@mlehto

Quote:

INTEGER & WRITING 512 Kb block: 121.04 / 1473.38 Mb/s INTEGER & WRITING 1024 Kb block: 120.61 / 161.64 Mb/s

He's got 512kB L2 cache

_________________
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.

@mbilla

...or you may even ask what movie, and in conjunction with what substances ?

_________________
XH558 - the worlds last flying Vulcan.
ok, its actually XL426 in the picture but you know what I mean.

If you have 256 kB cache and you transfer 512 kB test packets, does half fit in the cache and run at high speed, while the other half does not fit in the cache and runs at normal speed?

tony

_________________
cheers
tony

Hyperion Support Forum: http://forum.hyperion-entertainment.biz/index.php

@tonyw

May be borderline case. Look down there.


@Olegil

:)


I think so also. He didn't bother dig his case open, I didn't bother to dig info from net .... don't know, how up to date are old adverts about uA1


Strange is ... :

INTEGER & WRITING 512 Kb block: 121.04 / 1473.38 Mb/s
INTEGER & WRITING 1024 Kb block: 120.61 / 161.64 Mb/s

INTEGER & READING 512 Kb block: 199.22 / 1587.60 Mb/s
INTEGER & READING 1024 Kb block: 197.49 / 1039.59 Mb/s

FL-POINT & WRITING 512 Kb block: 122.93 / 1914.02 Mb/s
FL-POINT & WRITING 1024 Kb block: 122.27 / 1008.87 Mb/s

FL-POINT & READING 512 Kb block: 209.41 / 2155.79 Mb/s
FL-POINT & READING 1024 Kb block: 207.71 / 1248.78 Mb/s

In case I'm not going insane, there may be some sw-misbehavior.

I'm last worked in quality situations in chemical and medical industry. With computer world style people will die for medication and papermill's will stop immediatly forrrrrreverrr ... ;) Just joking, nothing amiga related, just generally.

Ok, I start install linux for testing.

Miikka

@kgrach

Quote:

What settings in ramspeed are you using as mine are way faster than yours. 800Mhz 750FX A1XE.

Try using the -b3 parameter. I use that because it helpfully gives an average instead of just a stream of numbers. My 600MHz SE gives these results :

4Gb per pass mode

INTEGER Copy: 130.65 Mb/s
INTEGER Scale: 129.66 Mb/s
INTEGER Add: 132.30 Mb/s
INTEGER Triad: 137.45 Mb/s
---
INTEGER AVERAGE: 132.52 Mb/s

4Gb per pass mode

INTEGER Copy: 144.48 Mb/s
INTEGER Scale: 143.12 Mb/s
INTEGER Add: 143.96 Mb/s
INTEGER Triad: 151.85 Mb/s
---
INTEGER AVERAGE: 145.85 Mb/s

_________________
Who do you serve, and who do you trust? - Galen

@DrBombcrater


@mlehto
Quote:


>Rather than any intrinsic performance problem with the Articia, I'd guess the >slow memory benchmarks are probably the result of U-Boot programming the >memory controller with the slowest possible timings in order to increase Dimm >compatibility. Which is completely the right thing to do considering how picky >the A1 is about Dimms.



You just made me happy, really

So it is possible to adjust timings from u-boot setup, if/when these additions come avail. And get most out from our computer. Where ever it is cabable. Yes!!!


Funny, I just learned, howto use these , these and these with IBrowse without copletely working javascript engine. So the fun begin...

What's this .... .-) ... ermh ...

EDITED:

.-) didn't make any effect...

And someone presumes, that UBOOT currently can't initialize our SE-mobos completely correctly, what you think ? This explanation makes more sense for me, than my expl. about poorly working PSUs, I think.

Miikka

Last edited by mlehto on 12-Jan-2005 at 12:56 PM.

@mlehto

Quote:

So it is possible to adjust timings from u-boot setup, if/when these additions come avail.

There are no memory timing options in U-Boot and I've no idea if they'll ever be put in. I wrote my own utility to change the settings (and, no, it isn't publically available yet).

Quote:

And someone presumes, that UBOOT currently can't initialize our SE-mobos completely correctly, what you think ?

U-Boot is initialising the board just fine but it is taking a conservative approach, which is what firmware should always do. A machine that runs a little slower than it is capable of is much more useful than one that won't boot because your memory can't handle aggressive timings.

Using registered memory imposes a speed penalty, too. I'd really like to see some benchmarks for an A1 with unregistered memory.

_________________
Who do you serve, and who do you trust? - Galen

@DrBombcrater

Quote:
So it is possible to adjust timings from u-boot setup, if/when these additions come avail.

Quote:
There are no memory timing options in U-Boot and I've no idea if they'll ever be put in. I wrote my own utility to change the settings (and, no, it isn't publically available yet).


Keep us informed, when it is. Please


Quote:
Using registered memory imposes a speed penalty, too. I'd really like to see some benchmarks for an A1 with unregistered memory.


That's something, what came to my mind allso. I decided to buy registered in late 2002, because it was in time more or less sure, that it works.

