LiXiangyang
Newcomer
I could careless about their L1 cache size until they increase the bandwidth of it significantly.
I think you meant to say "I couldn't care less" what you said means the opposite.
LiXiangyang
Newcomer
Well, I mean the bandwidth.
The bandwidth of Kepler's L1 is less than that of Fermi, if 64-bit access is disabled, and will be slightly more than Fermi if 64-bit access is enabled.
I hope Maxwell is not like Kepler in this regard.
The bandwidth of Kepler's L1 is less than that of Fermi, if 64-bit access is disabled, and will be slightly more than Fermi if 64-bit access is enabled.
I hope Maxwell is not like Kepler in this regard.
Looks like UVM-lite comes with CUDA 6.
http://www.anandtech.com/show/7515/nvidia-announces-cuda-6-unified-memory-for-cuda
http://www.anandtech.com/show/7515/nvidia-announces-cuda-6-unified-memory-for-cuda
Just the kind of code I want to see, free use of C-style pointers w/o contexts is a lovely thing and probably much faster too, although the cudaDeviceSynchronize() line gives me some reservation for nVidia's architecture vs. what's probably possible w/ Kaveri for a single threaded program.
Last edited by a moderator:
It's not going to be any faster. This is just a code cleanup exercise. nVidia is still doing the same stuff behind the curtain that developers had to do themselves previously.
It should be faster on Kaveri though as it shares the physical memory space.
LiXiangyang
Newcomer
Actually when talking about the new memory feature, NV tell it will be actually slower than programmer-controlled way of handling memory when they get it right.
NV tell the audience the NVM is productivity feature, not a performance one.
Its not hard to imagine a programmer who know what he is doing will do better, or even far better, than compiler at managing memory resources, otherwise there wont be so many tricks/algorthims on how to increase cache-hit rates.
NV tell the audience the NVM is productivity feature, not a performance one.
Its not hard to imagine a programmer who know what he is doing will do better, or even far better, than compiler at managing memory resources, otherwise there wont be so many tricks/algorthims on how to increase cache-hit rates.
http://www.fudzilla.com/home/item/33362-maxwell-geforce-cards-shipping-by-march-2014
Maxwell might launch in February or March, after Chinese New Year Jan 31st which is 1 week public holiday in China and Taiwan.
Maxwell might launch in February or March, after Chinese New Year Jan 31st which is 1 week public holiday in China and Taiwan.
Hopefully, it's a 20nm part, the timing could be aligned.
I don't know how reliable this is:
http://focustaiwan.tw/news/atod/201312050035.aspx
I wonder if we'll see something similar to a Tahiti/Kepler launch where the high cost, but low volume parts trickled out first on the new node, with the volume mid tier parts not coming until a quarter or so later.
I don't know how reliable this is:
http://focustaiwan.tw/news/atod/201312050035.aspx
I wonder if we'll see something similar to a Tahiti/Kepler launch where the high cost, but low volume parts trickled out first on the new node, with the volume mid tier parts not coming until a quarter or so later.
A1xLLcqAgt0qc2RyMz0y
Veteran
I noticed that AMD is missing from that list of industry giants.
Does that mean AMD will be late to the 20nm party?
A1xLLcqAgt0qc2RyMz0y
Veteran
I believe so as Nvidia needs a new high-end to replace the aging GK110 for their supercomputer push and it takes time (4-6 months I think) for validation of said parts. If Nvidia waits to produce the GM110 then that would put their release into 2015.
iMacmatician
Regular
There were some reports of GM107 and GM108 ES parts a few months ago, although those parts are notebook ones. I wonder if the retail parts referenced in this rumor are desktop GM107/GM108 or another chip.
"booked up with orders from industry giants including"
It's not a full list, but Nvidia to be in that list and AMD to be excluded is not a good sign.
AMD may very we'll be going with GF instead of TSMC as GF's 20nm process looks similar to TSMC's.
DavidGraham
Veteran
Don't think NV is launching Maxwell that soon, that would be stupid.
It might be the point where AMD makes the jump to GF's 20nm since they're going Gate Last for 20nm. Also is TSMC's 20nm SOC and GF's 20nm LPM a good fit for high performance GPU?
Unknown Soldier
Veteran
Yeah, would be a smart choice for Nvidia to get Maxwell out asap to extend the lead that the 780 Ti has. AMD won the console wars so Nvidia might as well move to kill the PC war. Ok it won't kill it but you get my meaning.
