@bernd_afa

Quote:

And in this benchmark pa6t show not good speed in compare to other.

It show that PA6T has really good speed in comparison with SPARC64 V, which IS fast CPU (or at least was in its time)...

Quote:

you choose not the right comparision.the fujitsu system is no embedded system.

I though we would like to use desktop/workstation computers, not embedded one...

Quote:

It is design to support lots of RAM, and so its possible that it perform not good on some embedded benches.

I really wonder how much weight you gave to this statement...

Quote:

Dont forget CPU developer know specint, specfp, they can optimize compiler and CPU to this to reach good performance for marketing.

Same for any other benchmark...

Quote:

Look at specint or specfp for Pentium4. It reach good values, but we all know P4 is a slow CPU in praxis

I have here Pentium 4 based computer and Core 2 based computer and I can assure you that my own experience with these computers confirms expected SPEC CPU numbers (eg. Core 2 Quad Q6600 2400 MHz is roughly 5 times faster - in sense Y=5*X - than Pentium 4 1400 MHz in most single core operations).

Quote:

specint /specfp is a 32 bit benchmark.normaly to compare better the coremark for Fujitsu bench have do with 32 bit.

I don´t think I understand what you wrote. However, both PA6T and SPARC64 are 64bit CPUs.

Quote:

maybe should say dont compare a multi CPU Server platform with a embedded system with coremark ?

Finally, we share the same opinion.

I´m looking forward to see some real benchmarks of X1000. Once the X1000 is on the market, we will know who of us was right.

@pavlor
Quote:

I´m looking forward to see some real benchmarks of X1000. Once the X1000 is on the market, we will know who of us was right.

As I said in my previous reply to Bernd_afa, you will see that only partially (just one core working in 32-Bit mode with no Altivec aside from Mplayer).

The X1000 is supposed to be a System that will prompt further evolution of the OS by utilizing what's on board, so while I'm sure that the initial speed will be great compared to current NG Amigas, it will only get better (and better and better etc.) as more features get support (something you will not see from day one).

_________________
SamFlex Complete 800Mhz System + AmigaOS 4.1 Update 4
Amiga 2000 DKB 2MB ChipRam GVP G-Force040 Picasso 2 OS3.9 BB2
AmigaCD 32

@DAX

Quote:

As I said in my previous reply to Bernd_afa, you will see that only partially (just one core working in 32-Bit mode with no Altivec aside from Mplayer).

Yes, these are exactly the benchmarks I´m interested in. I like many emulators and emulation isn´t area where one can benefit from AltiVec or second core.

@bernd_afa

Quote:

you can also look in the GCC compiler settings for instruction cost time.pa6 +/* Instruction costs on PA6T processors. */ +static const +struct processor_costs pa6t_cost = { + COSTS_N_INSNS (6), /* mulsi */ + COSTS_N_INSNS (6), /* mulsi_const */ + COSTS_N_INSNS (6), /* mulsi_const9 */ + COSTS_N_INSNS (8), /* muldi */ + COSTS_N_INSNS (36), /* divsi */ + COSTS_N_INSNS (68), /* divdi */ + COSTS_N_INSNS (8), /* fp */ + COSTS_N_INSNS (8), /* dmul */ + COSTS_N_INSNS (17), /* sdiv */ + COSTS_N_INSNS (32), /* ddiv */ + 64, /* cache line size */ + 64, /* l1 cache */ + 2048, /* l2 cache */ + 12, /* prefetch streams */ +};

Looking at the GCC settings to learn of the PA6T CPU is a very clever idea!
My compliments for this idea.


Regarding the values:
36 clocks for 32bit DIV and 68 clcoks for a 64bit dives are a lot.
If these values relate to the real divide speeds then this CPU has a extrem slow divide unit.

6 clocks for a 32bit mul is real slow also.


If these values are true then this will put the clockrate into a different relation.



BTW: the 5200B is no e300 Core but a 603 core.
The 5200 needs 3 times longer than the e300 core for a mul.
The successor of the 5200 the 5121 whihc is a e300 core.

The 5200 has some limits compared to the e300.
It has halve the perfomrance on logical and shiuft operations and muls take 3 times longer. And the biggest issue is that it can only do one outstanding memory load.


Cheers

_________________
APOLLO the new 68K : www.apollo-core.com

@DAX:
A good Benchmark is Blender. It doesn´t matter, if only one core is supported, because you can render with just one core in Blender. And 64-Bit isn´t about speed, you can address more RAM with 64Bit.

I think with Blender we will have have a good indicator for the raw CPU-speed of the PA6T. I am sure that the PA6T is faster than previous PPC-Amigas, but it will be interesting, how fast it is in comparison to G5-Macs or even to current Intel/AMD-solutions. We are talking about a computer, which costs around 2.000 EUR. At this prize-range I want to know for sure how fast the CPU is.

_________________
AmigaOne X5000/40, 2.2 Ghz, 4 GB RAM, Radeon R9 280X, M-Audio Revolution 5.1, 240 GB SSD

Bernd,

I once read that the PA6 has a 1st Level caceh latency of 4 clocks.
Does anyone know if this is true?

