Hi Tarbos.

Thanks for the link to Sergei's test. Its very interesting!

I think you can not compare those two systems.
The IBM blade is not a normal computer.
The blade has two complete memory busses each with interleaved DIMS. So you need always to have 4 DIMS to be equipped in this board.
The IBM blade is a real server design while the FREESCALE design is more suited for something like a MAC MINI or a maybe a laptop.
I would have no problem with 3 GB/s
Both the new XEON and OPTERON servers that we got the other days reach just barely more than 2GB/s memory bandwith.

I think if these quite expensive server boards do not achieve more than 2GB/s then a theoretical limit of 3 GB in this desktop CPU shouldn't bother us.

Cheers
Gunnar

Arno,

You are right of course that the 970 is a very powerfull chip.
If I understand IBM's test on the JS21 blade correctly,
then they have achieved this very good result using a board with two 970FX CPU chips each with its own hypertransport ram bus. So the results of 7GB/sec was for both busses together.

A 4-core 970MP system does have same memory throughput as the above system - so with the same type of memory you will get the same result of around 7 GB/sec.


Sergei's test showed a result 1.5 MB per CPU core with 400 MHz memory. With 600 Mhz memory the 8641D should get around 2.1 GB/ per core.
In total chip does have 4.2 GB for both cores together.

Two 8641 CPUs having 4 cores and two memory busses like the IBM system will get a total of around 8 GB/sec.

I think the result of the Freescale chip is very good indeed.
So if you look at these number we see that a duo core G4 and a duo core 970 (G5) have about the same memory throughput.

There is no doubt that a 970MP kicks ass.
I think we can be sure that the freescale chip is very powerfull too. I think the 8641 is the same performance range of a similar clocked G5 or Intel duo core. So nothing to sneeze about.


I think 8641, 970, or the PAsemi they are all good chips.
The 970 is the most powerfull and hottests of them.

The question is not which chip to use but the we need to mind is that we need a system soon for the developers.

Without PPC system used by developers there will be NO more testing of the developers on PPC. Without testing and optimizing on PPC the software will run badly on PPC.
Without testing/developing the open source server on PPCs it will perform badly and maybe even run unbreliable on it.

If products like MySQL or Apache will perform dissapointing on PPC - then why should someone buy PPC hardware at all?

Lets say IBM might bing out a cheap consumer version of the Power7 in two years. Even if these chips will rocks - will people buy it?
No they won't!

The reason is simply that if we loose the developers now the software will perform badly or not nat all on PPC chips in the future.




Cheers
Gunnar

That is exactly what we tried to explain to Freescale and IBM executives for years. That said, there are folks at Freescale that are starting to really listen.

In a year from now, let's see what has happened...;-)

R&B :-)

_________________
http://bbrv.blogspot.com

Is POWER.org the right place to move this forward?
I think I should start a thread there.
Power.orgs mission is to address this, isn't this?


There are so many exciting PowerPC chips out there:
- Freescale 86xx are real nice.
- IBM 970 is still a hot one
- PAsemi chips look really promissing too.
They would make lovely laptops/entry server chips, wouldn't they?
- And IBM has the trio of super chips with: Power7, CELL, and Q7
I think I've read somewhere that there might be a stripped down version of the Power7 for blades and small servers.


So the hardware side is very promissing but unfortunately no machines for the developers.

I think in the end its the software support that will call a platform a winners or looser.
The "en vouge" software stacks: Linux, Apache, PHP, MySQL, (othersoftwareofyourchoice) - they need to have optimization for Power - if they do not have this then we can forget about PPC.

The problem right now some developers and companies start to see the linux as dead on PPC.

Gunnar

August 29, 2006

Hi Gunnar, if IBM does that it will probably be for a nextgen XBox. If that happens it will likely have a 970-related core, whereas, the PS3 and CELL is more designed/optimized for multi-threading and not suited for a desktop environment. In other words, if CELL has non-conflicting and non-overlapping threads it is great, but launch a spreadsheet program and you will find a G4 does better. It would be a smart move from the XBox folks if they sorted this out with IBM and made that happen. In the meanwhile, it seems like the 8641D is not moving ahead as we had hoped. Too bad!

R&B :-)

_________________
http://bbrv.blogspot.com

Thats interesting.
Where can I find this benchmark?
And can you point me to the test results for the 8641?

Cheers
Gunnar

We have 8641D systems here, but they are the version before the revisions just made and that is being shipped now by Avnet/EBV (we think they are finally shipping the new version). If you would like us to ship one to you so you could give it a test run, we would be more than willing to do that. Perhaps, you will be less disappointed than we were. The 7448 and 8641D have been development sinkholes for us. This is one reason why you never saw the 8641D Developer Program move ahead and the reason the Dual 7448/HDB/Pegasos III was never produced -- the chips ran too hot for the performance required (and advertised).



