AMD's southbridge are buggy ? ahci and usb do not work well.

It did?
A-Link really is just PCI Express. It's not different enough (not like the ULi implementation which is just "quirky") to even give a name to it.. but this is what these people do these days.

There's not been a custom southbridge link outside of Intel for a long time now.
That's not been true since AMD's southbridge were AMDs and ATI's southbridges were in the 5xx series (the SB500 was completely broken, and it annoyed the hell out of several motherboard manufacturers who simply went out and used ULi chips instead - which fits neatly with the above, this is only possible because all these southbridges are simply PCI Express)

The AMD SB6xx series really was rock solid outside of extremely early versions with an odd RTC bug, and the SB7xx series are quite simply the best southbridges outside of Intel's. Regardless of anything else, AMD makes the *lowest power consumption* bridges, and publishes figures to prove it.

_________________
Matt Sealey

I found a little article about the SB800 series.

It's PCI Express 3.0, but AFAIK this is backwards compatible to previous implementations? Although more controllers on chip than earlier, I guess it won't exactly benefit from the PCI Express 1.0 (?) bandwidth of the 8640?

.. looking a little too far ahead aren't we.

One of the primary factors in picking components HAS to be availability or ease of development of drivers. If we can't get the datasheets or give them out, or the drivers are poor, it's a lot of work to bring up this support.

This is one of the reasons all the boards that 'want' to use AmigaOS 4 ended up with the exact same Silicon Image SATA chip :)

AMD's late M690T/E series are primed for embedded use right now, the SB600 southbridge component of which is pretty good. 700 series would be a good choice depending on price.. but I'm not sure if there are really any better features that are worth the move.

6xx/7xx support in Linux is pretty well maintained and actively developed. Support for MorphOS could probably be rolled out pretty quickly (and the vast majority of the driver support - USB, AHCI SATA, etc. would be portable to any other chipset).

_________________
Matt Sealey

Hmm, yeah, the SB600 seems like a good choice (and it seems to be PCI-E 1.x as well). This one coupled with the 8640 would make a great machine!

So what's stopping you? ;-)

We'd need more customers than just you alone, and some development funding to do the design in the first place.

Right now we're concentrating on the low-end low-power thin client (Efika style) and the potential for a MPC8610 Netbook.

The MPC8640D may fit into the Netbook category or it could be, like the MPC8610 too, a candidate for a higher performance desktop model, if we can convince anyone but between ourselves that it's an awesome idea.

For now we're concentrating on the Netbook idea, not the high-end desktop.

_________________
Matt Sealey

Yup. That's what Uli told me anyway. back when we were all excited about 8641d + M1573 (before 1575 was out), uli's rep said M1573 might work connected direct to one of the CPU PCI-E ports, but it would not work with a PCI-E switch. Alink didn't require some recognition logic or somesuch to make it _fully_ PCI-E compatible across a switch. AMD do not say anything abotu this at all, neither yes nor no. Considering M1575 was made to give this capability to their previous part, I'm assuming that something about Alink does not require something that full standard PCI-Express does.

The only thing I can find in AMD's datasheet for SB600 (can be found on their web site publically, I can't find anything for SB7xx anywhere) the pin names look like PCI-Express pin names, but other than that there's no mention of the thing working with anything different than their northbridges. Considering that Uli made it sound like they added to what Alink requires to make it a full PCI-E standard product, I'm hoping to hear of someone actually doing this with an AMD part.

If it truly is vanilla PCI-Express, why do they feel the need to play the rename game and call it Alink Express something instead of just PCI-Express?

And they fixed it for M1575, but it still has a weird "southbridge mode" built-in which.. I guess is there so boards come out just as quirky as with the M1573?
I had a small discussion with someone about 2 years ago? And they confirmed it should act like a PCI Express endpoint if that is all you have.

Anecdotal evidence (including a little hint on a Wikipedia article) suggests that the Northbridge A-Link II Express on AMD chipsets is essentially a PCI Express x4 link and will accept any PCI Express chipset or even chip (singular). This is backed up since when the original ATI SB500 (I think?) was completely broken on release and delayed a number of motherboards, they used the M1575 instead on the ATI northbridge.