If this is true this might be certainly the biggest performance problem of all.

_________________
APOLLO the new 68K : www.apollo-core.com

@BigGun

Quote:

I once read that the PA6 has a 1st Level caceh latency of 4 clocks. Does anyone know if this is true? If this is true this might be certainly the biggest performance problem of all.

Yes, but I don't think it'll be a big problem, Core i7 also has a 4 cycle latency.

_________________
Whyzzat?

Article on on PA6T at RealWorldTech.

That article goes into considerable detail. It's probably the best you'll get without paying for a journal article.

_________________
Whyzzat?

And some real benchmarks [PPT] see page 20 and 22.

_________________
Whyzzat?

@biggun
>I once read that the PA6 has a 1st Level caceh latency of 4 clocks.
>Does anyone know if this is true?

The pa6t GCC patch is send from a pasemi Email Adress you can seen when you click on link i post.

So i think they know best what there self developed CPU can do.
But wy the pa6t support is not add offical in GCC source i dont know.there is nothing find about pa6t in the GCC history.

>BTW: the 5200B is no e300 Core but a 603 core.
>The 5200 needs 3 times longer than the e300 core for a mul.

here are the values of 603.not much worser as e300 but much better as pa6t

/* Instruction costs on PPC603 processors. */
static const
struct processor_costs ppc603_cost = {
COSTS_N_INSNS (5), /* mulsi */
COSTS_N_INSNS (3), /* mulsi_const */
COSTS_N_INSNS (2), /* mulsi_const9 */
COSTS_N_INSNS (5), /* muldi */
COSTS_N_INSNS (37), /* divsi */
COSTS_N_INSNS (37), /* divdi */
COSTS_N_INSNS (3), /* fp */
COSTS_N_INSNS (4), /* dmul */
COSTS_N_INSNS (18), /* sdiv */
COSTS_N_INSNS (33), /* ddiv */
32, /* cache line size */
8, /* l1 cache */
64, /* l2 cache */
1, /* streams */
};

@pavlor
>I don´t think I understand what you wrote. However, both PA6T and SPARC64 are >64bit CPUs.

but for pa6t the coremark test is compile for 32 bit.and pa6t have more ram clock as 100 MHZ,

look at the specs it use
400 MHZ DDR2 CL5.

fujitsu sparc64 have only 100 MHZ ram clock in the test.

for sparc 64 its compile for 64 bit.
this compiler options show that.

O2 -m64 -DPERFORMANCE_RUN=1

pa6t compiler option is this

-O2 -DPERFORMANCE_RUN=1 -lrt

so its 32 bit.

You tell sparc 64 is a fast CPU, do you know other benchmark results as specint ?
maybe blender etc ?

as i told a P4 is fast in specint specfp

Last edited by bernd_afa on 11-Jul-2010 at 02:57 PM.

@minator
Hi,

Well, it will certainly depends on the usage case.
If you compile number crunching code especially for this CPU then the compiler will hide the latency.

But its easy to imagine that if you execute depending loads and instruction - like typical 68K code - then the latency will be a big drawback.
Also the latency will hurt you extra on mispedicted branches or on any computed jumps.
And the cache latency can ruin many memory processing algorithms.

And the cache of the PA-Semi CPU is relative "simple" for a non-embedded CPU-.
Its only 2-way associative. The 68040 was already 4-way. So cache aliasing and related penalties can be relative high for the PA-Semi.


But I'm sure that the PA-Semi CPU will be a nice CPU.
Maybe the G4 will outrun it in some cases and certainly the G5 does outrun it in mostly all cases but does this really matter?

No, I don't think so.
The AMIGA OS is so slim that you do not need a super G5 or ATHLON or CORE2 CPU to run it.
The 68060 did make AMIGA OS fly and I'm sure the PA Semi will make it fly too.

_________________
APOLLO the new 68K : www.apollo-core.com

@BigGun
>And the cache of the PA-Semi CPU is relative "simple" for a non-embedded CPU-.
>Its only 2-way associative. The 68040 was already 4-way. So cache aliasing and >related penalties can be relative high for the PA-Semi.

another thing i see on pa6t.it have cache line size of 64 bytes.
most other PPC have only 32 bytes.so cache misses cost more or there must use both ram channels.

>Maybe the G4 will outrun it in some cases and certainly the G5 does outrun it in >mostly all cases but does this really matter?

>No, I don't think so.
>The AMIGA OS is so slim that you do not need a super G5 or ATHLON or CORE2 CPU >to run it.
>The 68060 did make AMIGA OS fly and I'm sure the PA Semi will make it fly too.

the Problem is when run this linux programs that are written on fast 2 or more GHZ X86 nobody look if that run fast on a 1 GHZ or embedded CPU.

and on NG amiga i see not that there are other big apps as the Linux apps.
and last not least the X1000 is called a High End computer and should revival the amiga.

with that high price and when the performance of the CPU is not comparable with this high price then its maybe not sell much

Last edited by bernd_afa on 11-Jul-2010 at 03:11 PM.
Last edited by bernd_afa on 11-Jul-2010 at 03:06 PM.

