It would make more sense to have a year of the older generation and 20+ months of the later, faster cards instead of say 16 months of the older gen and 16 months of the newer gen.
AMD: Sea Islands R1100 (8*** series) Speculation/ Rumour Thread
It would make more sense to have a year of the older generation and 20+ months of the later, faster cards instead of say 16 months of the older gen and 16 months of the newer gen.
Silent_Buddha
Legend
Nvidia has been off the yearly cadence for quite a while now, but AMD has thus far been ~1 year per GPU series. It'll be interesting to see if they also fall off the 1 year cadence with this upcoming gen.
Regards,
SB
Maybe, but AMD could always go for a mid-life refresh to their Sea Islands lineup, much like the 7970 GHz Edition (although with better marketing, hopefully) or the 4890, a few years ago.
It seems to me that when competition is tough, it's always best to release improved products as soon as possible.
Is there anything that either future drivers/gpu can do to free up hardware threads so that a game restrained to just one thread could play more efficiently? For example (look at the 'update' portion towards the end):
http://www.pcper.com/reviews/Graphi...and-PhysX-Comparison-GTX-680-and-HD-7970/GPU-
Which indicates that BL2 is waiting for a single thread to finish which results in lower performance.
http://www.pcper.com/reviews/Graphi...and-PhysX-Comparison-GTX-680-and-HD-7970/GPU-
Which indicates that BL2 is waiting for a single thread to finish which results in lower performance.
I wonder if using a ramdisk could elevate some of the performance issues using a single HW thread.
I read something not long ago that AMD might be using a ramdisk. At what capacity would this be available? Would GPU users have access?
http://news.softpedia.com/news/AMD-Radeon-RAMDisk-Will-Make-Games-525-Faster-295452.shtml.
Edit:
I found more information but they are talking about it's use with memory products...
http://corporate.dataram.com/compan...xecutes-agreement-with-amd-for-radeon-ramdisk
Last edited by a moderator:
The cpu fallback of physx is just some old single threaded x87 code, coded by a retarded monkey on crack (or well, probably just made by ageia 5+ years ago and therefore not suited for today).
Making the cpu implementation of an interface made for the gpu multithreaded should be pretty damn easy, if nvidia wanted to supply a serious physics middleware. But with about 1 decent title pr year I guess noone bothers..
Making the cpu implementation of an interface made for the gpu multithreaded should be pretty damn easy, if nvidia wanted to supply a serious physics middleware. But with about 1 decent title pr year I guess noone bothers..
It's disconcerting to see that BL2 used physx 2.8.4 when physx 3.0 supposedly fixed those issues was ready in May of last year. I wonder if the newer version was released if AMD could take advantage of it's CPU performance?
Im more concerned by the fact PCper use a benchmark provided by Nvidia ( bugged with PhysX who stay on High on 2 scene ( even if you set it at low )
saying, this is what you will see as difference when playing .. obviously, the benchmark is only made for show PhysX effects .. ( just look the video ).
I like the article.. " as a new good game is release, we have decided to do an article about it and PhysX"... " what an hasard, Nvidia just send us a benchmark tool for it who will be certainly included in the game soon"
I dont understand why they have not just use fraps, on a choosen Game sequences, knowing the cpu thread problem, do it 3-4 times in a row. We dont have any information on the cpu used and at what speed.
If you want to mesure the impact it will have, you choose a typical game sequences and you go with it, not using large plans showing the additional PhysX effect in a raw sequences of 1min30
saying, this is what you will see as difference when playing .. obviously, the benchmark is only made for show PhysX effects .. ( just look the video ).
I like the article.. " as a new good game is release, we have decided to do an article about it and PhysX"... " what an hasard, Nvidia just send us a benchmark tool for it who will be certainly included in the game soon"
I dont understand why they have not just use fraps, on a choosen Game sequences, knowing the cpu thread problem, do it 3-4 times in a row. We dont have any information on the cpu used and at what speed.
If you want to mesure the impact it will have, you choose a typical game sequences and you go with it, not using large plans showing the additional PhysX effect in a raw sequences of 1min30
Last edited by a moderator:
UniversalTruth
Veteran
Here, because obviously the other thread is for something else.
Ok, I think about two possibilities. Either these 15% are wrong and we will be surprised.
Or they will significantly reduce pricing of current line-up and put these 15% more for the same price as it is the HD7970 now.
iMacmatician
Regular
If this is the case, then maybe the 88*0s and 87*0s will have closer to 30% performance gain over their respective 7**0 cards (especially if that Oland rumor is anywhere near accurate).
Taking it a step further, maybe the 89*0s were planned to be ~30% above the 7970 GHz Edition and the 7950 Boost, but due to power they scaled them back to ~15%.
