TT and NemBench

Cyprian · Post by **Cyprian** » Fri Nov 13, 2009 1:31 pm

Can someone post result of NemBench v2.1 test for TT030?
I'm just interested to see how really fast is TT.

Similar topic for Falcon: http://www.atari-forum.com/viewtopic.php?f=27&t=23543

dexterslab · Post by **dexterslab** » Sat Nov 14, 2009 3:35 pm

Code: Select all

NemBench v2.1 - precision CPU/FPU profiler.

Integer multiply (16bit)     -> 1.226 Mips (~200%)
Integer divide (16bit)       -> 0.731 Mips (~202%)
Linear (stalled) integer     -> 16.094 Mips (~202%)
Interleaved (piped) integer  -> 16.094 Mips (~202%)

Float multiply (64bit)       -> 0.541 MegaFlops (~204%)
Float divide (64bit)         -> 0.351 MegaFlops (~203%)
Linear (stalled) float       -> 1.006 MegaFlops (~188%)
Interleaved (piped) float    -> 1.003 MegaFlops (~188%)

16bit read (100% hit)        -> 15.873 MByte/sec (~202%)
16bit write (100% hit)       -> 8.169 MByte/sec (~135%)
32bit read (100% hit)        -> 31.746 MByte/sec (~202%)
32bit write (100% hit)       -> 16.339 MByte/sec (~245%)

Linear 32bit read (ST-Ram)   -> 7.867 MByte/sec (~148%)
Linear 32bit write (ST-Ram)  -> 7.867 MByte/sec (~121%)
Linear 32bit copy (ST-Ram)   -> 3.947 MByte/sec (~122%)

Linear 32bit read (FastRAM)  -> 12.615 MByte/sec (~237%)
Linear 32bit write (FastRAM) -> 15.772 MByte/sec (~244%)
Linear 32bit copy (FastRAM)  -> 7.881 MByte/sec (~244%)

Edited 7-JAN-10 to correct FPU results

wongck · Post by **wongck** » Sat Nov 14, 2009 4:41 pm

Sorry not sure how to read this.

And 100% is a ST @ 8MHz?

dexterslab · Post by **dexterslab** » Sat Nov 14, 2009 5:22 pm

i think % is based on a stock falcon

wongck · Post by **wongck** » Sun Nov 15, 2009 1:17 am

dexterslab wrote:i think % is based on a stock falcon

Makes more sense that way

But still, some performance of the TT is on par with a Falc?
It sure don't feel that way on a stock Falc, the TT seems a lot snappier.

Cyprian · Post by **Cyprian** » Sun Nov 15, 2009 11:40 am

dexterslab wrote:
Code: Select all
NemBench v2.1 - precision CPU/FPU profiler.

thanks dexterslab,
can you specify video mode, and CPU (16MHz od 32MHz) used for that test?

dexterslab · Post by **dexterslab** » Sun Nov 15, 2009 7:16 pm

ST-High, 32mhz cpu and fpu, fastram is a MegTT with 64mb of 60ns ram

Desty · Post by **Desty** » Wed Nov 18, 2009 12:36 am

wongck wrote:But still, some performance of the TT is on par with a Falc?
It sure don't feel that way on a stock Falc, the TT seems a lot snappier.

Not really - the only area in those results that's comparable is the floating point operations.... which is probably not used for 95% of things you and I would do with a TT/Falcon (AFAIK most games, at least in the TT's time, would have used integer operations for everything by using fixed point arithmetic and other tricks to avoid the slow floating point routines/instructions... even raytracers did this).