Via also didn't talk about their old southbridges (VT823x), but if you look at the implementation, it's just PCI (it wouldn't work on the Pegasos if it wasn't. They since remarketed it as "V-Link" for some odd reason)
SB7xx is a bit too recent (and still in use on mainstream motherboards) to be dropping public datasheets out in the wild, it might affect partner's marketing and suchlike.

You're right, though, they don't like to talk about it. Check the factsheet I linked above though, you'll note the southbridge has a little arrow for the connection that says "PCIe". Big hint, I think. The SB diagram doesn't even mention A-Link.
Marketing is very, very powerful. If someone says they have V-Link Advanced Connectivity, you would want to say you have something just as "special".

There's also the possibility that

* A-Link possibly doesn't scale - maybe it's fixed to x4 lanes and doesn't adapt like PCI Express should if you use different number of lanes to connect the chip up.

* A-Link possibly eschews parts of the PCI Express specification to simplify the design, perhaps fixing some features that are usually configurable. But as a link between a root complex and an endpoint, works fine. You can't market it as PCI Express, then, but you can use it that way :D

I'm actually surprised AMD don't use HyperTransport for it these days. But while they're using "A-Link Express II", the evidence suggests that it's just PCI Express x4 with some minor differences.

_________________
Matt Sealey

If not for hardware reasons, then obviously for marketing reasons. They want customers to buy the northbridges and southbridges in pairs. A custom name chains them together, one "A-link" in one end and another "A-link" in the other end. This in order to try to prevent customers from choosing one NB and then a SB from a different manufacturer, like in Matt's example with the SB500/M1575.

The same with the VIA, their NB/SB products "should" (according to the seller) be bought in pairs and connected to each other through the "V-link". That didn't stop the Marvell II of the Pegasos from using a "V-link" SB though.

The IEEE 1394 interface is also known under various brands (FireWire, i.Link, Lynx) for marketing reasons. Apple customers buys FireWire products to connect to it, Sony customers buys i.Link products, etc.


Looking at the spec sheet of the 600 chipset (the "Root Complex" bar in the NB part), it seems like the 690 NB is very similar to the "NB" inside the 8641/8640 chips in the sense that they have a total of 16 PCI-Express v1.x lanes. 8 lanes goes to the PCI-e x16 GFX slot, 4 lanes goes to the Southbridge, and the last 4 lanes goes to whatever expansion slots/on board controllers you choose to use in your design.

This is just how things are set up on the 8641/8640 dev board BTW; 8 lanes to the 16x GFX slot, 4 lanes to the ULI 1575 Southbridge, and the remaining 4 lanes available through a slot in line with the 16x slot on the board, if I read things correctly.
The SB600 spec sheet together with that article about the SB800 suggests that:

"A-Link Express I" = 4x PCI-Express v1
"A-Link Express II" = 4x PCI-Express v2
"A-Link Express III" = 4x PCI-Express v3

At least if I interpret things correctly.

It would be interesting to see more evidence of different NB/SB combinations in motherboards like the SB500/M1575 examples...

It's more complicated than that. You can't "split" the 8-lane controllers coming out of the MPC86xx - they are either x1, x2, x4 or x8, but you can't have two x4 out of an x8. Since PCI Express is point to point (unlike PCI which is a multidrop bus), that means you get two "chips" and if those two chips are x1, then you are wasting 14 lanes.

The remaining 4 lanes on the "southbridge port" to the M1575 are for a weird byte-reversed mode (a magic feature of the Freescale SerDes) where it disables the southbridge (an onboard FPGA handles this), then you get an 8x slot that connects backwards for debugging and you can even configure it as RapidIO and build a $6,000 sample board to run stuff off it :)

Basically if you go PCI Express on the MPC8640D you plug x8 into the southbridge (but only use x4), x8 into the graphics port, and now you've run out of PCI Express. Luckily there is normal PCI.. so a southbridge and graphics is your only choice, *or* you could have x16 graphics (with x8 connected), x8 pci express slot, and all the peripherals on PCI instead (USB, SATA, Ethernet, whatever) for user expansion - using a PCI southbridge (which usually has more PCI slots bridged off it) or discrete chips (it's only really USB, SATA and Ethernet the thing needs..)

