NVIDIA Maxwell Speculation Thread

Why not? It is the most important performance segment for them (mobile/OEM and large numbers) and most other segments got quite recent updates (GTX 760, 770, 780 (Ti)).
 
A GM107 being somewhat slower than a GK106 salvage part isn't much of a surprise.

GTX 660 is not a salvage part, it's fully enabled GK106, 960 CUDA cores and 192bit memory bus.

http://www.anandtech.com/show/6276/nvidia-geforce-gtx-660-review-gk106-rounds-out-the-kepler-family

Diving right into the guts of things, the GeForce GTX 660 will be utilizing a fully enabled GK106 GPU. A fully enabled GK106 in turn is composed of 5 SMXes – arranged in an asymmetric 3 GPC configuration – along with 24 ROPs, 3 64bit memory controllers, and 384KB of L2 cache.
 
Why not? It is the most important performance segment for them (mobile/OEM and large numbers) and most other segments got quite recent updates (GTX 760, 770, 780 (Ti)).

Perhaps they are not satisfied with the overall performance of the chip on the new architecture, or that the new architecture doesn't offer the 'promised' performance per watt improvements

Even tho they claim:

The rumor has it that GTX 750 Ti is supposedly the first Maxwell based video card. If those benchmark results are even remotely true, then first Maxwell GPUs won’t be revolutionary in terms of performance, but that’s nothing surprising

Go figure
 
This is dumb. I rather think this 750ti is a GK106 part on a 128bit bus.

It is also consistent to expect the GM107 to show in laptops first, where the power improvements are needed and better justify the high cost. Not throwing it out right away, stealthily in the low margin high volume desktop arena.

It could well show only in laptops even, we have a precedent, that was GF117. That one was arguably low volume, "risk production" or "pipe cleaner" as their first 28nm GPU.
(/Yes there's far less "process risk" from 28nm now, still a new GPU is a bit involved. That makes me want to delete the above, lol. Nvidia does introduce new GPU revisions or architectural updates on old model names too, e.g. 8400GS and GT640)
 
Last edited by a moderator:
AMD started GCN 1.1 + TruAudio with Bonaire HD7790 - a card that is actually a competitor to GTX 650 Ti in price and performance.

Maxwell doesn't have to differ that much from Kepler. It could perfectly be something similar to GCN 1.0 -> 1.1.
 
Why not? It is the most important performance segment for them (mobile/OEM and large numbers) and most other segments got quite recent updates (GTX 760, 770, 780 (Ti)).
I'm honestly not surprised that the first desktop Maxwell part is rumored to slot in the space directly below the existing 700 series desktop GPUs. I can see a cut-down part or two soon afterwards to replace the 650 and 650 Ti.

If the 750 Ti uses the GM107 chip then that's a pretty decent performance jump over the GK107 given that they are likely to be on the same process (although there's the power consumption factor).
 
Maxwell doesn't have to differ that much from Kepler. It could perfectly be something similar to GCN 1.0 -> 1.1.

That would be a pity.. Actually I haven't followed the speculations for GM up till now (if there was anything to follow).

What remained stuck in mind is some slide dated ages ago when Dally came to nV. Back then, he emphasized the perpetual need for increased memory bandwidth available to top end GPUs. And somehow we were then supposed to believe that we will see this focus come to fruition around maxwell time. Guess it's all gone to dust with the prioritizing of mobility over anything that seems to have infected nV as well. And guess the 'maxwell' back then has little in common with what will be launched the following months.





Well, at least on the CPU front i'm hoping that Skylake will be interesting. Gotta have at least one future architecture to be excited about while waiting :p
 
http://www.xbitlabs.com/news/graphi...well_Architecture_Will_Break_New_Grounds.html

“Number one for Maxwell, that is likely something that we are doing that breaks new ground in visual capability, something that is even more beautiful. […] Number two, it is likely that Maxwell breaks new ground in programmability, ease of programmability, because we want to expand the general purpose nature of the processor without sacrificing its speedup relative to a microprocessor. […] The last thing, the energy efficiency of Maxwell, it is going to crush Kepler. […] We know exactly how to measure it now and we know what it means to be good,” said Jen-Hsun Huang, chief executive officer of Nvidia, during a Q&A session at the company’s investor day conference.

www.youtube.com/watch?v=iq202LKkeHI#t=67

Perf per watt equals perf

See from 1:07 to 1:40, it makes a lot of sense, you can see why Intel and Nvidia is fanatically increasing their power efficiency. Stop complaining about mobile, mobile focus improves performance. If you go crazy without caring about power usage like what Sunnyvale did, you get Bulldozer & Hawaii.
 