In pretty much all other areas of that benchmark, the TT seems to destroy the Falcon!

dexterslab · Post by **dexterslab** » Wed Nov 18, 2009 8:43 am

yes, TT is faster. The one place the falcon does redeem itself is the dsp and blit which dont show on that test.

it's a shame we never saw any 040 or 060 upgrades for the TT like is available for the falcon

wongck · Post by **wongck** » Wed Nov 18, 2009 11:32 am

ST ram reads is only 25-50% faster not double as the speed of the CPU/Bus is.
I was expecting 100% faster (double).
Must be the implemetation of the alogorithm rather than brute force speed.

dexterslab · Post by **dexterslab** » Wed Nov 18, 2009 3:44 pm

TT bus is 16mhz

that with higher CPU speed and possible cache differences might explain it

Cyprian · Post by **Cyprian** » Wed Nov 18, 2009 10:23 pm

couple of cycles steals Dram Refresh.

Code: Select all

$FFD000xx-$FFD000xx|Set FastRAM refresh rate and generate a bus error

wongck · Post by **wongck** » Thu Nov 19, 2009 12:37 pm

dexterslab wrote:TT bus is 16mhz

That's explain it.
Atari "hardware lamers" at work again I see.

Cyprian · Post by **Cyprian** » Thu Nov 19, 2009 9:03 pm

wongck wrote:Atari "hardware lamers" at work again I see.

three years later, Amiga's hardware guru produced A1200 with impressive memory bandwidth:

32bit read (chip) 3.7 MB/s
32bit write (chip) 5.6 MB/s

who is the real "hardware lamer" ?

Cyprian · Post by **Cyprian** » Sat Jun 09, 2012 11:13 pm

here is a result for my TT:

NemBench v2.1 - precision CPU/FPU profiler

Integer multiply (16bit) -> 1.229 Mips (~200%)
Integer divide (16bit) -> 0.730 Mips (~201%)
Linear (stalled) integer -> 16.094 Mips (~202%)
Interleaved (piped) integer -> 16.094 Mips (~202%)

Float multiply (64bit) -> 0.263 MegaFlops (~101%)
Float divide %bit) -> 0.175 MegaFlops (~101%)
Linear (stalled) float) -> 0.528 MegaFlops (~99%)
Interleaved (piped) float -> 0.564 MegaFlops (~106%)

16bit read (100% hit) -> 15.873 MByte/sec (~202%)
16bit write (100% hit) -> 6.544 MByte/sec (~108%)
32bit read (100% hit) -> 31.1695 MByte/sec (~201%)
32bit write (100% hit) -> 13.1054 MByte/sec (~195%)

Linear 32bit read (ST-RAM) -> 7.867MByte/sec (~148%)
Linear 32bit write (ST-RAM) -> 7.867MByte/sec (~121%)
Linear 32bit copy (ST-RAM) -> 3.849 MByte/sec (~122%)

Linear 32bit read (FastRAM) -> 14.835 MByte/sec (~279%)
Linear 32bit write (FastRAM) -> 12.615 MByte/sec (~195%)
Linear 32bit copy (FastRAM) -> 4,849 MByte/sec (~150%)

FPU in my TT is twice slower than yours dexterslab, it seems its clocked with 16MHz.
And there is another interesting difference in figures for 16bit write (100% hit), 32bit write (100% hit), Linear 32bit copy (FastRAM).
I thought that all TTs have the same spec, but it seems that they are differ slightly.

wongck · Post by **wongck** » Sun Jun 10, 2012 1:39 am

I think dexterslab have sold off both his TT by now.
Anyway, there is also another thread by him (or someone else) that shows how to make the FPU run at 32MHZ.

DarkLord · Post by **DarkLord** » Sun Jun 10, 2012 3:21 am

Just to make it interesting, here are the nembench results from my PAK 68/3 equipped
STacy. This is med rez, TOS 3.06, 4 megs of ST RAM, but no FAST RAM. The board is
running the CPU and FPU at 32mhz here:

NemBench v2.1 - precision CPU/FPU profiler.

