Click Here
home features news forums classifieds faqs links search
6071 members 
Amiga Q&A /  Free for All /  Emulation /  Gaming / (Latest Posts)
Login

Nickname

Password

Lost Password?

Don't have an account yet?
Register now!

Support Amigaworld.net
Your support is needed and is appreciated as Amigaworld.net is primarily dependent upon the support of its users.
Donate

Menu
Main sections
» Home
» Features
» News
» Forums
» Classifieds
» Links
» Downloads
Extras
» OS4 Zone
» IRC Network
» AmigaWorld Radio
» Newsfeed
» Top Members
» Amiga Dealers
Information
» About Us
» FAQs
» Advertise
» Polls
» Terms of Service
» Search

IRC Channel
Server: irc.amigaworld.net
Ports: 1024,5555, 6665-6669
SSL port: 6697
Channel: #Amigaworld
Channel Policy and Guidelines

Who's Online
22 crawler(s) on-line.
 139 guest(s) on-line.
 1 member(s) on-line.


 OlafS25

You are an anonymous user.
Register Now!
 OlafS25:  54 secs ago
 BigD:  12 mins ago
 retrofaza:  15 mins ago
 kolla:  29 mins ago
 edwardsjethro:  1 hr 21 mins ago
 joeyunderwood:  1 hr 22 mins ago
 Sikharubel:  1 hr 25 mins ago
 Musashi5150:  1 hr 48 mins ago
 MagicSN:  2 hrs 16 mins ago
 Gunnar:  2 hrs 23 mins ago

/  Forum Index
   /  Free For All
      /  GFLOPS
Register To Post

PosterThread
KimmoK 
GFLOPS
Posted on 16-Jun-2015 13:04:40
#1 ]
Elite Member
Joined: 14-Mar-2003
Posts: 5211
From: Ylikiiminki, Finland

more meaningles info...? ;P

From web:

- i7 920 3.4ghz 69GFLOPS (double precision, all threads)
- Intel Core i7 2600K @5.2GHz, RAM 2134MHz 7-10-7-27 GFLOPS:140 (double precision, all threads)

"Theoretical peak is complicated on the new generations of processors but it is basically ....
the Core i7 5960X we expect a theoretical peak of,
3.0GHz * 8-cores * 8 DP-vector ops * 2 from FMA3 = 384 GFLOPS "
Core i7 5960X (Haswell E) 8 core @ 3.0GHz AVX2 linpack actual result: 354

T4240 1,8Ghz 216GFLOPS (single precision, 42,4 for double precision, all threads)
((18sp and 3,6dp GFLOPS per core))

According to seti@home the fastest computer so far has done only 43GFLOPS (i7-4790K CPU @ 4.00GHz) ... so, reality is far from benchmark result?

...

From various sources, I think with T4240 is as fast as Core i7 920 4 cores @ 2.66GHz SSE4.2 (linpack actual result:40)

My fastest x86 is AMD FX4300 (y2013 model), it should do only 10...12GFLOPS. (sisoft sandra results from web)
Weird... is AMD really that pathetic vs intel?
Or do they just rely on their GPU for GLOPS?

++++
((btw, I read that on intel HW, the single precision results should be 2x of the DP result.))

Anyway, i7 is not energy efficient but it can be 8x faster than T4240 if FP math.... I think.

I see thousand(s) GFLOPS listed for GPU results via OpenCL.
I wonder if those are single or double precision results?

++++
Also BayTrail is weak in GFLOPS test:
http://www.sisoftware.co.uk/?d=qa&f=cpu_byt


+++
Bed time readings for myself:
http://www.cnet.com/news/opencl-2-0-brings-new-graphics-chip-power-to-software/
https://www.ibm.com/developerworks/community/wikis/home?lang=en#!/wiki/Wbf059a58a9b9_459d_aca4_493655c96370/index
https://www.khronos.org/opencl/
I became sleepy allready...

+++

After quick look, OpenCL is unavailable for PowerPC Linux and freescale has not shown any interest towards OpenCL. x86, x64, Power and ARM developments exist.

