range are what I've heard through the grapevine, and make far more sense economically than some of the more ridiculous estimates out there.
Similar threads
