| Apollo Hardfpu Results | page 1 2 3 4 5 6 7 8
|
|---|
|
|---|
| | 
Gunnar von Boehn
(Apollo Team Member)
Posts 6214
04 Oct 2017 18:43
| Update
again speed increase
| |
| | 
Steve Ferrell
Posts 424
04 Oct 2017 21:57
| This is identical to yesterday's screenshot.
| |
| |
Gunnar von Boehn
(Apollo Team Member)
Posts 6214
04 Oct 2017 22:03
| Steve Ferrell wrote:
|
This is identical to yesterday's screenshot.
|
we replaced the picture - maybe the browser cached it and
you need refresh/reload your browser?
Its 67 MFLOPS now.
We started yesterday with 55 MFlops
Then tweaked it to 60 MFlops
Now we have 67 MFlops
In SySINFO
An 68882 FPU scores ~ 1.0 MFLOPS
An 68040@25 MHz FPU scores ~ 5 MFlops
An 68040@40 MHz FPU scores ~ 8 MFlops
An 68060@50 MHz FPU scores ~ 28 MFlopsThe SYSINFO MFLOPS code is sequential prevents APOLLO to use its full power.
Nevertheless I think our current score of 67 MFlops is very good for this test.
| |
| |
Steve Ferrell
Posts 424
04 Oct 2017 22:38
| Nice!
| |
| | 
Thierry Atheist
Posts 644
05 Oct 2017 04:03
| So... assuming it's a Vampire 2... and extrapolating from: 68060@50 MHz FPU scores ~ 28 MFlops
Motorola Apollo Core
Vampire 2
68060@50MHz * FPGA=~119.6MHz (equivalent)
~28 MFlops * 67 MFlops --- FPU results = TOTAL AWESOME!!!!!
| |
| | 
Kolbjørn Barmen
(Needs Verification)
Posts 219/ 2
05 Oct 2017 09:51
| Thierry Atheist wrote:
|
So... assuming it's a Vampire 2
|
Why would you assume this is Vampire 2?
| |
| |
Thierry Atheist
Posts 644
05 Oct 2017 10:48
| Kolbjørn Barmen wrote:
| Thierry Atheist wrote:
| | So... assuming it's a Vampire 2 |
Why would you assume this is Vampire 2? |
Hi Kolbjørn,The MIPS rating is just 114.77. P.S.sssst (Gunnar, you should have blurred that out. ;-) )
| |
| |
Gunnar von Boehn
(Apollo Team Member)
Posts 6214
05 Oct 2017 11:19
| Thierry Atheist wrote:
|
The MIPS rating is just 114.77.
|
A) The Mips rating of SYSINFO is wrong.
SYSINFO only scores 50% of real MIPS on 68060 and 68080. B) APOLLO on V2 and V4 is the same core!
V4 has some advantages as:
- bigger caches
- more memory
- faster memory
These improvements are of real advantage in real live but
are not visible in SYSINFO as fits completely into cache and does not work with memory. So from the SYSINFO score you can not see if its a V2 or V4
| |
| | 
Szyk Cech
Posts 191
05 Oct 2017 17:36
| Gunnar von Boehn wrote:
|
V4 has some advantages as:
- bigger caches
- more memory
- faster memory
|
How much impact on speed all these features?!? (I mean: how much faster is V4 than V2?)
| |
| | 
Mr Niding
Posts 459
05 Oct 2017 18:07
| Szyk Cech wrote:
|
Gunnar von Boehn wrote:
|
V4 has some advantages as:
- bigger caches
- more memory
- faster memory
|
How much impact on speed all these features?!? (I mean: how much faster is V4 than V2?)
|
Isnt that hard to measure beyond pure benchmarks? I would think it depends on the program(s) being programmed to be fully utilizing the core features etc...?
| |
| | 
Markus Horbach
Posts 35
05 Oct 2017 21:23
| An intel FAE once said :
make your code fit to the caches or your program will be extremely slow. Bigger caches mean you can do more compex code with more complex data structures. Not every algorithm can fit in any cache size.
You can lift more heavyweight problems with good performance when you increase the cache size, but you stall again when you reach the limits of your machine.