+++
Video processing CPU vs GPU:
http://imagescdn.tweaktown.com/content/6/4/6489_52_exploring_sony_vegas_pro_13_rendering_with_a_visiontek_radeon_r9_290.png
http://imagescdn.tweaktown.com/content/6/5/6539_71_lenovo_thinkpad_w540_mobile_workstation_laptop_review.png
http://imagescdn.tweaktown.com/content/6/4/6489_54_exploring_sony_vegas_pro_13_rendering_with_a_visiontek_radeon_r9_290.png

Here we see some RadeonHD cards vs the rest: http://images.anandtech.com/graphs/graph7503/59859.png

Then, it would be interesting to find info of how much PCIe speed affects to OpenCL results...

Last edited by KimmoK on 16-Jun-2015 at 02:01 PM.
Last edited by KimmoK on 16-Jun-2015 at 02:00 PM.
Last edited by KimmoK on 16-Jun-2015 at 01:58 PM.
Last edited by KimmoK on 16-Jun-2015 at 01:38 PM.
Last edited by KimmoK on 16-Jun-2015 at 01:30 PM.
Last edited by KimmoK on 16-Jun-2015 at 01:21 PM.
Last edited by KimmoK on 16-Jun-2015 at 01:09 PM.

_________________
- KimmoK
// For freedom, for honor, for AMIGA
//
// Thing that I should find more time for: CC64 - 64bit Community Computer?

 Status: Offline
Profile     Report this post  
olegil 
Re: GFLOPS
Posted on 16-Jun-2015 14:41:41
#2 ]
Elite Member
Joined: 22-Aug-2003
Posts: 5895
From: Work

@KimmoK

GPUs are typically 16, 24 and 32 bit (or at least they were last time I checked). Most math problems actually DON'T need a 52 bit fraction and 11 bits of exponent, but it requires THINKING to scale down from a DP to a 32 bit fixed-point calculation.

I'm an MCU programmer at heart, so I like to work in fixed-point. I even extended my Atmega printf to support dynamic precision of integers, now I do printf("%02.3d",variable) all the time, cause I didn't need the flexibility of a float, I just needed to not have to do printf("%02d.%03d",variable/1000,variable%1000). The implementation was simple, but still took me a couple of attempts to get right.

Going back to ANSI C sucks so bad after getting used to that

But I digress. GPUs don't normally do double precision floating point, but they are VERY good at fixed-point calculation which can mostly be used instead (you know, if you add a large number to a float all those small bits you used to have are lost anyway). A float is a 53 bit number that you can shift an awful lot of places up or down.

_________________
This weeks pet peeve:
Using "voltage" instead of "potential", which leads to inventing new words like "amperage" instead of "current" (I, measured in A) or possible "charge" (amperehours, Ah or Coulomb, C). Sometimes I don't even know what people mean.

 Status: Offline
Profile     Report this post  
cdimauro 
Re: GFLOPS
Posted on 19-Jun-2015 4:00:16
#3 ]
Elite Member
Joined: 29-Oct-2012
Posts: 3621
From: Germany

@KimmoK: after MIPS now GFLOPS? They suffer from the same problems which I reported for the MIPS.

Please, stop looking at theoretic numbers, and take some real world applications which are useful for you.

@olegil: GPUs use single precision floats from very long time (around 10 years; more for FP16-FP24).

 Status: Offline
Profile     Report this post  
KimmoK 
Re: GFLOPS
Posted on 25-Jun-2015 9:54:36
#4 ]
Elite Member
Joined: 14-Mar-2003
Posts: 5211
From: Ylikiiminki, Finland

@cdimauro

MIPS ratings are the most common performance estimate that I find from CPU/SoC specs, for integer.
It's the most commonly used to get rough idea what a core is capable to do, especially in embedded solution.

GFLOPS values are little bit similar, even if less common.

>take some real world applications which are useful for you

HandBrake could be that. But I doubt I find any results for other than x86 & x86 version of OpenCL.
Then I can not compare.

_________________
- KimmoK
// For freedom, for honor, for AMIGA
//
// Thing that I should find more time for: CC64 - 64bit Community Computer?

 Status: Offline
Profile     Report this post  
cdimauro 
Re: GFLOPS
Posted on 25-Jun-2015 18:22:00
#5 ]
Elite Member
Joined: 29-Oct-2012
Posts: 3621
From: Germany

@KimmoK

Quote:

KimmoK wrote:
@cdimauro

MIPS ratings are the most common performance estimate that I find from CPU/SoC specs, for integer.