_________________
Matt Sealey

Ah, OK, but it would be worth it IMHO. In the SB600 there are a lot of bandwidth extensive controllers (PATA, 10x USB2, 4x SATA2), and it also offers up to 6 regular PCI 2.3 slots/controllers as well, so I think you can live with that! :-)

The SB600 looks OK really, especially if it's reasonably priced as you say, and it's well supported. The only little quirk is perhaps the AC97 instead of HD Audio, but I guess people that treasures audio quality can always add some real sound to it through PCI (as long as there are slots available), so...
Well, the 8641/8640 has 4x Gigabit Ethernet on chip, so you won't need that at least... :-)

Are they quirky? How so? I guess I always assumed that any special "southbridge mode" would be some weird thing not standard to PCI-Express that makes it work as "Alink Express" instead. Alink move vs standard PCI-Express endpoint mode... I have a PC with an M1575, but am not aware of anything quirky about it. I don't do benchmarks or anything to see some of that. Plus, as I understand, M1575 has ethernet but the AMD SB does not. (newly added in SB800?) One of my crazy ideas would prefer as few components and as much integration as possible due to limited board area.
Cool. I take it you have more direct experience than some anecdote on wikipedia. :)
I didn't know that was a Vlink thing or that Vlink was PCI. Interesting to know. If all they market it as is Vlink or Alink or whatever, how does one go about determining that it is compatible with standard PCI or standard PCI-Express, or if it is some not fully compatible subset of them?

I remember distinctly some very confusing output from Freescale about changes they made to PCI Express from revision 1.0 to 2.0 of the MPC8641D, which hinted at the M1575 being quite odd. The concensus was "we will fix our PCI Express controller to act odd too" when it was also hinted that the M1575 could simply be put into a different mode.. which may or may not work... but who knows. PCI Express was completely busted when we were doing our development, and I have one on my desk right now which is of some fixed revision (an MPC8641D I mean).

If I ever get the damn thing booting (this is a time issue more than anything else, life keeps getting in the way of spending a couple days kicking a SUSE release onto it) it will become Genesi's second SUSE build service box to go with Peter's MPC8610. I'd love to check the performance on it with some real apps.. the example Linux was frighteningly fast.
This is not too bad in the end, it means you're free to pick any ethernet and phy you want (especially, one you already have a driver for :)
Put it on a board and find out. I dunno. I find it hard to believe they would use PCI Express pin namings but make it totally incompatible, use "PCIe" as the link out of the southbridge on a fact sheet (rather than the marketing term..)

In any case if we ever go down this path and need a southbridge, I'd say AMD is the best way to go (since ULI is dead and nVidia have "issues" with people who want to make less than 100,000 systems) if not discrete chips. Since the AMD southbridge supports all the SMI, temperature monitors and even power management stuff that would be exposed through the MPC8610, plus the AMD southbridge is actually available, ready for embedded use, fully documented in an open fashion (try that for the M1575 or VT8231) - and if we're telling them we'll buy 1,000 chips in 6 weeks, they are sure to tell us "it is just PCI Express".

The only thing more exciting than using an AMD southbridge that's less than a year old (at least from the point of view of the last PC motherboard to use it before the SB7xx came out) and still very well supported would be dropping an r500-class Mobility Radeon on there. Little point for an MPC8610 Netbook but for the dual core, it'd need some graphics, and r500 is where all the Linux development and AMD documentation is right now (and practically the lowest PCI Express graphics solution that AMD would support with a binary driver if that was necessary).

On a marketing side of things, using an AMD southbridge means we can be "ATI certified" again, just like the ODW was ATI certified.

_________________
Matt Sealey

Thank you Neko for all the insight. Takemehomegrandma had posted on amiga.org I think the idea of using an AMD southbridge like this and my quest for more info brought me to this thread here. I'm looking forward to hearing if you guys find a use for it on PowerPC, and I'm very very interested in the 8610 netbook as well.

No problem. I found another thing too; if you check the AMD SB600 Databook (the one where it describes pinouts and things) and look at the example diagram of the "Albacore" reference design (page 12), there, too, it says the link between Northbridge and Southbridge is PCI Express x4.

Nevertheless.. on page 52 for the USB architecture it says A-Link Express again. Confusing.
.. we're actively looking into it as of today, first order of business is fix this confusion.

_________________
Matt Sealey