In contrast, a benchmark is to showcase the theoretical maximum of your machine in best conditions/circumstances. That does not mean that your application can benefit of the maximum performance when it does not fit to the machine or does not use the full potential.
| |
| | 
Andy Hearn
Posts 374
06 Oct 2017 09:39
| Ok, so, let’s have a little quake FPS demo on that nice new cpu/fpu ;)
Come on you know you want to :) .... ok, I want you to :D
| |
| | 
Rod March
Posts 119
06 Oct 2017 12:09
| Nixus Minimax wrote:
|
Not necessarily. The hard-FPU and everything else fit the v4 FPGA easily but even though these tests are being done on the v2 (a well-known reliable hardware suitable for testing), some features planned for v2 Gold3/Platinum are probably missing or removed in order to make space for the FPU in the v2.
|
This is all simply awesome.
Soft FPU is already great, but if it is possible to get hard FPU in the V2 it would be brilliant beyond words to have the option.
Sorry, have probably said all this before. Guess I'm just super-excited about doing 3D on my A2000...Awesome work Gunnar, keep it up!
| |
| |
Gunnar von Boehn
(Apollo Team Member)
Posts 6214
06 Oct 2017 12:37
| Rod March wrote:
|
Soft FPU is already great, but if it is possible to get hard FPU in the V2 it would be brilliant beyond words to have the option.
|
Let us clarify this.
Yes, 68080 HARD-FPU is working!
Yes, 68080 HARD-FPU is faster than 68060.
Yes, software making utilizing the fully Pipelined and Super-Scalar capabilities of the 68080 HARD-FPU can reach speed of 10 times 68060. Yes, 68080 Hard-FPU, fits nicely in VAMPIRE-4. While we have in the team builds with partial HARD-FPU for VAMP2 - please understand we can _not_ promise this for a customer release for the VAMP-2 as of today.
| |
| |
Gunnar von Boehn
(Apollo Team Member)
Posts 6214
06 Oct 2017 21:59
| Another benchmark, some different numbers 
| |
| | 
Nicolas Sipieter
(Needs Verification)
Posts 115/ 1
07 Oct 2017 08:42
| this screenshot must be, the partial-hard-fpu core for v2 i suppose?
(considering lower score: 40 < 68 mflops and also the amount of ram: 128mb)
if so, then it's better than femu score: 10 < 40 mflops.
that would be good for v2 owners.
| |
| |
Gunnar von Boehn
(Apollo Team Member)
Posts 6214
07 Oct 2017 09:13
| nicolas sipieter wrote:
|
this screenshot must be, the partial-hard-fpu core for v2 i suppose?
(considering lower score: 40 < 68 mflops)
|
Different tests give different scores.
The CPU supports hundreds of different operations.
The FPU supports also hundreds of operations. The execution time of different operations is different. Its obvious that a DIVIDE is more complex and takes longer than an ADD. So if a benchmark does only ADDs its will score a different result than another benchmark which e.g. does only DIVs. Which number is now correct?
| |
| |
Mr Niding
Posts 459
07 Oct 2017 09:18
| Well, I guess thats why the Wintel market reviewers post pages of different benchmarks when they review CPUs and GPUs. So including the same for Apollo seems like a good idea :) Sysinfo, SysSpeed, rendering on Cinema4d, FPS on xyz game(s) etc. This takes time to do tho, so thats best left for general users to do when cores are released so developers dont feel the need to spend time producing performance demostrations instead of developing the core. (But ofcourse I DO love reading/watching benchmarks/demostrations!)
| |
| | 
Lorenzo Pistone
Posts 22
07 Oct 2017 09:24
| Gunnar von Boehn wrote:
|
nicolas sipieter wrote:
|
this screenshot must be, the partial-hard-fpu core for v2 i suppose?
(considering lower score: 40 < 68 mflops)
|
Different tests give different scores.
The CPU supports hundreds of different operations.
The FPU supports also hundreds of operations.
The execution time of different operations is different.
Its obvious that a DIVIDE is more complex and takes longer than an ADD.
So if a benchmark does only ADDs its will score a different result than another benchmark which e.g. does only DIVs.
Which number is now correct?
|
There is a program to test the real power of the cpu/fpu that we can use instead of sysinfo and that maybe is multiplatform so we can compare the results with other systems?
| |
| | 
Asaf Ayoub
Posts 26
07 Oct 2017 09:33
|
Some people use this FPU test, it is part of HPC Challenge Benchmark.wget EXTERNAL LINK mv linpackc.new linpack.c
gcc linpack.c -o linpack -lm
./linpack EXTERNAL LINK
| |
|
Performance and Benchmark Results!