Pegasos G4 Upgrade

[bloodline]

I always wondered about CPU slot architecture, if you actually get the same performance out of it as if it were surface mounted.

AFAIK, Slots tend to be more (electrically) noisy than sockets.
This means you have to run the memory slower to keep it reliable. The best option is to include DIMM slots on the CPU card.

How much of this applies to the current Peg design I do not know. I am unaware of any drawbacks in the Peg design.  ;-)

[magnetic]

THere is one drawback to Peg design: USE OF ARTICIA CHIPS !!! Though that wasnt their fault... now we know though.. Peg2 will be great.
mag

[MiniBobF]

THere is one drawback to Peg design: USE OF ARTICIA CHIPS !!! Though that wasnt their fault... now we know though.. Peg2 will be great.

The only drawback of the Pegasos I design is... the DESIGN of the Pegasos I. The chipset incorporated into the Pegasos design works okay on other boards (and no, I'm not refering to AmigaOne/Teron - OTHER boards). The problem with the Pegasos is a layout, termination or signal integrity issue. That is an artwork issue, i.e the design of the Pegasos PCB. Granted though, perhaps the chipset is not so tollerant of poorly formed waveforms as perhaps more mainstream PC chipsets.

From the TECHNICAL standpoint, does anyone know what the April interposer does? Answers like "It makes it work", "it makes it work better than Teron", "there's no Mai without April" or "It doesn't matter because Pegasos II will replace it" are not considered concise or technical.

Neil Thomas, AKA MiniBobF

[Dietmar]

only drawback of the Pegasos I design is... the DESIGN of the Pegasos

I have to admit I'm very impressed by your ability to diagnose systems you have not designed, out of your armchair, not even knowing what bplan's fix fixes. As to MAI's Arctica, it seems to work reasonably well in MAI's evaluation boards which should not surprise anybody. It also works well in designs derived from those boards. Big surprise. But it did not work well in the independendly designed Pegasos before fixes were added. There are at least two conclusions: Pegasos is poorly designed. Your conclusion. Or: the Arctica does not meet its specifications. I don't design boards (obviously) but, generally speaking, assume that boards are not designed by trial and error but based on the chips' specifications. If that is the case, the Arctica may very well be full of problems that just did not show up with MAI/Teron-derived boards. It will be interesting to see how the Marvell chipset works in a bplan design: it's better spec'ed and should require more design skills than an Arctica-based board.

[Dietmar]

> AFAIK, Slots tend to be more (electrically) noisy than sockets. This means you have to run the memory slower to keep it reliable.

On the purely logical level, this statement makes no sense: noisy it may be, whatever that is, but if the memory has to be slowed down or not would depend on the tolerance of the design for "noise".

On a purely empirical level, this statement obviously makes no sense, too: slot boards were the dominating design for a while and those boards did not run memory slower to keep it reliable (as far as I can remember).

[MiniBobF]

I have to admit I'm very impressed by your ability to diagnose systems you have not designed, out of your armchair, not even knowing what bplan's fix fixes. As to MAI's Arctica, it seems to work reasonably well in MAI's evaluation boards which should not surprise anybody. It also works well in designs derived from those boards. Big surprise. But it did not work well in the independendly designed Pegasos before fixes were added.

Pegasos is not singularly the only independantly designed board, which is my point. Make a search on www.armada.ch and you will find their are other developers, independent of Mai, who produce marketable goods. What additional skills or information does that firm have that Genesi doesn't have? Probably no difference in access to information or skilled personnel. The difference is, they got it working, without blaming one of the few most important device in a system.

it's better spec'ed and should require more design skills than an Arctica-based board.

What do you mean by "more design skills"? I assumed also, that by Articia-based board, you refer specifically to Articia S? Articia P is of course similalry speced to Disco 2, but with the addition of AGP.

Neil Thomas, AKA MiniBobF

[Dietmar]

What do you mean by "more design skills"?

Compared to the existing Pegasos, just about everything is higher spec'ed: PCI-X rather than PCI, the FSB will be faster (183 MHz?), the memory bus will be faster (and DDR vs SDR), etc. I was just assuming that faster components require more design skills.

