Based on measurements, then confirmed by Nvidia. However they wouldn't say what exact part of the process is the limitation (vertex fetch ?). Note that this high throughput can only be achieved when a single vertex has to be fetched per primitive. If you need to fetch 2 vertices, rate is one primitive per SMM every 6 clocks etc.
Regarding the 750 non-Ti, I did measure culling rate as well, but I didn't show it in the graph.
Well, the Cuda Programming Guide's been updated, with information about Maxwell. http://docs.nvidia.com/cuda/index.html
And....
1/128 double precision throughput. REALLY Nvidia??
And....
1/128 double precision throughput. REALLY Nvidia??
OpenGL guy
Veteran
GM107 has 1/32nd rate, not sure where you are seeing 1/128.
LiXiangyang
Newcomer
I believe its a tempo, based on my measurements.
However from the pg, its rather disapointing that Maxwell have not give any boost to the much needed general integer multiple operations, instead, the integer multipity/fma performance become even weaker than Kepler's.
However from the pg, its rather disapointing that Maxwell have not give any boost to the much needed general integer multiple operations, instead, the integer multipity/fma performance become even weaker than Kepler's.
That table of sm_50 throughputs has many errors. I don't trust it at all.
In addition to the wrong fp64 throughput, the SIMD video instructions are no longer native operations sm_50, yet that table shows them not only still there, but at double throughput to Kepler. This is wrong.
The new Maxwell Tuning Guide says that integer multiplies and compares are faster than Kepler. The new throughput table says they're slower than Kepler.. multiple instructions even. Benchmarks indicate they are probably the same per SP.
iMacmatician
Regular
From PC Tuning: "No GPU chip 20nm this year doesn't work, new graphics, but yes!" (original):
.png)
Also, 3DCenter gives probable release timeframes of the GM204 as December 2014, the GM206 as December 2014 to January 2015, and the GM200 as H1 2015 (ideally "as early as February 2015").
Also, 3DCenter gives probable release timeframes of the GM204 as December 2014, the GM206 as December 2014 to January 2015, and the GM200 as H1 2015 (ideally "as early as February 2015").
My Microsoft translated Czech leaves much to be desired, but are they first blowing their own trumpet about the whole world linking to their article last week? You know, the one about Maxwell being 20nm. Which they now contradict 100% by saying it will be 28nm?
One thing is sure: they'll be able to write after the fact that, at some point, they wrote the truth. Journalism...
One thing is sure: they'll be able to write after the fact that, at some point, they wrote the truth. Journalism...
Rumor mongering is no more journalism than astrology is science.
It starts with a shred of information which gets enriched with speculation and ends up being plagiarized in many colourful forms. Funniest of them all being an April tape out and a June release. I call it SOLS speed of light silicon.
So it looks like Erinyes' sources/insiders are correct. The GM204 (and 206?) were taped out a few weeks ago and the GM200 is going to be taped out soon and they will be all on 28nm.
I don't think they'll be a 20nm refresh of Maxwell and Pascal 16nm FF will be out in 2016 with a GP107 as a 'test run' chip.
I don't think they'll be a 20nm refresh of Maxwell and Pascal 16nm FF will be out in 2016 with a GP107 as a 'test run' chip.
UniversalTruth
Veteran
So, will it be in August/September ("before autumn") or in December (winter)?
UniversalTruth
Veteran
What happens during this so long period?
It lies in the sun to dry out and its engineers next to it to get a tan
http://en.wikipedia.org/wiki/Integrated_circuit_development
Yes. My apologies.
All I heard was that GM204/206 are being targeted for a release in the holiday season, with GM206 being slightly behind GM204. Holiday season usually implies Q4 so it should be sometime then.
Depends on your definition of tape out but yea I think thats about right.
It lies in the sun to dry out and its engineers next to it to get a tan
http://en.wikipedia.org/wiki/Integrated_circuit_development
UniversalTruth
Veteran
They will send the design on digital carrier to the fab and then the first photomasks will be created for subsequent production.
A check google and on wiki says it is Late November Until Christmas Eve/Day so I guess they'll release by then. Also will the GM200 be taped out in June and likely release in Feburary since it is around 8 months. I hope I can get a GM200 and play the Witcher 3 @ 2560x1440 at max (minus MSAA, hopefully SMAA starts to pick up)
.
