NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
+ 2 times the cores.
PSU-failure
Newcomer
And on which basis? "chip for chip" or "flop for flop"?
If 150-400% is just the raw throughput and not respective to the theorical throughput, that's not exactly exciting as it's effectively about 75-200% as efficient.
200% is quite low for a caching bonus and remember GF100 is huge (>2x normalized wrt the processes used) and power hungry (1.25-1.5x depending on the board) compared to GT200, if that's "better efficiency", I don't know what would be a decrease.
I'm yet to find any use for this chip, let's hope process refinements/change will be able to reduce its static leakage substantially as it seems to be the biggest issue.
If a program runs 4 times as fast on Fermi than on a GT200b, then I'd say that > 2X size and 1.5X power is pretty reasonable. You don't really get that in the CPU world.
PSU-failure
Newcomer
At the same time, in the CPU world we're not supposed to work with massively parallel workloads... there's absolutely no point in using many complex OoO cores for that.
Even if we do use multiple complex cores, they're often clustered in a single die, thus removing some redundant and useless bits, which translates into better efficiency compared to multiple "complete" CPUs.
About GF100's efficiency, any workload slower "flop for flop" than its predecessor would just be very bad, even if it's twice faster in other areas since overall power efficiency would end up similar, not significantly better.
It's the fastest single GPU ever made, lets get back to reality shall we?
nApoleon at chiphell just posted
....interesting the die size at below 300mm2 seems to contradict cfcnc and edison previously saying it was somewhere in the 65nm G92 region or a bit above(ie 330-350mm2)
Re the rose colored glasses comment previously: that was in relation to the reason the chip is initially going up against the 5830 in cutdown form is to protect sales of GTX470...would have to be on drugs to believe that...the chip is initally starting in crippled form principally to get inital yields up to a money making level, not enough full versions can be made yet.
....interesting the die size at below 300mm2 seems to contradict cfcnc and edison previously saying it was somewhere in the 65nm G92 region or a bit above(ie 330-350mm2)
Re the rose colored glasses comment previously: that was in relation to the reason the chip is initially going up against the 5830 in cutdown form is to protect sales of GTX470...would have to be on drugs to believe that...the chip is initally starting in crippled form principally to get inital yields up to a money making level, not enough full versions can be made yet.
That doesn't mean much per se. +15% of performance, +40% price and +50% power consumption aren't very convincing numbers. I'm not saying GF104 can't change it (I hope it will), but lots of people aren't willing to accept small performance advantage at all costs. That era is over for many years...
High-end enthusiast market is very small and populated by people caring for one thing:
Performance.
A lot (if not must) of people bashing Fermi don't run a 5870, they run a lesser card. Must don't even run a GTX28x/4890.
Which is like watching the "Morris Minor Owners Club" complain that Lamborghini's are noisy, not fuel efficient and way to pricy.
I only notice the +15% performance win.
Only reason I still have a GTX285 is because I am buying a house and moving in with the GF & bonusdaughter this summer.
After that I plan to upgrade my rig like this:
i7 920 -> i7 980X
GTX285 -> GTX480
2x80GB RAID 0 -> X25-M SSD
Air cooling -> water cooling.
Performance being guideline here, not fluffy none-enthusiast concerns
But what about Eastern Islands?
Jawed
Legend
If the DP is implemented the way Aaron suggested a while back then it could only make a notable difference to die size if the int MUL is also deleted. Doing that would hurt compute stuff.
The caches are too small to make any notable difference.
Jawed
Jawed
Legend
In theory it will end-up competing with HD6xxx... OK, so it gets ~ one quarter (2 if it's really lucky) in the sun...
Computex will bring us GF104 and HD6000 (although the latter not a launch), actually.. more people here know about the AMD lineup than about GF104/6/8.
GTX465 die:
Performance:
Farcry 2 1920x1200 8x AA Max:54fps
Crysis Warhead 1920x1200 VHQ 4x AA:19.9fps
3Dmark Vantage:X5741
Heaven Benchmark 2.0 1920x1200 4xAA:18.8fps
Source:
http://translate.google.com/translate?js=y&prev=_t&hl=en&ie=UTF-8&layout=1&eotf=1&u=http%3A%2F%2Fwww.enet.com.cn%2Farticle%2F2010%2F0512%2FA20100512652722.shtml&sl=zh-CN&tl=en
Performance:
Farcry 2 1920x1200 8x AA Max:54fps
Crysis Warhead 1920x1200 VHQ 4x AA:19.9fps
3Dmark Vantage:X5741
Heaven Benchmark 2.0 1920x1200 4xAA:18.8fps
Source:
http://translate.google.com/translate?js=y&prev=_t&hl=en&ie=UTF-8&layout=1&eotf=1&u=http%3A%2F%2Fwww.enet.com.cn%2Farticle%2F2010%2F0512%2FA20100512652722.shtml&sl=zh-CN&tl=en
Last edited by a moderator:
Which would fit nicely with the goals of the GF104, IMHO.
I beg to differ about the GT212+DX11 part. For me the biggest difference between GT200/G80/G92 and GF100 is not really the cache structure. Its the chip strucutre with its "almost own GPU clusters" (im aware some people here dont believe the GPCs and all that jazz...) each with its own "Polymorph Engine". Now, if GF104 doesnt mantain this structure you are absolutely right about GT212+DX11...
Similar threads
- Deleted member 2197
- Architecture and Products