Integer multiply (16bit) -> 1.229 Mips (~200%)
Integer divide (16bit) -> 0.726 Mips (~200%)
Linear (stalled) integer -> 15.968 Mips (~200%)
Interleaved (piped) integer -> 15.968 Mips (~200%)

Float multiply (64bit) -> 0.536 MegaFlops (~202%)
Float divide (64bit) -> 0.347 MegaFlops (~200%)
Linear (stalled) float -> 1.054 MegaFlops (~197%)
Interleaved (piped) float -> 1.123 MegaFlops (~211%)

16bit read (100% hit) -> 15.748 MByte/sec (~200%)
16bit write (100% hit) -> 3.503 MByte/sec (~58%)
32bit read (100% hit) -> 31.347 MByte/sec (~199%)
32bit write (100% hit) -> 3.698 MByte/sec (~55%)

Linear 32bit read (ST-Ram) -> 2.622 MByte/sec (~49%)
Linear 32bit write (ST-Ram) -> 3.573 MByte/sec (~55%)
Linear 32bit copy (ST-Ram) -> 1.575 MByte/sec (~48%)

Just for the fun of it. Looks like no FAST RAM, a 16bit path, and
the Ultrasatan does slow things down compared to a "real" TT.

wongck · Post by **wongck** » Sun Jun 10, 2012 3:38 am

DarkLord wrote:Just for the fun of it. Looks like no FAST RAM, a 16bit path, and
the Ultrasatan does slow things down compared to a "real" TT.

But Integer calculations are similar as expected.

lp · Post by lp » Sun Jun 10, 2012 4:50 am

My TT030 is up and running again. ST High with a clean boot as the doc stated.
Specs are here: http://gfa.atari-users.net/atari/html/about_me.htm

Code: Select all

NemBench v2.1 - precision CPU/FPU profiler.

Integer multiply (16bit)     -> 1.236 Mips (~201%)
Integer divide (16bit)       -> 0.731 Mips (~202%)
Linear (stalled) integer     -> 16.094 Mips (~202%)
Interleaved (piped) integer  -> 16.094 Mips (~202%)

Float multiply (64bit)       -> 0.270 MegaFlops (~101%)
Float divide (64bit)         -> 0.175 MegaFlops (~101%)
Linear (stalled) float       -> 0.528 MegaFlops (~99%)
Interleaved (piped) float    -> 0.564 MegaFlops (~106%)

16bit read (100% hit)        -> 15.873 MByte/sec (~202%)
16bit write (100% hit)       -> 8.116 MByte/sec (~134%)
32bit read (100% hit)        -> 31.746 MByte/sec (~202%)
32bit write (100% hit)       -> 16.181 MByte/sec (~242%)

Linear 32bit read (ST-Ram)   -> 7.867 MByte/sec (~148%)
Linear 32bit write (ST-Ram)  -> 7.867 MByte/sec (~121%)
Linear 32bit copy (ST-Ram)   -> 3.947 MByte/sec (~122%)

Linear 32bit read (FastRAM)  -> 12.566 MByte/sec (~236%)
Linear 32bit write (FastRAM) -> 15.641 MByte/sec (~242%)
Linear 32bit copy (FastRAM)  -> 7.820 MByte/sec (~242%)

DarkLord · Post by **DarkLord** » Sun Jun 10, 2012 6:37 am

lp wrote:My TT030 is up and running again. ST High with a clean boot as the doc stated.
Specs are here: http://www.bright.net/~gfabasic/html/about_me.htm
Code: Select all
NemBench v2.1 - precision CPU/FPU profiler.

Looks pretty consistent across different machines.

(and I like the last 2 lines on your page Lonny)

lp · Post by lp » Sun Jun 10, 2012 10:45 am

I have 2 more TT's in various states of dis-repair to try NemBench on and among the 3 machines they all have different fast ram boards from different makers. I'm curious how consistent the results will be just among my machines. The other machines have a Mighty Mic 32mb and an Atari made 16mb board.

It is good to see consistent values, you sort of get a feeling you have configured things correctly. After all these years I had to relearn a few things while getting the good ole TT up and running again. The computers memory remains the same, where as my memory gets worse with age.

Moulinaie · Post by **Moulinaie** » Mon Jun 11, 2012 10:46 am

Here are some results I though they could be of some interest.