I'm not really nothing hw-guru, but intrested anyway.


Miikka

@DrBombcrater

Quote:

Using registered memory imposes a speed penalty, too. I'd really like to see some benchmarks for an A1 with unregistered memory.

Are you sure?
A read or write access to registered SDRAM is generally delayed by one clock cycle due to the additional register, but the maximum transfer speed can be achieved anyway only with large burst transfers. IMHO a delay of one clock cycle should not decrease the memory performance that much.
(Although I have to admit, that I never measured the performance of normal/registered SDRAM)

_________________
A1SE: G3@600MHz, 2GB, 1GBit network card
A1XE: G4@933MHz, 2GB, refitted AC'97 codec
microA1: G3@800MHz, 1GB

- A1 Linux support -

@DrBombcrater

Quote:

Quote: What northbridge is Adam going to use? Your guess is as good as mine on that one, but the scanty information that's been released suggests it could be one of the Marvell Discovery II family, because the specs (AGP 1x, DDR) are quite similar to the Pegasos 2.

Ha, ha, ha. Don't you know which chip is "used" in this brain-project of Adam? Really?

Please recall WHEN Adam made his announcement. It was SIX days after Elbox's news about the Dragon-ColdFire board. Right?

In the discussion on Dragon (here is the link) it was YOU who mentioned the name of the Freescale chip, which could be built in the Dragon-PPC version. Right?

There were some very nervous, unfavourable comments of some people in this discussion. It looks like they took Dragon as a competitor to their business. A solution to this "Dragon-problem" was necessary. So, this A1 guy--Adam--made up an announcement about new PPC cards for A1200 using the PPC chip suggested by YOU and its simplified version.

Do you know now what is this mysterious "northbridge" chip with DDR support?
And why this chip name was not mentioned in the Adam's announcement?

Simple, isn't it?

@tjaoz

Quote:

In the discussion on Dragon (here is the link) it was YOU who mentioned the name of the Freescale chip, which could be built in the Dragon-PPC version. Right?

Yep, no doubt about that.

Quote:

Do you know now what is this mysterious "northbridge" chip with DDR support?

No, but I'd still guess it's a Discovery II or Discovery LT. You'll notice, if you've bothered to read Adam's announcement rather than just skimming it, that the product listed as having DDR memory slots also uses a PPC750CXe processor.

Not an 8245, 8555E or any other SoC chip. The 750CXe. That's a standard PPC processor containing no memory controller or PCI host. It requires a North Bridge chip (or the functional equivalent) to handle those tasks. The AmigaOne uses the MAI ArticiaS for these purposes, the Pegasos2 uses the Marvell Discovery II.

I'm afraid your nutty conspiracy theory is nothing but a paper tiger, with not a shred of validity to it. Sorry...

For the record, I have no connection at all to Adam, and I know nothing of these products except what was in the public announcement. I'm just an interested observer who is looking forward to both devices arriving (particularly the PPC750 busboard) and I can't see anything to suggest that won't happen.

_________________
Who do you serve, and who do you trust? - Galen

@Geri

Quote:
A read or write access to registered SDRAM is generally delayed by one clock cycle due to the additional register, but the maximum transfer speed can be achieved anyway only with large burst transfers. IMHO a delay of one clock cycle should not decrease the memory performance that much.
(Although I have to admit, that I never measured the performance of normal/registered SDRAM)


I think, that in real world operations there is difference and unreg is bit faster.

But when you transfer big chunks, it may be just same (?).

Miikka

@Geri

Quote:

(Although I have to admit, that I never measured the performance of normal/registered SDRAM)

Me neither. The gap probably depends very much on the application being used to measure it. Code that does non-linear accesses will suffer much more than an application that is doing block accesses.

_________________
Who do you serve, and who do you trust? - Galen

@DrBombcrater

Quote:

You'll notice, if you've bothered to read Adam's announcement rather than just skimming it, that the product listed as having DDR memory slots also uses a PPC750CXe processor.

Adam did NOT write 'The CPU card will initially include a 750CXe processor,'
he wrote: 'The CPU card will initially be a 750CXe product,'

750CXe = G3
The official Freescale name of the G3 is e500.
MPC8555e chip, you mentioned in the Dragon discussion, includes an e500 core, DDR controller and two PCI buses.

I guess Adam did not want to use the chip name mentioned by you in his announcement. Why? Probably because he was afraid that it would exhibit that his A1200 PPC cards idea is 6-days old. That's why Adam wrote 'a 750CXe product' instead.

Simple, isn't it?

@DrBombcrater

> I'd really like to see some benchmarks for an A1 with unregistered memory.