15% more than? Is it per W? Per SP? Or e.g. 8770 vs. 7770?
caveman-jim
Regular
I'm thinking CI is to SI what NI was to Evergreen. Tweak and tuck, bigger, more power, bit more performance.
SI is a 2012 line, CI is a 2013 line, 15 months between launches isn't too bad, especially with another 15 on the cards availability for next TSMC node. On a marketing slide that's still a yearly cadence.
SI is a 2012 line, CI is a 2013 line, 15 months between launches isn't too bad, especially with another 15 on the cards availability for next TSMC node. On a marketing slide that's still a yearly cadence.
Sounds about right. History seems to repeat itself in GPU manufacturing the last several years. The top GK110 part will probably be a smidge faster than HD 8970, albeit much larger and power hungry, similar to GF110 vs the top Cayman part.
Personally I want to see how Oland(8870) performs against GK104(114) since both will be relatively similar size...
UniversalTruth
Veteran
ROPs will without doubt stay at the same level all along 28nm node. And yet we really dont need more of them, rather higher core all along
why o why we always have mine is biggger issue here
afai saw most of S'thern Aylands MC could bear (be OCed to) stunning 1900MHz GDDR5. And envy's GPU only breed has only 6GHz 256-bit wide MC (thats 1500Mhz GDDR5 MC)
So even i highly doubt they would need anything much larger chip than current 350mm2 (TahitiXT) for redesigned 512b wide memory bus it's obwious that even 384b is far more than really neede. With proper redesign they'll probably could offer standard 1800MHz GDDR5 on 384-bit bus or even reduce it and use smaller chips as GK104 is smaller and more effective GPU competition
I hope for larger chip that could be clocked further than 1100Mhz and have less leakage and 3048 SU and same TMU ROP numbers just improved functionality.
But probably we wont see neither but quite opposite like in GK104. we could see smaller 192b wide chips leading way for dual GPU crown yet again .... 1536S/96TMU/24ROP .... all ati needs is final improved algorithm for "picture patching" now that have more simpler SIMD based shaders. And this beast could be highly clocked (1300Mhz) delivering double performance with two smaller dies
I just hope they wont cripple down double precision functionality from Tahiti's succesor with dummy explanations how its unnecessary just to keep their pace regarding GPU die sizes with competition. Something we saw in Evergreen -> Barts move on same 40nm node.
We'll se same here 28nm->28nm with same improvements. So they'll had to cut it somewhere if they dont go to multichip strategy
Last edited by a moderator:
It's never that easy. The guy behind chip-architect.com (i think here) once said that all the new idea about Bulldozer is in fact resurrected idea behind one of the K8 prototypes. That prototype should be AMD64 chip based on two separate double pumped buses (original SMT) found in then contemporary P4 architecture that even Intel disliked and abandoned.
AMD now refined that idea that they scrapped as it was too huge to be implemented on 130nm SOI, but it's pretty radical change in times they need to cut down their expenses after they bought ATi and need budget for two huge separate R&D. And BD suffered, been delayed lacking funds ... And in fact all that doule pumping idea not being brightest idea we could say it is quite decent that Bulldozer didnt sunk us down like P4 did after P3 architecture.
Some major improvements except few fiddling & tweaking with pipeline optimization we wouldnt see on 32nm node, nor in current Piledriver incarnation, as we could see today. Hopefully IPC could be improved, but there's two way to do that. %1 more instruction schedulers inside core whic is old way and they could do that w/o BD or %2 by adding more cores inside a module and even lighten them up to be more module friendly. So fatter modules with 4-cores seems appealing. And tha would be "Bulldozer Next Generation"
I somehow believe that better front end arbiter that well saw in R1100 would be a part of R&D they already did in development of Steamroller CPU @22nm (BD further refinement)
And other thing their really need to work on is speculative cache management and branching of those monster while they still didnt perfected 2-core modules, adding more of them seems insane but they should do that sooner rather than later because we really dont have time to wait DAMNs tick-tocking like intel did for millenia.
UniversalTruth
Veteran
off/
Is the problem really transistor size?
Ok, I don't think it's a good idea to resurrect an old scrapped one. Obviously it had been scrapped because it was worse.
Which actually gives an expected result- Bulldozer is slow.
Should have shrunk Thuban and add more speed through frequency increase. (headbang)
endofoff/
Is the problem really transistor size?
Ok, I don't think it's a good idea to resurrect an old scrapped one. Obviously it had been scrapped because it was worse.
Which actually gives an expected result- Bulldozer is slow.
Should have shrunk Thuban and add more speed through frequency increase. (headbang)
endofoff/
Similar threads