We don't think Power.org is the place for these discussions. Have a look at the Power.org Forums. They are not maintained and the leaders of the organization do not participate in them. Your latest "Lets face it we have a very big problem!" post won't change anything. If you count up the total number of IBM posts (not counting Site Admin who does his best) there might be 50. We are guessing there might be 10 Freescale posts. The CTO of Power.org (mpaczan) has made a total of 12 posts in since the Forums have been open (December 2005). They just don't understand or care about a real developer community. There are plenty of spam accounts. The poor guy who tries to make that Forum work does his best, but the support he gets is not much more than lip service. We know, as Genesi maintained the Power.org Forums until Jan/Feb 2007. There is nothing substantial happening there. It is too bad, the concept of Power.org is good, but the people managing it don't have a clue about community, developers or what Power should do next to be relevant. Sorry to be so harsh. We wasted a lot of time, effort and money in Power.org. It needs leadership and so does the architecture if it is going to be as successful as it can be. We would love to see a vibrant, successful Power.org, but we don't think enough changes can be made to make that happen any time soon.

R&B :)

_________________
http://bbrv.blogspot.com

This is the system you are looking for:



R&B :)

_________________
http://bbrv.blogspot.com

I'm not BadAndy ;-)
For further informations you have to ask him :-)
Doesn't make sense. NewerTech already released 7448 upgrades.

If you need to present comprehensive cost, efficiency and performance analysis of a product and it shows that all in all, it draws too much power, needs a heatsink like a Pentium 4, and doesn't run nearly as fast, you aren't going to really justify that product's existence in the world.

Tapping into the Mac fanboy market who refuse to get rid of their Sawtooth tower, isn't the same thing. They'll buy a dual 7448 upgrade because they have no other choice.

What you have to look at, is the same question we looked at with the Pegasos - if we made a 1.4GHz 7447A card or a 1.6GHz 7448 card, would spending $500 on a CPU upgrade net them $500 worth of performance? Consider most people paid $500 for the board in the first place, is doubling the overall cost of the system worth a 50% speed increase, 80% power consumption increase, 200% noise increase?

I doubt NewerTech, or Sonnet, or any of the Mac Fanboy suppliers sell very many systems at all. They can buy 7448's in quantities of 2000 or so, and we know they were taking rerated chips (1.2GHz and 1.4GHz processors which when given 0.2V extra voltage, didn't burn up, and were rebadged as 1.6GHz and 1.8GHz chips with 75C junction maximums and a 4-year lifespan..)

The expectation is the CPU card will last far longer than the Mac it's installed in.

We would need to make a production system, from scratch, that lasted a normal product lifetime - for a blade server this might be the 10 years most Freescale chips are rated for - and be useful for all of those 10 years, with a decent cost ratio for each of those years. We *can* justify the HDB system on a density vs. other systems, we *can* justify low-end chips with easy cooling there if you go for very very high numbers of systems.

Rather than buy 5 IBM JS20 racks full of blades, you could buy 11 Genesi racks full of blades, for the same space, and 7x the aggregate performance, for about twice the price, IF you had the ability to use that aggregate performance through clustering. The number of applications that it works in are very few, and none of them apply to people who read this forum and want a developer box :)

It certainly doesn't work on a single board on a single desktop for a single developer.

_________________
Matt Sealey

Okay, I see. But I think heat wasn't the issue. Maybe not suitable at 1.8GHz for the PIII, I think.

No offence!

>I have ars testbench 2.0 data now for MPC8641D, at 1.8 Ghz.
>On almost everything it is outrunning the dual (not dualcore) 1.8 970.

According to this paper the 1.8 GHz version was a shaming 70 W monster.
Motorola's original target for a dualcore G4 was the 15-25 W typical power range.
That's why they don't provide a 1.8 GHz version.


@Gunnar:

>The blade has two complete memory busses each with interleaved DIMS. So you need always to have 4 DIMS to be equipped in this board.

To be precise, it has 128 Bit memory access, so a memory feature consists of (just) two DIMMs.

>If I understand IBM's test on the JS21 blade correctly, then they have achieved this very good result using a board with two 970FX CPU chips

970MP (only one core alive)

>each with its own hypertransport ram bus. So the results of 7GB/sec was for both busses together.

Yes, only it's not Hypertransport but IBM's Elastic Interfaces with plenty of bandwidth to the CPC945 northbridge
while Freescale limits throughput by using a single MPX bus for two G4 cores (with a single load/store unit each).
So you could say the IBM solution is RAM bandwidth limited while Freescale has a FSB limitation.
Supposedly the Power.org G5 devmachine can use DDR2-667, up from JS21's DDR2-533.

>Sergei's test showed a result 1.5 MB per CPU core with 400 MHz memory.

And just over 2 GB/s for both cores - they don't scale 100% for reasons mentioned above.
According to your site an iBook with 142 MHZ FSB exhibits 958 MB/s already...

>Two 8641 CPUs having 4 cores and two memory busses like the IBM system will get a total of around 8 GB/sec.

This would not be a desktop system anymore. Two 8641 CPUs are not meant to play with each other like Opterons.


>And IBM has the trio of super chips with: Power7

Isn't Power7 a 2010 GA chip? :-)

I'm not talking about the 8641D ;-)

We will ship Gunnar one of our 8641D based systems. He has some interesting plans for it and we wish him the best on his Project.

R&B :)

_________________
http://bbrv.blogspot.com