GTX 1070 8gb review
 

[url]http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1070-8gb-pascal-performance,4585.html[/url]
[quote]
In every test we ran, GeForce GTX 1070 is as fast as, or faster than the $1100+ GeForce GTX Titan X.
[/quote]

Faster everywhere except for 64 bits double precision floating point of course.

sieving/trial factoring/factoring isn't using that at all.

Yet we do not have anything here on the forum that's really using doubles. I might have a shot at it some time from now though :)
Right now toying a tad to get sieving riesels to work at a GPU (starting with my GTX580 yet should scale upwards).

Isn't CUDA-LL using Doubles?

Yes, and LLRcuda. Primality testing, any GPU calculation using FFT, uses doubles.

It's not so efficiently using the double resources.

If card could deliver 6 Tflop then all sorts of factors will limit it to deliver say 400 gflop output as it probably is bandwidth limited (didn't investigate it).
Didn't investigate it but would suspect memory bandwidth larger impact than double resources - unless you have too few double resources...
What's it using effectively from a Titan X in double resources? Just 20% of it or so or 10%? Maybe less?

So that's not really a double precision resources dependant program. It's not a joke that program - it exists!

Thats something worth investigating once a forum member gets his/her hands on one (or a GTX1080) .

1000 series gamerscard have only 4 double precision execution units in each Multiprocessor array (SM) of 128 cuda cores.

So a 1070 with 15 SM's has just 60 double precision execution units.

[url]http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#maximize-instruction-throughput[/url]

Compare with oldie Tesla's that are for pickup at ebay for 100 dollar, they can deliver 650+ gflops.

At a Titan however, not to confuse with newer Titan X, 1/3 of all cuda cores made up its FP64 resources.

[QUOTE=VictordeHolland;438414]Thats something worth investigating once a forum member gets his/her hands on one (or a GTX1080) .[/QUOTE]

We do have stats from my 1080, though we are still waiting for NVIDIA to release the final fix that allows mfaktc to work properly.

How does the GTX 1060 compare since it is in more of my price range?

[url]http://www.newegg.com/Product/Product.aspx?Item=N82E16814125879[/url]

[QUOTE=RichD;438421]How does the GTX 1060 compare since it is in more of my price range?

[url]http://www.newegg.com/Product/Product.aspx?Item=N82E16814125879[/url][/QUOTE]
The GTX1060 should be comparable in performance or outperforming the GTX980 and AMD RX480, at least in games. And it does so using less power.

GTX1060: 1280 CUDA cores
GTX1070: 1920 CUDA cores
GTX1080: 2560 CUDA cores
All three cards run at similar clocks speeds. So the GTX1060 should be performing at around 2/3 of a GTX1070.

I hope the availability of the newest cards (GTX1080, GTX1070, GTX1060 and RX480s) is solved soon. They are either out of stock or priced ridiculously higher than their advertised launch price. But maybe that is a europe/Netherlands thing.

It would appear as either nVidia design is more power efficient or TSMC 16nm FINFET node is superior to GlobalFoundries 14nm.

I'm really torn about which card to get.

The 1060 uses 80% of the power of the RX 480 and performs better, especially OpenGL, which matters to me. That 20 watts per year will pay for the difference in price in a couple of years.

On the other hand, AMD is doing great work with their drivers, open-sourcing almost everything and getting them integrated into the kernel, and I want to support that.