Dont think the 7870 will have only 50% the performance of a 7970.
mmm..ok, we'll see :smile:
7970 is about 40-50% faster than than 6900 series which is about where I expect the 7800 series to end up.
Expectation from performance ASICs, Pitcairn and GK104 is that they offer upto 3/4 performance of the top-end at roughly half the price.
Is that necessarily a sign of incompetence on the engineering side or could that also be explained with conscious design decisions. For example, power gating is not free in terms of transistors/die area from what I've heard.
Charlie says it's 324-361mm², but he said nothing about transistor count.
Probably basing it on Ail's comment earlier about the transistor count (3.4/6.4b) being close for one of the ASICs.
Well if it's not that small and not that fast I sort of doubt it will be very power efficient neither (though if nvidia sacrifices some OC potential it could certainly beat Tahiti there).This is getting pretty funny.
"GK104 is really small, really fast, really power-efficient".
"OK, turns out it's not all that fast."
"OK, turns out it's not small either."
What's next?
If true, I like them specs (along with their performance estimates), but the prices could be a tad better.
The lesser framebuffer seems to be able to provide Nvidia with a little better price/performance ratio than what AMD offers atm. That will be until the 7950 1.5GB comes out anyway. I wonder if AMD will consider bringing a 1.5GB 7970, if this table is correct.
I just hope none of the 680/670/660 Ti will come with some silly axial HSF solution.
A higher transistor density doesn't come only with advantages. Just for the record's sake you're comparing the transistor count of a pityful fake specification table to a probably close guess for the die area. If you want to go with lenzfire's crap you should use their hypothetical 290mm2 die area with their hypothetical transistor count. Not that it'll reflect reality in the end, but no one said that the lenzfire transistor count is 100% correct either.
Considering some of the numbers there are simply impossible (224bit memory interface) and others simply dumb (performance difference between some GPUs) I wouldn't really take that table all that seriously.
We also don't know if both companies count transistors the same way (see Bulldozer).
Charlie's earlier die size and power consumption guesses seem to be completely based on the cooler he saw on A2 silicon. That doesn't say much since we've already seen that you can cut Tahiti's power consumption a lot if you get a good undervolting sample.
Charlie's earlier die size and power consumption guesses seem to be completely based on the cooler he saw on A2 silicon. That doesn't say much since we've already seen that you can cut Tahiti's power consumption a lot if you get a good undervolting sample.
There's counting things differently, and then there's just not doing a good job of it.
The ASIC side of AMD seems to have a much better idea of the count than whoever was in charge of getting Bulldozer's count the first and second time, going by the little joke they had in their GPU slide deck.
The ASIC side of AMD seems to have a much better idea of the count than whoever was in charge of getting Bulldozer's count the first and second time, going by the little joke they had in their GPU slide deck.
Last edited by a moderator:
Why is 224bit memory interface impossible? The gddr5 chips are all 32bit devices so this should be doable. Granted all amd/nvidia chips currently use 64bit MC channels but this is not something you couldn't change.
(That's not saying I believe the 224bit interface of course just I don't see why it would be impossible.)
With so many different numbers around, what if some of that confusion is due to a similar architectural split between Kepler high-end and midrange chips as we've seen with Cayman vs. Barts?
Maybe GK110 will be the only "real" next-gen chip - and GK104 basically is Fermi evolved/streamlined?
What really got me thinking was that - according to Charlie - the GK104 die is almost SQUARE. Very unusual for an Nvidia midrange chip (which usually is "chopped down" from the top dog chip - and ends up rather rectangular). Makes me think that GK104 actually is a rather "independent" design.
So ... just carrying that Fermi-evolved thought for GK104 a little further for the sake of speculation: How much Fermi could they possibly squeeze into a ~3.5 billion transistor chip?
Maybe GK110 will be the only "real" next-gen chip - and GK104 basically is Fermi evolved/streamlined?
What really got me thinking was that - according to Charlie - the GK104 die is almost SQUARE. Very unusual for an Nvidia midrange chip (which usually is "chopped down" from the top dog chip - and ends up rather rectangular). Makes me think that GK104 actually is a rather "independent" design.
So ... just carrying that Fermi-evolved thought for GK104 a little further for the sake of speculation: How much Fermi could they possibly squeeze into a ~3.5 billion transistor chip?
I don't recall performance chips before GF104/114 as rectangular but could be wrong. And no neither the GF104 nor the GF114 is any sort of "chopped down" from the top dog chip for heaven's sake. A GF114 has per SM 3*16 SPs with 8 TMUs/SM, while a GF110 has per SM 2*16 SPs with 4 TMUs/SM. It's in fact as "chopped down" that io and behold the performance part has suddently twice the TMU amount as the top dog there.
"double confirmed" = Le wild guess
"our sources" = random post on chiphell
"XXX will lose to YYY" = there's this one benchmark where YYY performs better
"internal performance estimate" = it's the best we could come up with
"our sources at XXX" = this guy got fired months ago
"We've seen the board" = someone described it in not too much detail
"it will be XX% faster" = In this one benchmark it is almost XX-1 faster
"leaked chart"= check out my guesstimates in excel
but that was obvious
Last edited by a moderator:
"double confirmed" = Le wild guess
"our sources" = random post on chiphell
"XXX will lose to YYY" = there's this one benchmark where YYY performs better
"internal performance estimate" = it's the best we could come up with
but that was obvious
You should start writing a full silly season glossary!
My guesses are the GK104 part coming out in April-May is this part:
GK104
~350mm^2
GTX 660 or Ti
1536 Cuda Cores
64 TMU's
32 ROP
1050-1100mhz Core
1050-1100mhz Shader (No hotclock)
5800mhz GDDR5 (2GB)
256bit Memory interface
~200-215w TDP
Bandwidth: 186GB/sec
4600 Mtris/sec
3.3 Tflops (SP)
35 or ~65 GPixels?
67.2 Gtexels
~On par with 7970 +5%
~X3000 3d Marks
Price cheaper than 7970
Smaller Die, low power consumption, good price. Cuda core complete overhaul compared to Fermi.
This chart from Dec/early Jan may be correct chart:
GK104
~350mm^2
GTX 660 or Ti
1536 Cuda Cores
64 TMU's
32 ROP
1050-1100mhz Core
1050-1100mhz Shader (No hotclock)
5800mhz GDDR5 (2GB)
256bit Memory interface
~200-215w TDP
Bandwidth: 186GB/sec
4600 Mtris/sec
3.3 Tflops (SP)
35 or ~65 GPixels?
67.2 Gtexels
~On par with 7970 +5%
~X3000 3d Marks
Price cheaper than 7970
Smaller Die, low power consumption, good price. Cuda core complete overhaul compared to Fermi.
This chart from Dec/early Jan may be correct chart:
