Haswell benchmarks show up back

(L) [2013/03/18] [ost by dbz] [Haswell benchmarks show up] Wayback!

Haswell benchmarked on [LINK http://www.tomshardware.com/reviews/core-i7-4770k-haswell-performance,3461-4.html Cycles] and [LINK http://www.tomshardware.com/reviews/core-i7-4770k-haswell-performance,3461-5.html 3ds Max]. Hardly any speedup over Ivy Bridge and definitively not worth upgrading. Remarkable that cpu performance seems to have basically flatted out from one generation to the next.

(L) [2013/03/18] [ost by stefan] [Haswell benchmarks show up] Wayback!

Performance should improve when applications make use of the FMA instruction. It did make a difference on the PPC. Also, depending on what their benchmark scenes were, the bottleneck could have been entirely in memory bandwidth/cache and cover up any arithmetic improvements.

(L) [2013/03/18] [ost by graphicsMan] [Haswell benchmarks show up] Wayback!

Yeah, I found that on Interlagos (AMD 16-core) FMA made about a 10% difference, which is pretty impressive when all the work required is to change a few compiler flags.  AMD has FMA4 though, which is a little easier for the compiler to optimize with than the FMA3 that Intel has.  Should be interesting to see what happens.  Also, are these apps optimized to use AVX?

(L) [2013/03/18] [ost by dbz] [Haswell benchmarks show up] Wayback!

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

(L) [2013/03/19] [ost by Dade] [Haswell benchmarks show up] Wayback!

>> 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.

(L) [2013/03/20] [ost by hobold] [Haswell benchmarks show up] Wayback!

>> Dade wrote:
the speed up is quite small: something in the 10%-25% range.
Please do not take this as an attack; and not as directed specifically at you. I just want to draw everybody's attention to the awful truth that ...


... the time of exponential performance improvements from new processor generations is over. Moore's Law will continue to be true to its letter for another while: transistor counts per chip area will continue to double. But the three main corollaries, which used to be true in the past:

1. chips will run at higher frequencies
2. chips will be cheaper
3. chips will consume less energy

will no longer follow from a new generation of fabrication technology. Those trends lost steam ever since the Pentium IV era.


We are now in an era of slow linear growth. An improvement of 20% is actually very significant these days. GPUs seemed to withstand this slowdown a bit longer, but only because their power consumption increased exponentially. That particular workaround has already reached its limits.

I advise you not to expect tomorrow's consumer grade hardware to be significantly faster than today's. Optimizing the software side is now the name of the speed game, or custom hardware specialized to particular tasks.

(L) [2013/03/20] [ost by Dade] [Haswell benchmarks show up] Wayback!

>> hobold wrote:Dade wrote:
the speed up is quite small: something in the 10%-25% range.
Please do not take this as an attack; and not as directed specifically at you. I just want to draw everybody's attention to the awful truth that ...

I have 3 GPUs in my home PC for a total of 10TeraFlops. About 9.7 of the TFlops are from the 3 GPUs. Where do you think I stand in the CPU Vs GPU holy war ?  [SMILEY ;)]
 >> hobold wrote:Optimizing the software side is now the name of the speed game, or custom hardware specialized to particular tasks.
OpenCL on FPGA, not conventional FP accelerator like Epiphany, Xeon Phi, etc. The time of running renderings on CPUs is over.

P.S. ok, ok, now I switch the fan boy mode off  [SMILEY ;)]

(L) [2013/03/20] [ost by mpeterson] [Haswell benchmarks show up] Wayback!

>> Dade wrote:the speed up is quite small: something in the 10%-25% range.
right, the same holds for gpus when we compare the last two generations.

(L) [2013/03/20] [ost by hobold] [Haswell benchmarks show up] Wayback!

>> 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?

(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.)

(L) [2013/03/20] [ost by Serendipity] [Haswell benchmarks show up] Wayback!

I think it will be very interesting to see how haswell performs with optimized AVX code since it is the first CPU where you can actually use it (before that the absent of integer instructions was an absolute showstopper since you were forced to constantly switch between AVX and SSE which, by definition, is slow).

And just for the fun of it  [SMILEY :twisted:] :

GPUs rock because:

A) they have massive driver problems
B) you can´t fix a bug yourself if the bug is in the driver so it slows down your developing speed
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...)
D) they have limited memory ( 6GB is ok for many users but definetly not for all)