Haswell benchmarks show up

Must read and other references.
Dade
Posts: 206
Joined: Fri Dec 02, 2011 8:00 am

Re: Haswell benchmarks show up

Postby Dade » Thu Mar 21, 2013 8:31 am

Serendipity wrote:GPUs rock because:

A) they have massive driver problems


Uh, don't tell me, I have found problems in every single OpenCL platform I have used (AMD, NVIDIA, Intel, Apple). Soon, I'm going to try an ARM Mali T-604 GPU ... so I can add ARM to the list too :lol:

Serendipity wrote:B) you can´t fix a bug yourself if the bug is in the driver so it slows down your developing speed


I have recently worked with an AMD guy to fix a problem on Apple platform with AMD GPUs. I had to do a lot of work to isolate the problem and provide a good example but was a positive experiences: they fixed the issue in 2 days (the major problem now seems to have Apple to release the patch).

The lesson for me was that I should probably report bugs to vendors more often (and vendors should encourage developers more).

Serendipity wrote:C) a new generation might mean older software doesn´t run anymore and it sometimes takes months until a working version shows up (do I need to mention iRay...)


Every one, here, exercise their constitutional right to write their own renderers so it isn't a problem ;)

Dade
Posts: 206
Joined: Fri Dec 02, 2011 8:00 am

Re: Haswell benchmarks show up

Postby Dade » Thu Mar 21, 2013 8:39 am

hobold wrote:
Dade wrote:OpenCL on FPGA

Do you happen to be one of the people using that kind of setup? If so, are you allowed to talk about your experience with that hardware?


I'm just following the developments (there is a twitter channel, the name should be OpenCLonFPGA). Altera (http://www.altera.com/) has posted a lot of slides and documentations.

It may not be possible to port a complete renderer to FPGA but, for instance, the ray intersection code alone should be doable.

hobold wrote:(BTW, thanks for your efforts regarding OpenCL. So far, MandelGPU has been the most useful example code for my own attempts to learn and tinker with OpenCL.)


MandelGPU was the very first OpenCL program I wrote :D Just in case someone needs, I'm trying to pack my old OpenCL demos in a single repository and to update the code with some better practice: http://code.google.com/p/ocltoys/ (still a lot of unfinished stuff).

dbz
Posts: 46
Joined: Wed Jan 11, 2012 10:16 pm
Location: the Netherlands

Re: Haswell benchmarks show up

Postby dbz » Thu Mar 21, 2013 4:14 pm

Dade wrote:
dbz wrote:Regarding AVX, creating 8-ary bvh's might be interesting where one ray is intersected with 8 bounding boxes during intersection.


It is a features already available in Intel Embree and, according the authors, the speed up is quite small: something in the 10%-25% range.

Embree only supports binary and quad bvh's? I only see code for bvh2 and bvh4 in the rtcore directory and the README.txt says

--- Setting Spatial Index Structure ---

The ray tracing core in Embree supports a BVH with a branching factor
of 2 (BVH2) and branching factor of 4 (BVH4).

...

Possible acceleration structures are:
bvh2 (BVH with branching factor 2)
bvh4 (BVH with branching factor 4)


Nothing about 8-ary bvh's?

Dade
Posts: 206
Joined: Fri Dec 02, 2011 8:00 am

Re: Haswell benchmarks show up

Postby Dade » Fri Mar 22, 2013 8:23 am

dbz wrote:Embree only supports binary and quad bvh's? I only see code for bvh2 and bvh4 in the rtcore directory and the README.txt says


Sorry, it wasn't Embree: viewtopic.php?f=3&t=45&p=266&hilit=avx+embree#p257

dbz
Posts: 46
Joined: Wed Jan 11, 2012 10:16 pm
Location: the Netherlands

Re: Haswell benchmarks show up

Postby dbz » Fri Mar 22, 2013 10:06 am

Dade wrote:
dbz wrote:Embree only supports binary and quad bvh's? I only see code for bvh2 and bvh4 in the rtcore directory and the README.txt says


Sorry, it wasn't Embree: viewtopic.php?f=3&t=45&p=266&hilit=avx+embree#p257

Thanks for making this clear. I guess AVX2, which Haswell supports, could deal with the limitations AVX1 has, which are addressed in that topic. So there is still a potential for some speedup with Haswell when using 8-ary bvh's.

dbz
Posts: 46
Joined: Wed Jan 11, 2012 10:16 pm
Location: the Netherlands

Re: Haswell benchmarks show up

Postby dbz » Sat Jul 13, 2013 8:58 am

Interesting Haswell review from the German website computerbase: "Das ist gut zu sehen in Cinebench x-CPU, in dem ein 4770K acht Prozent schneller ist als ein 3770K, das Gesamtsystem aber auch 13 Prozent mehr Leistung benötigt" . Haswell 8% faster on Cinebench but draws 13% more power. Really discouraging the complete lack of progress in cpu development.

Zelcious
Posts: 42
Joined: Mon Jul 23, 2012 11:05 am

Re: Haswell benchmarks show up

Postby Zelcious » Sat Jul 13, 2013 2:47 pm

I got my new asus gtx 780 with directcu ii cooler yesterday and with a heavy overclock it's twice as fast as my gtx 480 with my own renderer. So far no tweaking have been able to remedy that. Quite a bit of disappointment actually.

Dade
Posts: 206
Joined: Fri Dec 02, 2011 8:00 am

Re: Haswell benchmarks show up

Postby Dade » Mon Jul 15, 2013 7:18 am

Zelcious wrote:I got my new asus gtx 780 with directcu ii cooler yesterday and with a heavy overclock it's twice as fast as my gtx 480 with my own renderer. So far no tweaking have been able to remedy that. Quite a bit of disappointment actually.


It probably fair to consider the 480 and 580 a single GPU generation, same for the 680 and the 780. You have to factor also that Kepler doesn't really shine when compared to Fermi architecture.

My AMD HD7970 is also "only" 2 time faster than my old HD5870 but, again, the HD6xxx was too similar to the HD5xxx (4-way VLIW Vs. 5-way VLIW) and built on the same manufacture density (i.e. 40nm).

I can not remember the last time I have seen a CPU 2 times faster than 1 or 2 older generation CPUs. Today, it seems something that could be achieved only by increasing the number of cores but it looks like no one want to take the route (we still have CPUs with only 4 cores :| ).

To be fair with CPUs, it looks like vendor are adding a GPU to their chip instead of adding more cores (like with Haswell).

ingenious
Posts: 279
Joined: Mon Nov 28, 2011 11:11 pm
Location: London, UK
Contact:

Re: Haswell benchmarks show up

Postby ingenious » Mon Jul 15, 2013 10:50 am

Maybe it's time to stop slacking off and make them algorithms a bit faster? :mrgreen:
Image Click here. You'll thank me later.

dr_eck
Posts: 45
Joined: Mon Dec 05, 2011 7:35 pm

Re: Haswell benchmarks show up

Postby dr_eck » Mon Jul 15, 2013 6:44 pm

@ingenious: Are you volunteering to take the lead on that? :lol:


Return to “Links & Papers”

Who is online

Users browsing this forum: No registered users and 1 guest