Two aspects are shown:
- the presence of the 16MB TT RAM from Atari vs 64MB TT Ram from MagnumTT
- the use of the stock 32MHz or the CaTTamaran 48MHz

Last thing: using or not the NOVA board or any of the TT resolutions doesn't affect the result.

TT 32MHz + Atari TT RAM

Code: Select all

NemBench v2.1 - precision CPU/FPU profiler.
TT 32MHz, NOVA graphic card, 640x480x32768.
No change in any resolution (NOVA or TT)

Integer multiply (16bit)     -> 1.236 Mips (~201%)
Integer divide (16bit)       -> 0.731 Mips (~202%)
Linear (stalled) integer     -> 16.094 Mips (~202%)
Interleaved (piped) integer  -> 16.094 Mips (~202%)

Float multiply (64bit)       -> 0.541 MegaFlops (~204%)
Float divide (64bit)         -> 0.351 MegaFlops (~203%)
Linear (stalled) float       -> 1.005 MegaFlops (~188%)
Interleaved (piped) float    -> 1.003 MegaFlops (~188%)

16bit read (100% hit)        -> 15.873 MByte/sec (~202%)
16bit write (100% hit)       -> 6.443 MByte/sec (~107%)
32bit read (100% hit)        -> 31.695 MByte/sec (~201%)
32bit write (100% hit)       -> 12.886 MByte/sec (~193%)

Linear 32bit read (ST-Ram)   -> 7.867 MByte/sec (~148%)
Linear 32bit write (ST-Ram)  -> 7.867 MByte/sec (~121%)
Linear 32bit copy (ST-Ram)   -> 3.947 MByte/sec (~122%)

Linear 32bit read (FastRAM)  -> 10.377 MByte/sec (~195%)
Linear 32bit write (FastRAM) -> 12.429 MByte/sec (~192%)
Linear 32bit copy (FastRAM)  -> 6.211 MByte/sec (~192%)

Now I just change the Atari TT RAM with the MagnumTT:

Code: Select all

NemBench v2.1 - precision CPU/FPU profiler.

Integer multiply (16bit)     -> 1.236 Mips (~201%)
Integer divide (16bit)       -> 0.731 Mips (~202%)
Linear (stalled) integer     -> 16.094 Mips (~202%)
Interleaved (piped) integer  -> 16.094 Mips (~202%)

Float multiply (64bit)       -> 0.541 MegaFlops (~204%)
Float divide (64bit)         -> 0.351 MegaFlops (~203%)
Linear (stalled) float       -> 1.006 MegaFlops (~188%)
Interleaved (piped) float    -> 1.003 MegaFlops (~188%)

16bit read (100% hit)        -> 15.873 MByte/sec (~202%)
16bit write (100% hit)       -> 8.169 MByte/sec (~135%)
32bit read (100% hit)        -> 31.746 MByte/sec (~202%)
32bit write (100% hit)       -> 16.339 MByte/sec (~245%)

Linear 32bit read (ST-Ram)   -> 7.867 MByte/sec (~148%)
Linear 32bit write (ST-Ram)  -> 7.867 MByte/sec (~121%)
Linear 32bit copy (ST-Ram)   -> 3.947 MByte/sec (~122%)

Linear 32bit read (FastRAM)  -> 12.615 MByte/sec (~237%)
Linear 32bit write (FastRAM) -> 15.772 MByte/sec (~244%)
Linear 32bit copy (FastRAM)  -> 7.881 MByte/sec (~244%)

16/32 bits WRITE increases and the Fast RAM access increases. Nothing else changes.