Because it's cheap. But it doesn't mean that it's a good measure, as I tried to explain several time.
Quote:
It's the most commonly used to get rough idea what a core is capable to do, especially in embedded solution.

It might be the most commonly used by people which knows of computer architecture as well as I know of home improvement.
Quote:
GFLOPS values are little bit similar, even if less common.

It's less common only because a limited number of architectures reached that threshold. In fact, previously MFLOPS were used. And before MFLOPS there was nothing, because CPUs lacked FPUs.

Nevertheless, it's a misleading unit of measure which is susceptible of the same critics of MIPS.
Quote:
>take some real world applications which are useful for you

HandBrake could be that. But I doubt I find any results for other than x86 & x86 version of OpenCL.
Then I can not compare.

Blender? POVRay? Cinema 4D? There are a lot of REAL applications which stress a lot the floating point ability of a CPU.

 Status: Offline
Profile     Report this post  
KimmoK 
Re: GFLOPS
Posted on 26-Jun-2015 8:33:47
#6 ]
Elite Member
Joined: 14-Mar-2003
Posts: 5211
From: Ylikiiminki, Finland

@cdimauro

>>It's the most commonly used to get rough idea what a core is capable to do, especially in embedded solution.
>It might be the most commonly used by people which knows of computer architecture as well as I know of home improvement.

Well, then you are perhaps better in home improvement than Freescale or APM/AMCC (etc.) people are at computer architecture, as it seems they mainly release MIPS info on their material.

>>>take some real world applications which are useful for you
>>HandBrake could be that. But I doubt I find any results for other than x86 & x86 version of OpenCL.
>>Then I can not compare.
>Blender? POVRay? Cinema 4D? There are a lot of REAL applications which stress a lot the floating point ability of a CPU.

I was not aware that those support OpenCL.

But for CPU benchmarking we have a few usefull results and hundreds of unusefull results. (of blender, where you must have same compiler version, blender version and OS)

To evaluate for some future project and a sampling (not in production) SoC, one usually can not run blender.

Last edited by KimmoK on 26-Jun-2015 at 09:24 AM.

_________________
- KimmoK
// For freedom, for honor, for AMIGA
//
// Thing that I should find more time for: CC64 - 64bit Community Computer?

 Status: Offline
Profile     Report this post  
cdimauro 
Re: GFLOPS
Posted on 26-Jun-2015 21:49:24
#7 ]
Elite Member
Joined: 29-Oct-2012
Posts: 3621
From: Germany

@KimmoK

Quote:

KimmoK wrote:
@cdimauro

>>It's the most commonly used to get rough idea what a core is capable to do, especially in embedded solution.
>It might be the most commonly used by people which knows of computer architecture as well as I know of home improvement.

Well, then you are perhaps better in home improvement than Freescale or APM/AMCC (etc.) people are at computer architecture, as it seems they mainly release MIPS info on their material.

I already reported why they do it. I quote myself:

"Because it's cheap"

Nevertheless:

"But it doesn't mean that it's a good measure, as I tried to explain several time."

In fact, the most important benchmark is the famous SPEC, which is more useful than the misleading MIPS (and MFLOPS/GFLOPS too). For HPC, instead, the most important one is LINPACK. For databases there's a specific one, but currently I didn't recall the name.

Again, it's not rocket science: that's how the industry works.
Quote:
>>>take some real world applications which are useful for you
>>HandBrake could be that. But I doubt I find any results for other than x86 & x86 version of OpenCL.
>>Then I can not compare.
>Blender? POVRay? Cinema 4D? There are a lot of REAL applications which stress a lot the floating point ability of a CPU.

I was not aware that those support OpenCL.

Because it doesn't make sense to talk about OpenCL applications, which mostly test the GPU and NOT the CPU "power".
Quote:
But for CPU benchmarking we have a few usefull results and hundreds of unusefull results. (of blender, where you must have same compiler version, blender version and OS)

To evaluate for some future project and a sampling (not in production) SoC, one usually can not run blender.

If you know other real-world applications, report them.

But, please, avoid useless synthetic benchmarks.

 Status: Offline
Profile     Report this post  

[ home ][ about us ][ privacy ] [ forums ][ classifieds ] [ links ][ news archive ] [ link to us ][ user account ]
Copyright (C) 2000 - 2019 Amigaworld.net.
Amigaworld.net was originally founded by David Doyle