and a lot more expensive...
and a lot more expensive...
Well, first, you don´t always need that bandwidth, second, comparing architectures between NV and AMD gives me a headache, because comparing between NV architectures is already quite a stretch. You have to keep in mind that GK104 is NV´s answer to AMD´s "high-end" (insert a LOL here) ASIC with a relatively low BOM and great performance for a great price, so there will certainly be some limitations, of course.But there have been rather significant architecture changes in between. Hard to expect similar between a GF114 and GK104. With a 256bit bus, it will be at best 50% behind in bandwidth compared to a 7970... how much more bandwidth efficient can it be?? Maybe they're using XDR2 after all?
If you want insane numbers, GK104 will probably already impress on serveral fronts, but what I´m really curious about is if NV can deliver a can of whoop-ass (tm) part 2 with GK110.
Last edited by a moderator:
Right but if Charlie is really implying Tahiti-ish performance then the competition is $549. Or it might not be about performance at all but simply better pricing.
Where does he imply that?
What else does he have to compare it to for him to start declaring winners? As far as I know there are no details on Pitcairn yet.
Unless its a big hoax and he's fucking with everyone
What else does he have to compare it to for him to start declaring winners? As far as I know there are no details on Pitcairn yet.
Unless its a big hoax and he's fucking with everyone
When did he ever need any proof or even evidence of anything to make a claim?
Ailuros hinted that the fastest GK104 will probably be called GTX 680, though. And I don't think Charlie knows the exact price yet, either.
btw, someone mentioned AMD would drop prices around that time-frame and probably release a performance driver as well.
The architectural change between R580 and R600 was much bigger than the change between R600 and RV770, yet RV770 delivered a superior performance jump.DarthShader said:
Maybe GF100 was Nvidias R600, and GK104 will be their RV770.
Jawed
Legend
I'd like to see something Larrabee like. Texture decompression as FF but everything else in ALUs. This balance minimises the stuff that's sat doing nothing in compute.
Proper task-parallelism in each core, too.
Since Charlie seems to like this, I guess that means I'll be disappointed yet again.
Proper task-parallelism in each core, too.
Since Charlie seems to like this, I guess that means I'll be disappointed yet again.
Or he has seen a selected benchmark were GK104 actually beats Tahiti (hence "wins in any metric") while running at a lower power consumption.
Compared to nV, Tahiti probably still leaves some performance on the table in the work distribution/front end even as it was improved. At the same time, process variations during the first batches for Tahiti probably caused the incredibly conservative frequency and voltage settings for the HD7970. Some people are running the chip at stock frequency with a 0.25V or even 0.3V lowered voltage! It's possible AMD also underestimated the qualities of TSMC's process. NVidia may have tuned their Kepler GPUs closer to the performance/power sweet spot of the process and AMD would need a full respin for that. Or AMD is indeed using HPL while nV uses HP, who knows? We will see.
Seriously, I dont know why people are shocked by the fact that GK104 might beat or be closer to Tahiti... GF114, which was just a fixed GF104 says hello? Go back to the GTX560Ti reviews and see where it stands against a HD5870. And a GTX580? Its just a bug fixed GF100. So, if GF100/GF104 would have been what GF110/GF114 were since the beginning, and they werent just because nVIDIA had problems with them, the landscape would be kinda similar to what we might have now...
Well the 560 Ti is between 15 and 30% behind the 6970. If the "680" is behind Tahiti by a similar margin then yeah nothing really changes. It's probably the pricing on the 7970 that's making the most difference but AMD can adjust that to suit.
If we assume that GK104 is 30% slower than a 7970 then the 680 moniker makes no sense cause that's no faster than a 580 (9800GTX notwithstanding). So either Charlie is bonkers or the 680 isn't showing up anytime soon and GK104 will be named something else.
DavidGraham
Veteran
I was comparing Fermi against Cypress, hello? GF114 IS what GF104 should have been, and its on the same performance ballpark...
GF110/114 were more than just bug-fixed GF100/104, they were silicon respins that not only benefited from a 40nm process that was more mature by a full year, but also (most likely) had been partially redesigned for higher clocks and/or lower power.
But that's not really the point. What's surprising here is that GK104 is supposed to feature a 256-bit bus. Unless it also has really fast memory, that means it has about 2/3 of Tahiti's bandwidth. So if it does manage to outperform it anyway, that's quite a feat.
Yes, but the underlying architecture is the basically the same, is it not?
And the respins were needed to get it to the performance estimated when the chip was designed, I guess. Thats the "estimated performance" im hinting at here: it was as fast as Cypress.
For real, it is a very strange thing indeed. I almost don't know what to make of it unless somehow he was upset about a lack of info regarding the recent release of AMDs card. Otherwise he seems to have suddenly joined the ranks of the reasonable (not b/c of what he says either, but in general b/c of how he said it, things like they way he described possible hitches and so on).
I have no idea. Just trying to piece things together. How fast would a GK104 need to be (and at what price) for Charlie to be so unconditionally impressed? It's weird.
Configuration wise maybe, certainly not clocks!
What if it is within, say 5%, of 7970?
And under $400? Then Charlie's article sorta makes sense. I would bet that a big part of it is price related.
No kidding. It was almost....normal. Maybe he's in shock. Or nVidia replaced him with a body double