[Dietmar]

Make a search on www.armada.ch and you will find their are other developers, independent of Mai, who produce marketable goods

On a purely logical level, that does not mean that the Arctica is not the culprit. It is perfectly possibly that it does not meet its specs and that the designers you are referring to have run into the same brick wall. Maybe they just had more money and redesigned their boards to meet Artica's bugs instead of having a rather visible  (and annoyingly expensive) kludge directly under MAI's chip? They would have ended up with the same result: a working mainboard.

[MiniBobF]

Point noted.

I wonder though, I doubt they'll make use of PCI-X, even though the chip has it. PCI is 133MHZ, and MUST be point-to-point. This means only one device, and the bus is limited to 3 inches in length. It's not as simple and to just not fit a card, as the bus is then shaped like a T, and reflections are set up.

In addition to that, are there any PCI cards available in PCI-X yet? Maybe there are, but more server orientated, not for the desktop user! There are plenty of PMC cards that do PCI-X, but it's a different ball game.

I would hope they make use DDR and Gigabit ethernet functions. Most important factor with DDR is to match the clock and signals tracks as closely as possible. Not tracking other busses above or below is important too, but on a 4 layer board, the scope is limited. Or has the defacto PC motherboard moved to 6 layer yet? Probably not.

Providing the Northbridge is located close to the CPU and tracks are matched, running the CPU bus at the higher speen shouldn't be a problem.

I wouldn't be supprised if the bus is not run at that higher rate. Genesi have problems with the socketed interface (hence the need to sell the latest G4 'upgrades' as matched sets, so you have to ssend you Peg back again). These problems will be exaserbated by the higher clock rate. Matching track length over the legth of a connector is much harder if that connector is long and thin, rather than square.

Ever wondered why Intel dropped those slot designs?

Depending on characteristic impedance of the motherboard, the skew in clocks over, lets say 3 inches is about 1ns - and at 533Mhz (or whatever the current fastes PIV FSB is) is about half that time!

Easy to make a bus not work then!

Neil Thomas, AKA MiniBobF

[Dietmar]

>Ever wondered why Intel dropped those slot designs?

Maybe it's more interesting to ask why to start with a slot design? CPUs are flat. Going the extra mile and soldering a CPU to a vertically mounted card seems to defy reason. In the case of Intel, I recall problems with getting sufficient cache into the CPU. When that was fixed, they reverted back to cheaper sockets. In the case of Genesi, I suspect it is their inability to grow pins to the PPC, leaving them little choice: solder or slot. The latter obviously provides more flexibility.

[MiniBobF]

One can solder a BGA to a PCB, and mount the PCB is a socket.

Of course, to package on a ceramic plate, and add an array of pins to the bottom cost more than to produce a BGA. Additionally, it's easier to pack balls in tight and retain their stregth than it is to pack in a pin grid array.

The reasons for dumping slot form factor is more technical than that too. Imagine trying to mount 200g of aluminium to the side of a chip mounted in a slot - it's not going to work..

Neil Thomas, AKA MiniBobF

[Dietmar]

> Imagine trying to mount 200g of aluminium to the side of a chip mounted in a slot - it's not going to work...

That puts the name of this passive "Eliminator" heatsink into a whole new perspective :-)

Eliminator 0.5kg heatsink

>One can solder a BGA to a PCB, and mount the PCB is a socket.

What would still speak for the slot design rather than an adaptor PCB is that it's easier to add the stuff needed in direct vincinity of the CPU (ie. L3 cache on the backside, the capacitors), while keeping the micro-atx form factor.

[MiniBobF]

A BGA can still be mounted on a PCB and be attached to a mother board in a way other than linear contacts along one edge.

Much like the AmigaOne XE/TeronPx.

(Maybe it's not the PX, I can't remember)

This way, A BGA chip can be mounted on a motherboard, can still be removable, is parallel to the motherboard (so no drastic increase in profile) and does not have a long thin line of connections (but a grid).

Neil Thomas, AKA MiniBobF

[Dietmar]

A BGA can still be mounted on a PCB and be attached to a mother board in a way other than linear contacts along one edge.

With two CPUs? With L3 cache on the backside? While keeping the micro-atx form factor? I find it hard to imagine any other design but a slot to offer so much flexibility (but I haven't seen the mainboards you refer to).