The test on an A1XEG4 1000MHz, 32k L1 cache, 256k L2 cache, 2M L3 cache and unregistered memory gives me these results:

FL-POINT & READING 32 Kb block: 7585.19 Mb/s
FL-POINT & READING 64 Kb block: 2327.27 Mb/s
FL-POINT & READING 128 Kb block: 2327.27 Mb/s
FL-POINT & READING 256 Kb block: 2275.56 Mb/s
FL-POINT & READING 512 Kb block: 276.01 Mb/s
FL-POINT & READING 1024 Kb block: 228.83 Mb/s
FL-POINT & READING 2048 Kb block: 225.55 Mb/s
FL-POINT & READING 4096 Kb block: 225.55 Mb/s

FL-POINT & WRITING 32 Kb block: 2528.40 Mb/s
FL-POINT & WRITING 64 Kb block: 2250.55 Mb/s
FL-POINT & WRITING 128 Kb block: 2250.55 Mb/s
FL-POINT & WRITING 256 Kb block: 2178.72 Mb/s
FL-POINT & WRITING 512 Kb block: 669.28 Mb/s
FL-POINT & WRITING 1024 Kb block: 519.80 Mb/s
FL-POINT & WRITING 2048 Kb block: 495.88 Mb/s
FL-POINT & WRITING 4096 Kb block: 487.62 Mb/s

INTEGER Copy: 163.97 Mb/s
INTEGER Scale: 164.37 Mb/s
INTEGER Add: 159.34 Mb/s
INTEGER Triad: 174.50 Mb/s
---
INTEGER AVERAGE: 165.54 Mb/s

FL-POINT Copy: 195.42 Mb/s
FL-POINT Scale: 162.41 Mb/s
FL-POINT Add: 153.68 Mb/s
FL-POINT Triad: 153.64 Mb/s
---
FL-POINT AVERAGE: 166.29 Mb/s


Kind regards,

Dennis

@tjaoz

Now you are really clutching at straws. 8555E != 750CXe. The chips may share a common heritage, but they are not the same and a CPU card described as being a "750CXe product" could not possibly be based on anything but a 750CXe.

_________________
Who do you serve, and who do you trust? - Galen

@tjaoz

Incorrect .

I heard about adams project quite some time ago. I think over a year ago.
There was allot of discussion about it on IRC.

The Elbox Dragon project forced him to announce his work earlier than anticipated.
I beleive he was planning on an official announcement only when it was ready for public release.

kgrach

The boards are not vapor and are a REAL design.

@mlehto

Quote:

If it just access L1/L2 cache, these high speeds are ok ??

Duh! I wasn't thinking about that.

I made the assumption that is was hiting the SDRAM directly.
Thats why I hate programs like that.

also It looks like Uboot does read the SPD as different SDRAM results in different speeds.

I would like to have people paste thier results up here with board and processor type

kgrach

Processor speed should not really make a speed difference if Drbombcrater is right.

But if it does then .....

@A3000T

OK What about these numbers:

AOne XE/G4 running at 1066 mHz 512 mb of TwinMOS unregistered ram

RAMspeed (UNIX) v2.3.0 by Rhett M. Hollander (Alasir Enterprises), 2002-04

4Gb per pass mode

INTEGER Copy: 169.75 Mb/s
INTEGER Scale: 170.60 Mb/s
INTEGER Add: 163.97 Mb/s
INTEGER Triad: 178.81 Mb/s
---
INTEGER AVERAGE: 170.78 Mb/s

RAMspeed (UNIX) v2.3.0 by Rhett M. Hollander (Alasir Enterprises), 2002-04

4Gb per pass mode

FL-POINT Copy: 193.48 Mb/s
FL-POINT Scale: 162.48 Mb/s
FL-POINT Add: 157.42 Mb/s
FL-POINT Triad: 159.96 Mb/s
---
FL-POINT AVERAGE: 168.33 Mb/s



Btw, was running DNet client at the same time but everything seemed to stop during the benchmark. Even time stood still!

@mlehto

I use the command -b 3 and -b 6

I doubt I have a secondrun XE as I seem to have the debounce problem with USB.

Last edited by Steff on 12-Jan-2005 at 11:58 PM.
Last edited by Steff on 12-Jan-2005 at 11:41 PM.

_________________
Fixed A1G4XE 7455 RX933PC with fried CPU
Sapphire Radeon 9100 128mb
ESI Juli@ 24bit 192kHz Envy24HT
Sil 680 Ultra Ata 133 E-ide
SeaGate Barracuda 120gb 8mb cache

@kgrach
Quota
I made the assumption that is was hiting the SDRAM directly.
Thats why I hate programs like that.


It makes comparing of results bit difficult. Here start to be couple of different processors with different L1/L2 and even with huge mount of L3 cache.



Quote:
also It looks like Uboot does read the SPD as different SDRAM results in different speeds.


Can't follow ... ??


Quote:
I would like to have people paste thier results up here with board and processor type


It's started allready, but we need more of them anyway. Info included with memory type (reg./non reg.) and used command string. And someone said, that there is made modifications to ArticiaS, wich are used for last batch of XE and uA1 (?). Do it affect speed of memory, don't know. So it may be better in XE:s to say, are they from first or second production round. That we have something, what we can compare Loads of variables.




Quote:
Processor speed should not really make a speed difference if Drbombcrater is right.
But if it does then .....


Look message #235 , It is made with G4/unreg. memory and is with better results. What is needed, is more info.


Miikka