Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Here comes Maxwell... on 20 nm
http://www.tomshardware.com/news/google-pixel-c-tablet,30202.html
Nvidia has now ported their Maxwell architecture to a 20 nm chip for the Tegra X1, as found in the Google Pixel C. That would be an integrated GPU with two Maxwell SMMs, alongside a boatload of ARM cores and various other SoC stuff.
TSMC's 20 nm process node, like all other 20 nm process nodes available, is built for low power parts. But a tablet needs a low power part, and Maxwell is the most energy-efficient architecture among high performance ones, so it's a reasonable fit.
But this also means that Nvidia has put in the work to port Maxwell to TSMC's 20 nm process node. That raises the possibility of launching discrete GPUs of 20 nm Maxwell parts. Will they?
You can make a high performance chip on a process node optimized for low power, but it's not guaranteed to work very well. Would a 200 W Maxwell card on 20 nm be better than one on a 28 nm process node designed for higher performance? It could easily be worse on 20 nm rather than better, and even if it is better, that's no guarantee that it's better by enough to justify the cost.
What I think is the more interesting possibility is a lower power 20 nm Maxwell card for laptops. If performance that used to take 70 W can now be had for 50 W, that's pretty nifty. Will 20 nm make that possible? Even if it does, will Nvidia do it? It's not like they need something new to catch up to AMD; Maxwell is already ahead of AMD on energy efficiency, which is a major reason why so few laptops have a discrete AMD GPU.
I'm not predicting that Nvidia will make such a card. But it's an interesting possibility to consider.
For what it's worth, AMD will not do that. AMD has publicly announced that chips that were planned for 20 nm have all been moved to a FinFET process node, presumably 14 or 16 nm, depending on the foundry. The considerations that led AMD to shy away from 20 nm for discrete GPUs could lead Nvidia to do the same. But that includes AMD's tablet chips, and Nvidia just delivered a 20 nm tablet chip.
Nvidia has now ported their Maxwell architecture to a 20 nm chip for the Tegra X1, as found in the Google Pixel C. That would be an integrated GPU with two Maxwell SMMs, alongside a boatload of ARM cores and various other SoC stuff.
TSMC's 20 nm process node, like all other 20 nm process nodes available, is built for low power parts. But a tablet needs a low power part, and Maxwell is the most energy-efficient architecture among high performance ones, so it's a reasonable fit.
But this also means that Nvidia has put in the work to port Maxwell to TSMC's 20 nm process node. That raises the possibility of launching discrete GPUs of 20 nm Maxwell parts. Will they?
You can make a high performance chip on a process node optimized for low power, but it's not guaranteed to work very well. Would a 200 W Maxwell card on 20 nm be better than one on a 28 nm process node designed for higher performance? It could easily be worse on 20 nm rather than better, and even if it is better, that's no guarantee that it's better by enough to justify the cost.
What I think is the more interesting possibility is a lower power 20 nm Maxwell card for laptops. If performance that used to take 70 W can now be had for 50 W, that's pretty nifty. Will 20 nm make that possible? Even if it does, will Nvidia do it? It's not like they need something new to catch up to AMD; Maxwell is already ahead of AMD on energy efficiency, which is a major reason why so few laptops have a discrete AMD GPU.
I'm not predicting that Nvidia will make such a card. But it's an interesting possibility to consider.
For what it's worth, AMD will not do that. AMD has publicly announced that chips that were planned for 20 nm have all been moved to a FinFET process node, presumably 14 or 16 nm, depending on the foundry. The considerations that led AMD to shy away from 20 nm for discrete GPUs could lead Nvidia to do the same. But that includes AMD's tablet chips, and Nvidia just delivered a 20 nm tablet chip.
Comments
Maxwell so far turns out worse in peformance than last generation AMD (refreshed to 3xxx). And, with few dx12 test that are available, AMD has decent advantage over NVidia. If NVidia wants competitive products it will need to look at performance again. Power is least of worries when you buy 150-300$ GPU.
So yeah, NV doesnt have much wiggle room.
That being said - we already have Maxwell on 28nm, although I seem to remember that it was originally destined for 20nm when it was first being rumored about. Not to say it's never been done before, but I don't anticipate another generation on Maxwell in the discrete market at different lithography unless Pascal has problems or we see HBM get pushed back farther.
Kind of interesting - it appears that Maxwell specifices 128 "cuda cores" per SMM for Maxwell, and 4 (initially 5) SMMs to make a single GPC (graphic processing cluster) - I don't know how flexible nVidia's GPC design is for SMM count, so it looks like they may have done some redesign right there to get that to work out. That makes me think we are looking at a push to get this into the SOC market, and not so much that we'll see another full generation on 20nm.
Also - nVidia does have some rumors for Pascal on 16nm FinFET, with HBM to get on that train, coming early-mid 2016. With that just around the corner, unless they run into design or supply issues, I don't see nVidia pushing a stop-gap generation of cards -- although that has happened before. The 500 series was basically a "Fixed Fermi" because the jump to 28nm Kepler was running behind and they couldn't afford to sit on the 400-series for too long.
Sure, there are still some niche applications that ~need~ a dGPU, and there are still a whole lot of scenarios where a dGPU would do better than an iGPU, but for 99% of all laptop users (making up random stastics), the iGPU in whatever CPU they get is more than enough, and the battery, weight and cost tradeoffs for a dGPU just don't make a lot of sense any longer, and most of that 1% that do need a dGPU could probably just perform that work on a desktop as effectively anyway.
When your CPU is pulling 35W with an integrated GPU, an additional 30W GPU is a lot of extra load to power and cool in a laptop.
Mobiles are already SOCs, and a lot of laptop-type devices are heading that way. Discrete components are largely being relegated to desktop and server class hardware that needs upgradability and interchangeability. Mobiles and Laptops have mostly always been upgraded as an entire device anyway.
Does the x1 run cooler? Like a lot cooler? Shield tablet was almost there for me. good media device. great speakers. good size. Looked good. To dam hot. Oh amazing price point for a tablet to! That can't be understated in the least. $299 -$399 for a shield tablet...
http://wccftech.com/nvidia-devs-computegraphics-toggle-heavyweight-switch/
Maxwell is history.
Only question is have they screwed up with new chip also, as that would be pretty pretty bad