Holy crap, if true then Nvidia really has weaved some magic here. Big Maxwell should be a monster!
Holy crap, if true then Nvidia really has weaved some magic here. Big Maxwell should be a monster!
It will probably happen after tessellation, but it's possible to perform some HSR prior to tessellation. If there's displacement mapping you probably need hints from software though. Or just let software do this level of HSR for you.
Blazkowicz
Legend
Less so for a render farm, esp. once 8GB version gets there.
Even physics meant for visual effects doesn't need double precision I think.
DuckThor Evil
Legend
If those numbers are legit, in that test it performs about equal to 780ti clock for clock or a tiny bit worse. Isn't that about in line with the expectations? GM204 isn't a small chip after all and the 750ti kind of points out the way already with regards to performance. The Maxwell tech does seem very strong, AMD is going to have its hands full.
Where do you get that from? It could very well be the case that the application states just random unit counts for one and as a second the GM204 seems to do quite well compared to a full blown GK110b in double precision tests at SiSoft.
In any case, if the chip theoretically has 16 clusters with 16 FP64 SPs (for each) as the application states, then it would be more like a 1:8 ratio.
Maxwell isn't nearly as dependent on ILP to achieve max performance. Each scheduler only has to worry about one ALU instruction per clock but retains the benefit of dual issue for load/store, SFU etc. It's a win-win. Couple that with the much bigger cache and it has a pretty good leg up on Kepler.
Let's wait for actual reviews before declaring victory though. The 980 still has to overcome a 50% bandwidth deficit to the 780 Ti.
AnarchX's post at the top of the page. It's just rumored right now of course, but if true it'd be highly disappointing for those who like to run DC GPGPU applications on our systems.
So there are people who need DP and need it to be fast but yet are unwilling to pay for the higher end compute focused cards? Who are those people?
The Titan Black should be right what they need...
That's what I was thinking. $1000 shouldn't be a big deal for anyone that "needs" fast DP.
I thought professionals buy Quadros/Teslas for that purpose. Outside of those (fair or not) what you pay is what you get for double precision. It's not like the 780ti scores all that much higher than a GM204 in those tests and I can imagine it's a fair bit higher than what GK104 GPUs would get.
From the Maxwell tuning guide:
http://docs.nvidia.com/cuda/maxwell-tuning-guide/#smm-latencies
Does anyone have hard numbers to back up those claims?
Eh, I should pay $3000+ for a firepro card why the fuck for exactly? It's the same god damn silicon as in the regular radeons. Same with nvidia's overpriced "higher end" junk, by the way.
Do you think Intel should deliberately gimp FPU performance in their core-series CPUs, forcing you to fork out a thousand bucks or more for a xeon 'if you need fast DP'? Maybe I shouldn't be giving you any ideas...
I don't know any of them personally, so I can't answer your question, but just for Folding@Home, they have over 170k users, providing nearly 40 PFLOPs dedicated to science and healing human beings. That's pretty serious stuff. For a lot of science stuff, single precision floats just don't cut it, and it simply isn't practical, reasonable or even affordable for institutions to build their own supercomputer clusters to crunch all their data. Hence, DC.
When it comes to segmentation via disabling random features, Intel is probably the worst offender.
PS: I think Folding@Home uses SP, but don't quote me on that.
It's not impossible to have separate silicon for bifurcation based on a single metric, in this case DP throughput.
However, if everything else is equal, you will stand a good chance of paying incrementally more for each, since there are now two different physical designs requiring implementation and manufacturing.
They tend to segment on a number of other axes, such as threading, security instructions, extensions, IO, and platform features for a specific die.
Intel tends to provide different die or architectures if the metric is floating point performance.
The way of the business is such that someone considers something "serious" when they are willing to pay for it. Other scenarios are "beggars can't be choosers".
Yes, I know. However, say, virtualization in a CPU for example is not comparable to a GPU's DP performance. But yeah, deliberately gimping hardware just to charge more for the ungimped version is crap, no matter who's doing it (and intel has even experimented with paid CPU "ungimp DLC", making them the absolute worst of the worst by quite a degree really.)
Isn't it comparable? I suspect more CPU users need virtualization than GPU users need (fast) DP.
But hyperthreading and Turbo boost are (for random Core i3/5 models).
For Sandy Bridge celerons it goes even further (no AES, disabled L3 cache slices)
Cheers
Because the supplier of that capability determined that's what it's worth.
Completely irrelevant. As a consumer you have no right to demand that a company sells you something cheaper. Especially when it comes to high tech gadgets. The fact that the silicon has the capability doesn't entitle you to it on the cheap.
Given that its their blood, sweat, tears and money that went into designing and producing those chips they can sell them for whatever they feel the market can bear.
If a big oil company or government research lab is willing to shell out big bucks for fast GPGPU DP then that's what it's worth.