Last edited by a moderator:
I think the more valid comparisons should be to the chip this is the successor to: GK107. 10-15% slower than the GK106 gtx 660 on the same node puts this new chip in a pretty good light when looking at the performance of the GK107 based gtx 650. Once again, SAME NODE. The big question is die size. If Nvidia was able to accomplish this kind of performance increase over GK107 without a huge increase in die size, then Maxwell will shape up to be a beast with a node shrink.

GK107 gtx 650 vs. GK106 gtx 660: http://www.anandtech.com/bench/product/681?vs=660
 
If GM107 is 28nm it would make perfect sense to launch it there as high-volume desktop 750ti (and below). Wouldn't rule out a GK106 rebrand either though..

The numbers are a little weird.. What's is so special about Valley compared to the rest? I don't think it's that tesselation/GPC limited.. Should probably compare with 650ti and 650ti boost...

Stop complaining about mobile, mobile focus improves performance
Why aren't we all running ARMs in our desktops then? Must have the best performance... :rolleyes:
 
If GM107 is 28nm it would make perfect sense to launch it there as high-volume desktop 750ti (and below). Wouldn't rule out a GK106 rebrand either though..

The numbers are a little weird.. What's is so special about Valley compared to the rest? I don't think it's that tesselation/GPC limited.. Should probably compare with 650ti and 650ti boost...

I'm liking the idea more now. GK107 is too old and slow, it gets phased out so now the GM107 replaces it promptly. Even the GK106 is being phased out or somewhat demoted, the GTX 760 replaced it, 650 Ti Boost doesn't exist and the 660 gets cheaper.
Eventually if GK106 and GK107 slowly go away, there won't be Kepler GPUs with less than 512K L2.

The "old" range was GT218 < GF119 < GF108 < GK107 < GK106 < GK104 < GK110
newer range would be GT218 < GF119 < GK208 < GM107 < GK104 < GK110

(GK106 ommitted because I want to make a point)
 
Perhaps they are not satisfied with the overall performance of the chip on the new architecture, or that the new architecture doesn't offer the 'promised' performance per watt improvements

Even tho they claim:

Go figure

I figure that it is on 28nm, and if it uses about the same power as the 650 Ti, then 30% better perf/W would quite impressive.
 
I figure that it is on 28nm, and if it uses about the same power as the 650 Ti, then 30% better perf/W would quite impressive.

The " rumored " performance, if true, place it at 650TI performance level.. im sorry but i dont where the 30% perf / watt is if the power is about the same. Its where my problem is with this leaked GM107 performace.. why outside need a pipecleaner (for new node ), you want release a gpu on new arch for get similar perf of the 650TI, call it it 750TI and so place it in under the 760.. ( yes i know, called "107", it should replace the GK107, but before we get the full spec, ... nvidia lately like to play on the "architecture code" (( it could never exist a gm 106--- GM104 and instead GM114 etc ).
 
Last edited by a moderator:
The " rumored " performance, if true, place it at 650TI performance level.. im sorry but i dont where the 30% perf / watt is if the power is about the same. Its where my problem is with this leaked GM107 performace.. why outside need a pipecleaner (for new node ), you want release a gpu on new arch for get similar perf of the 650TI, call it it 750TI and so place it in under the 760.. ( yes i know, called "107", it should replace the GK107, but before we get the full spec, ... nvidia lately like to play on the "architecture code" (( it could never exist a gm 106--- GM104 and instead GM114 etc ).

No, the rumored performance places it at GTX 650 Ti + 30%:
Average the performance increase of the 660 over the 750 Ti = 18.5%
660 is about 56% faster than the 650 Ti. Thus the 750 Ti should be around 30% faster than the 650 Ti on average.

I know you don't like Nvidia, but at least do us the courtesy and do your math next time before complaining and being pessimistic.
 
Back
Top