Olimex A10S/A20-OLinuXino boards quite BBB-like

Morgaine Dinova wrote:

 

The datasheet did remind me though of a point you mentioned some time back, that the i.MX6's implementation of gigabit Ethernet is constrained by  a design fault which limits its maximum throughput to 400-470Mbps. 

That's supposedly improved somewhat in more recent steppings, but I agree it's not good.  It doesn't help that seemingly that limitation is aggregate of transmit+receive throughput.

Using PCIe may help, or not. A lot of the desktop gigabit add-in cards (PCIe x1) have a reputation for not being able to reach Gigabit speeds and aparrently this is due to the PCIe connection using small packets and the additional latency this causes. Indeed I have a two port card that uses PCIe x4 obviously in some attempt to mitigate the problem.  The iMX6 only has a single lane PCIe. So you can't expect wonderful performance, especially from a multi-port card.

Solve those issues and you'll still have to work out if the CPU, Memory or other busses within the SoC are up to the task - it's not so long ago that high end servers had problems with gigabit ethernet due to simply not enough cpu horsepower, badly arranged bus topologies etc. They're now having similar sorts of issues with 10Gb ethernet. 

I'd be very surprised if the same sort of issues don't arise for Arm devices. The iMX6 issue is likely just one aspect of this overall problem.  I'm also sure that, in time, there will be a device that's positioned as a communications processor that resolves them. Whether those sorts of SoCs ever appear in a BBB like device is a very different question.

selsinork wrote:

 

I just wonder if there's something more to it, was John right in that other thread suggesting you need two ports ?

Well the only requirement is a physical path to get frames from one EtherCAT device to another --- any physical topology that links masters to slaves will do.  Having two NICs allows you to cascade boards without additional glue hardware so there's a clear cost benefit, but you still have a couple of options available if your boards have only a single NIC.  One is to use them only in pairs, since the start and end of a chain of EtherCAT devices need only one NIC, although that does limit your range of application.  And the other is to use switches configured to replicate traffic on multiple ports to do all the routing between single-NIC EtherCAT boards.  Port replication is available quite widely even on low-end managed switches.

The ICE board also uses a different PHY from BBB, simple cost savings or something not supported ?

I don't know, but my understanding is that there is nothing special about PHYs used for EtherCAT.  As a physical layer device, you would expect it to just encode/decode the standard 802.3 frame without looking at the ethertype.  The choice of PHY is mainly determined by the interface provided by the host MAC, which is typically RMII on application SoCs to keep the pincount low.  The BBB's SMSC LAN8710 PHY supports both MII and RMII, and is supposedly compliant with the various 802.3 specs.

I expect it's all academic though, fairly clear that we simply don't know enough to implement it ourselves right now - at least not without a lot of help from someone who does.

It would indeed be nice to attract some expert on the subject here.  Seeing as Farnell sells so many BBBs for CircuitCo, how about someone asking Gerald Coley to pay us a short visit?  This is actually relevant to current product, because CircuitCo switched over to AM3358 on the original BeagleBone (white) with board revision A6A, so EtherCAT is now gone there, and undoubtedly will disappear soon on BBB as well.  I would like to know the physical reason that prevents the new SoC from supporting it.  It shouldn't be left as a mystery.

Morgaine.

PS.  And while we're at it, it would be good to know whether there is anything physically preventing single-NIC EtherCAT operation on AM3359-based BBBs, to avoid disappointment in case someone embarks on an EtherCAT project with expectations of fun.  There are a lot of AM3359-based BBBs out there, as well as A5A BBs.

It's Allwinner chips with everything at Olimex!

• Announcement from their blog.
• Allwinner chips in Olimex shop.

That's a very good idea from Olimex to sell the Allwinner devices directly.

Not only does it support their official position on open hardware by making it easier for people to acquire the necessary components for their own PCBs, but by selling more of the devices, Olimex gains access to more volume from Allwinner and so brings down their own product BOM costs too.  Good plan.

Morgaine Dinova wrote:

 

That's a very good idea from Olimex to sell the Allwinner devices directly.

Interesting, but I wonder how many sales of the BGA devices they'll make. Unlike the RPF, I don't presume that it's impossible for hobbiests to handle BGA's at home, but I do think that the number who will have invested the time and effort will be fairly low. Divide that by the number who will want these particular devices and what's left ?

The other avenue is if they can supply commercial volume, but it wouldn't be unreasonable to think that anyone buying in commercial volumes will either be able to go direct to Allwinner, or will get their outsourced Chinese subcontract manufacturer to just source them locally.

So good plan, yes. But I must be missing something for this to be much more than interesting PR.

Are there places that will just fit a supplied BGA package to a PCB, possibly as part of printing the board?

If so, I can understand enthusiasts in the West being happy to pay Olimex a bit more to avoid liaising with China themselves.

Morgaine Dinova wrote:

 

Are there places that will just fit a supplied BGA package to a PCB, possibly as part of printing the board?

I dare say there are. Most of the places I've looked at in the past fall into two broad categories:

1. PCB Only
2. Give us the gerbers & BOM we'll source everything and assemble it for you

So I'm sure there's some middle ground if you can find somewhere willing to deal with very low volume prototype stuff.  I've seen places that will supply a laser cut solderpaste stencil almost for free even when ordering relatively small quantities of PCB's.  It starts to become a question of whether the setup of the paste machine, pick&place machine & reflow oven is worth it for one-off or even 10 off type quantities though.

I do seem to remember reading about some newer solderpaste machines that are more like an inkjet printer, so you can selectively do only what's required.

From memory, pcbtrain (excellent manufacturer) will put on a one-off single SMD part (e.g. BGA) for £50, so not cheap, but not completely unaffordable either. I've been working on a reflow machine for over a year now! I might finish it one day. Pcbtrain can supply a metal mask, they sent a sample a while back, it looked very nice (it was steel).

shabaz wrote:

 

From memory, pcbtrain (excellent manufacturer) will put on a one-off single SMD part (e.g. BGA) for £50, so not cheap, but not completely unaffordable either.

Yeah, but whatever you're doing would need to be pretty special to bother. Added to the cost of a PCB and an A20 + PMIC from Olimex and you're likely already at least double the cost of a BBB or RPi.  As a way to prove a prototype for a much bigger production run then sure.

Sounds like there's a niche for a self-soldering BGA breakout board, containing an on-board heater and clamp.

Or a socket: http://www.selwyn.co.uk/index.php?option=com_k2&view=itemlist&task=category&id=13:bga-sockets-adapters

I can imagine those sockets are probably bga's themselves and I'd be surprised if they're not insanely expensive, so 50 quid to get a one-off done may not be so bad after all.