NVIDIA Maxwell Speculation Thread

10nm FinFET 2016 ready for what? Mass production? Customer Sampling? Process Demonstration via SRAM-Cells? Serious question, I really don't know.
 
AFAICS FDSOI has the best chance to compete with Intel ... unlike non-Intel finfets products have actually shipped with FDSOI.
 
and will be a node behind in area scaling over intels full front and backend 10nm process

http://www.extremetech.com/computing/171477-intels-14nm-milkshake-its-better-than-yours/2
They'll likely be closer, Intel's earlier backing off of scaling metal pitch aside, preliminary numbers from some of the foundries indicates at least some density gains with FinFET, even if the metal pitch doesn't improve.
Intel had to wring density improvements out of slowly scaling metal pitch at one point as well.

AFAICS FDSOI has the best chance to compete with Intel ... unlike non-Intel finfets products have actually shipped with FDSOI.

For some reason, those in the position of deciding what the foundries will use haven't shown much public commitment to it. What gives?
 
The M in GM107 sure doesn't stand for Kepler.
Introducing the Mepler architecture?

There's been quite a bit of talk lately saying that the upcoming GM107 part is a Kepler refresh (so it's like GM107 is "GK207" or something similar). Let's assume that the GM107 codename is accurate, and it is the same chip that is rumored to show up later this month. Then I don't see how it wouldn't be based on the Maxwell architecture. Sure, NVIDIA's given new model numbers to chips based on older architectures, but when have they given a chip of an old architecture (Kepler) a codename of the form that would (presumably) be used for chips of a new architecture (GMxxx)?
 
Nvidia GTX 750 & 750 Ti Benchmarks

750gpuz.jpg


750
7503dmark.jpg

750Ti
750ti3dmark.jpg
 

Quite strange that a computer shop is jumping the gun like that. So it really seems that neither card needs a 6pin power connector. I would not call having more than 1k more 3DMark score "barely" exceeding the GTX650Ti. Especially not when the GTX750 non-Ti does it as well. In percentage would be ~24% faster. Still, its 3DMark, not exactly reliable to take conclusions, I guess....
 
Another interesting feature is neither card has a power connector drawing all the required power from the PCI-E Slot.

Thats impressive power efficiency, considering the GTX 650 Ti is 110W and these are 75W powered off the PCIe bus only. Hopefully no one thinks this is a Kepler, there's no way Kepler can even hope to be this power efficient.

And this is on the same 28nm process, not even any die shrinks involved yet.
 
According to the link, both cards are factory OCed but neither card has a power connector… interesting. I wonder how much they were OCed.
 
Thats impressive power efficiency, considering the GTX 650 Ti is 110W and these are 75W powered off the PCIe bus only. Hopefully no one thinks this is a Kepler, there's no way Kepler can even hope to be this power efficient.

And this is on the same 28nm process, not even any die shrinks involved yet.

Just to be clear, I don't think this is Kepler, but mobile Kepler cards do reach similar power-efficiency levels.
 
Just to be clear, I don't think this is Kepler, but mobile Kepler cards do reach similar power-efficiency levels.

When compared to GK107's fastest product (same hierarchy) - the gtx 650, GM107's 3dmark scores are 85% higher, supposedly within a very similar power envelope. Die size is up a little over 30%, but for being on the same fundamental node process, that is a massive improvement.

I don't know if games will scale that well though, the bandwidth is just so low. Early rumors said an integrated Denver core is somehow helping to alleviate bandwidth constraints, but the past several pages of comments seem to doubt a CPU core is integrated, let alone able to function in such a manner.
 
Back
Top