Now, the TT at 48MHz with the Atari TT RAM:

Code: Select all

NemBench v2.1 - precision CPU/FPU profiler.
TT with CaTTamaran 48MHz and NOVA graphic card, 640x480x32768.
No change in any resolution (NOVA or TT)

Integer multiply (16bit)     -> 1.858 Mips (~303%)
Integer divide (16bit)       -> 1.096 Mips (~302%)
Linear (stalled) integer     -> 24.165 Mips (~303%)
Interleaved (piped) integer  -> 24.165 Mips (~303%)

Float multiply (64bit)       -> 0.811 MegaFlops (~306%)
Float divide (64bit)         -> 0.526 MegaFlops (~304%)
Linear (stalled) float       -> 1.338 MegaFlops (~251%)
Interleaved (piped) float    -> 1.332 MegaFlops (~250%)

16bit read (100% hit)        -> 23.809 MByte/sec (~303%)
16bit write (100% hit)       -> 8.000 MByte/sec (~132%)
32bit read (100% hit)        -> 47.619 MByte/sec (~303%)
32bit write (100% hit)       -> 15.974 MByte/sec (~239%)

Linear 32bit read (ST-Ram)   -> 7.963 MByte/sec (~149%)
Linear 32bit write (ST-Ram)  -> 7.839 MByte/sec (~121%)
Linear 32bit copy (ST-Ram)   -> 3.938 MByte/sec (~122%)

Linear 32bit read (FastRAM)  -> 12.459 MByte/sec (~234%)
Linear 32bit write (FastRAM) -> 15.438 MByte/sec (~239%)
Linear 32bit copy (FastRAM)  -> 7.728 MByte/sec (~239%)

You can see that the FastRam access is sped up with the CaTTamaran!
Last, the best config, the TT at 48MHz and a MagnumTT fast ram:

Code: Select all

NemBench v2.1 - precision CPU/FPU profiler.

Integer multiply (16bit)     -> 1.858 Mips (~303%)
Integer divide (16bit)       -> 1.096 Mips (~302%)
Linear (stalled) integer     -> 24.165 Mips (~303%)
Interleaved (piped) integer  -> 24.165 Mips (~303%)

Float multiply (64bit)       -> 0.811 MegaFlops (~306%)
Float divide (64bit)         -> 0.526 MegaFlops (~304%)
Linear (stalled) float       -> 1.356 MegaFlops (~254%)
Interleaved (piped) float    -> 1.345 MegaFlops (~252%)

16bit read (100% hit)        -> 23.809 MByte/sec (~303%)
16bit write (100% hit)       -> 10.869 MByte/sec (~180%)
32bit read (100% hit)        -> 47.619 MByte/sec (~303%)
32bit write (100% hit)       -> 21.739 MByte/sec (~326%)

Linear 32bit read (ST-Ram)   -> 7.972 MByte/sec (~150%)
Linear 32bit write (ST-Ram)  -> 7.848 MByte/sec (~121%)
Linear 32bit copy (ST-Ram)   -> 3.938 MByte/sec (~122%)

Linear 32bit read (FastRAM)  -> 15.791 MByte/sec (~297%)
Linear 32bit write (FastRAM) -> 20.954 MByte/sec (~324%)
Linear 32bit copy (FastRAM)  -> 10.485 MByte/sec (~324%)

I have another TT with a 32MB fast ram that seems to be even faster than this one. But never tested with the CaTTamaran.

Guillaume.

lp · Post by lp » Mon Jun 11, 2012 7:40 pm

The documentation for the aixTT claims its faster than Atari's burst mode. Perhaps this is true of the Magnum as well. Interesting results.

Cyprian · Post by **Cyprian** » Mon Jun 11, 2012 10:41 pm

It seems that my TT Ram is the slowest one:
Linear 32bit copy (FastRAM) -> 4,849 MByte/sec (~150%)
This is Mighty Mic 32b with nibble mode SIMMs, I need to check another faster SIMMs.

wongck wrote:I think dexterslab have sold off both his TT by now.
Anyway, there is also another thread by him (or someone else) that shows how to make the FPU run at 32MHZ.

that is easy, 15 years ago I overclocked FPU in my Falcon

wongck · Post by **wongck** » Tue Jun 12, 2012 12:47 am

It not much of an hack actually.
IIRC it a user selectable thing... was just a jumper that select 16MHz and 32MHz.

Of course you can feed in a clock at 48MHz or 50MHZ into the FPU.

Atari-Forum

TT and NemBench

TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench

Re: TT and NemBench