384 Sp units vs 320 SP units . Just as a hunch.
if 320 Units in cypress = 320 perf points, 384sps x 96% = 368.64 perf points for cayman. This is the rumoured 6950 that is.
Printable View
384 Sp units vs 320 SP units . Just as a hunch.
if 320 Units in cypress = 320 perf points, 384sps x 96% = 368.64 perf points for cayman. This is the rumoured 6950 that is.
no, he said a SP made of 4 units (VLIW 4) is around 96% of a SP made of 5 units (VLIW 5). So that's why i multiplied the 384 sps of the rumoured 6950 to 96%
now that is senseless.Quote:
Originally Posted by Picao84
If you wanna compare them than you have to take equal numbers of SPs
meaning
320 of VLIW 4 SPs = 0,96*320 of VLIW 5 SPs
No, it makes total sense.
384 SPs in 6950 would be the equivalent in Cypress SPs of 368 SPs since each Cayman SP is 96% of a Cypress SP, so that's why i multiply 384 with 0.96. This why i find out how what their equivalent in Cypress SPs would be, or, to put it easier, 384 Cayman SPs = 368 Cypress SPs.
@Florinmocanu
FYI
rumored 1536sp of HD6970 is equal to 1600sp of HD 5870
So we are not getting performance boost from new gen cards.
Great! ;)
the rumours i read say that 1536 would be the SPs of 6950, and it would be higher than the perf of a 5870. You can find them 1-2 pages behind.
All I wanna to say that VLIW4 SP performs not almost the same as VLIW5 SP
and
Quote:
Originally Posted by From Cayman analysis
translated with googleQuote:
Originally Posted by From Cayman analysis
A ~20% increase in performance over a 5870 would be weak. That's not even enough to beat the competitor's current stuff.
Going to VLIW4 could either save you power and die space for similar performance, or give you power/die headroom to also add extra shaders. Considering what they are already doing for the low-power segment; and also considering the size of the chip on that leaked board design. What is the likely scenario?
All these numbers are pointless to think about.
Look at it this way:
Fermi was supposed to be released with a full 512 core monster last year. AMD forecasted that, and knew that their next generation, NI had to beat Fermi 512 and be prepared to fight against Nvidia's 2nd gen of DX11 cards.
32nm gets cancelled, so AMD probably has to scale NI back a bit for 40nm, but they see Nvidia has stumbled out the gate. Still, they know Nvidia no doubt has the 512 card and possibly bigger refreshes available.
So what would you do in that situation as AMD? Release a new 69xx series, first time a single GPU is using the x9xx moniker since way back when big GPUs were AMD's thing, where its only 20% faster than Cypress and just beats a GTX 480, a card that wasn't even meant to be the high end card from a year ago?
Of course not, you're planning to have Cayman compete with the 512 Fermi with Antilles to take on any dual GPU / refresh of Fermi. It makes no logical sense for AMD to sit around 14 months later with a card that barely beats a card (the 480) that wasn't even meant to be their single GPU crown a year ago
"1536sp of HD6970..." If that's true than the 6970 is going to be a great card indeed. If they found a way to make the card that efficient there is no telling how powerful the card will be with what other improvements they've made. Add that with whatever the clock rate is, off chip buffering, what appears to be a bigger die, 2Gigs of vram, etc could explain the higher TDP! The HD6970 is shaping up to be a great card so far. I hope it's not to long before benchmarks start showing up.
This is true if the number of SPs are equal. meaning that one VLIW4 SP is 20% faster than one VLIW5 SP.Quote:
Originally Posted by Solus Corvus
Though I don't have confirmed info on SP number for both 6970 and 6950 cards.
If some one has, please, let me know. I'll buy him beer :toast:
Comparing individual SP counts is the wrong way to go about it. If current rumours stand up you're looking at 24 cores for Cayman and 20 for Cypress. That's already a 20% increase in core count without talking about clockspeed increases on top of that.
QFT
you have to compare the amount of SIMD units, not the amount of SP
Cypress: 320
Cayman: 384
that is a 20% increase, if you count in the higher clockspeed and new front end already known from barts you beatt the 480 without any problems....
580 is another story and i think that the 384simd card is going to be the 5950 not the 5970
What happened to the whole 1920 SP thing? How come it's suddenly 1536? Why would they have slightly less shaders that are slightly better performing for a flagship single GPU card by going from 5D to 4D and have a higher TDP? This doesn't make any sense. As of late I've been seeing so many Cayman rumors being twisted and turned and now everything is blown out of proportions with all these rumors and it seems now that the 580 is perceived to be a faster card, but there is so much crap information going around it's hard to see anything.
Do you think maybe the 1536 shaders could be refering to the 6950 as suggested? Because if the 6950 is 384x4 = 1536 then the 6970 would either be 384 x 5 = 1920 or 480 x 4 = 1920?
Also, what is up with the memory count? Is it 1GB or 2GB? Or is it only 2GB on the 6970?
I don't know what the hell I am talking about.........:confused:
It's 2GB only for cayman.
6870/6850 is 5D shaders, and cayman is 4D shaders.
So maybe the 1536 rumour is true, but the number is not very important, because you can't compare to an old ATI/AMD gpu. The thing i can be sure, that my estimation about cayman will be corect.
I say cayman will be 35 to 40% faster than cypress. ( XT to XT. ) Pro will be different i think, and a i believe it's gonna be lower.
The thing the most important is not the number of shader but the improvement about the weaks of cypress and bart. That's i wait a lot for.
The thing that would be very nice, is a portal 2 key in package, and amazing, a software that do HD video high speed encoding with multipass, and high number of settings.
I think a lot of gamers here have some TB/s of video data, and don't have time to covert one file, after one file. Just select a folder and convert in high speed with good quality.
I have a lot of old file in old codecs, and i would like use much more advanced codec to lower the file sizes without loss or a very small.
No one is saying the 6950 is the one with 1536 shaders at beyond 3d, it is the 6970. This number came from 64* 24. Pro is supposedly 1404 or 64 * 22. This forrest guy gave the true specs(not the cut down numbers) of cypress and juniper 2 or 3 weeks before release and he shows that he had production cards in the past so this can be taken with alot less salt than usual.
But with 1536 shaders, my performance expectations have gone done substantially.
No, a VLIW4 SP is most likely a tad slower than a VLIW5 SP. But a 1536 (384) shader VLIW4 chip will have 20% more real useful SPs than a 1600 (320) VLIW5 chip.
A 320x5=1600 chip will have a lot less useful power than a 384x4=1536 chip.
Even if a Cayman shader is 96% of a Cypress one performance wise, there are still 20-50% more of them depending on which rumor to trust.
Your number of 20% higher performance is based on 20% more shaders.
You also have to look at where the cards are performing relative to one another
If the 6870 is ~5-10% slower than the 5870, how can the 6970 be only 20% faster than the 5870 if there's a 6950 to slot there too. Assuming the 6950 and 6970 have a 20% gap in performance (akin to 5850 and 5870), you'd certainly have to expect the 6950 and 6870 to have a sizeable gap as well. From there one ought to extrapolate where 6950 and 6970's are going to perform