@bernd_afa

Quote:

bernd_afa wrote: @BigGun >And the cache of the PA-Semi CPU is relative "simple" for a non-embedded CPU-. >Its only 2-way associative. The 68040 was already 4-way. So cache aliasing and >related penalties can be relative high for the PA-Semi. another thing i see on pa6t.it have cache line size of 64 bytes. most other PPC have only 32 bytes.so cache misses cost more or there must use both ram channels.

Does the CPU use 64bit wide DDR3 memory?

_________________
APOLLO the new 68K : www.apollo-core.com

@pavlor

So capture the background of AmigaOS 4.1 and use it as background on Windows :P

And you have a FASTER Computer, Looks like AmigaOS and actually also have software :D

@BigGun

>Does the CPU use 64bit wide DDR3 memory?

no only DDR2

@bernd_afa

Bernd,

Of course the typical problem of higher clocked CPU is their memory latency.

And to be able to saturate the longer latency memory interface these CPUs need to have many bursts in flight.

By making the lines longer the burst get longer too which makes saturating the interface easier as the core need to track less burst in flight to achieve this.
So I think the 64byte are long lines are good for the CPU and improve its performance.


You are of course right that long cache lines increase the latenc.
A 64bit line takes 8 memory beats to transmit.
Which means the transfer of courses takes twice as long as a 32byte long would takke.
But is relation to the total memory latency which is probably in the magnitude of 200 clocks this is not a big difference.

Last edited by BigGun on 11-Jul-2010 at 03:31 PM.

_________________
APOLLO the new 68K : www.apollo-core.com

@BigGun: >Maybe the G4 will outrun it in some cases and certainly the G5 does outrun it in >mostly all cases but does this really matter?

I don´t know, but from an HiEnd-Amiga that costs at least EUR 1.800,- I would expect that it outperforms at least a G4 in *every* case. If not, than it makes no great sense to switch from a Pegasos G4 (or you can buy a SAM 460 for less money).

>the AMIGA OS is so slim that you do not need a super G5 or ATHLON or CORE2 CPU to >run it.
>The 68060 did make AMIGA OS fly and I'm sure the PA Semi will make it fly too.

Sure, but when it comes to raw-CPU-power and applications like Blender, than a really fast CPU would be nice. The OS my fly, but when you wait for rendering this doesn´t matter so much.

_________________
AmigaOne X5000/40, 2.2 Ghz, 4 GB RAM, Radeon R9 280X, M-Audio Revolution 5.1, 240 GB SSD

@BigGun
on branch prediction the above linked article (real world tech) says:

"For all CPU designs, the branch prediction is a key element; however, for low power designs, it is even more essential. Mispredicted branches waste a substantial amount of power, because any work done past the branch is squashed and cannot be used. As a result, the PA6T is lavishly equipped with both a 16-entry next line predictor and a tournament predictor that uses a 32Kbit branch history table. "

On the cache problems, shouldn't size (twice the G4) and also L2 availability (4 times the g4) be also considered in the overall performance mix? They say the L2 cache is pretty advanced in that same article:

http://www.realworldtech.com/page.cfm?ArticleID=RWT102405055354&p=3

_________________
SamFlex Complete 800Mhz System + AmigaOS 4.1 Update 4
Amiga 2000 DKB 2MB ChipRam GVP G-Force040 Picasso 2 OS3.9 BB2
AmigaCD 32

@AP
On your Blender example, it must be said that aside from testing per MHZ single core performance (VS other Amiga CPUs), such test won't be able to tell you how Blender would fare if both cores and vector units would be in use.
That would be the real PA6T performance, and rest assured there are no PegasosII that could compare.

_________________
SamFlex Complete 800Mhz System + AmigaOS 4.1 Update 4
Amiga 2000 DKB 2MB ChipRam GVP G-Force040 Picasso 2 OS3.9 BB2
AmigaCD 32

@DAX: >On your Blender example, it must be said that aside from testing per MHZ single core >performance (VS other Amiga CPUs), such test won't be able to tell you how Blender would fare if >both cores and vector units would be in use.

But you can at least compare how fast one core is performing and compare it to other CPUs (as I said before, you can just use one core in Blender for rendering). When I compare a PA6T with a Multicore-G5 and use just one core on both machines, than I can at least see, how fast the X1000 is per core.

I also can compare it to single-core G4s (like the Macmini with MOS or Pegasos wirh AOS4). You can even compare it with Intel/AMD-CPUs if you like. Just use only one core on all these machines and you get an impression how fast the PA6T is in comparison to other CPUs. As I said before: When I spend EUR 1.800+ I want to know how fast the X1000 performs in comparison to other systems (at least previous Amiga-systems).

Last edited by AP on 11-Jul-2010 at 04:13 PM.

_________________
AmigaOne X5000/40, 2.2 Ghz, 4 GB RAM, Radeon R9 280X, M-Audio Revolution 5.1, 240 GB SSD