PI 4 Compute module with pci-e?

[red_mamba]

Oh, and another thought. One user pointed this out already and I have similar idea.

Few years back we say the rise (and fall ) of PCI expanders that were used to connect PCIe-x1 signals to mining GPU cards.
What if you simply make a board taller and add USB3 connector exposing PCIe signals.

There are already PCIe hubs out there that split signals to 3 or 4 PCIe-x4 connectors (with x1 signals) for people that want to use it.
But also USB3 kables can be used to create custom PCIe boards for any use case you desire

regards

[jamesh]

I like this discussion and I hope RPI foundation will listen to real world users.

We always listen. If I see stuff on here that I think is interesting and useful, I pass it up to the HW team. It's very rare that they haven't already considered anything I pass on.

James,

I had a plan for quite some time now to make rPi clustering easier for people to learn this sorts of things on a budget.
I'm working on CM3 version at the moment which is OS and can be seen and followed by anyone interested here: https://workspace.circuitmaker.com/proj ... M3-Cluster

The primary idea was to give users cheap way of testing and playing with docker/k8s and growing popularity of microservices.
But I see that 1GB of ram on rPi3 makes it unusable for any docker based OS. I tried balena OS and lack of ram is a big issue.

What I would like to see for CM4 is the ability to buy modules with different RAM sizes. 1, 2 and 4GB for example. Where for clustering 4GB would be the preferred choice and probably at least 8GB of storage, maybe 16Gb for OS.

The next thing that would be great to have is an easy way to route out signals for networking to connected it to a network hub that is integrated on the motherboard. I personaly thing that exposing PCIe signals should be given a serious thought as it gives users high speed access to SoC.
It doesn't matter what you connect to it at the end of the day as long as you have the choice to connect something

Also have you guys considered keeping the board size the same, but making PCB taller and maybe add gigabit LAN and USB chip on there and just route that signals to SODDIM connector. What ever can not be exposed directly can be exposed via board-2-board connectors. Which gives users the abillity to still connect extra monitor and stuff. I've seen SoM boards like this before. I just wouldn't go only b-2b connectors way as that eliminates the possibility of mounting the modules vertically and stack them tightly for clustering

my 51 sents

Obviously I cannot comment on any future CM4 features, but have passed on your comments.

[Gavinmc42]

I think someone said a Model Pi4A has been ruled out?
And a Compute4 is still ????

That PCI-e hack that uses a USB3 cable gave me an idea.
Why not a Model 4A but with a USB 3 connector connected to the PCI-e pins.
USB2 via the USB-C like the existing Pi4B.

Still keep the Model 3A+ footprint, but with PCI-e for those that want to play with that feature.
Also have the dual HDMI?

Would a 4A+ wih PCI-e be of more use than a footprint same CM4?

[trejan]

Why not a Model 4A but with a USB 3 connector connected to the PCI-e pins.
USB2 via the USB-C like the existing Pi4B.

You're asking for trouble with that configuration. It looks like USB 3 but it is actually PCIe and you need a non-standard adapter to use it. If you want to plug in your USB keyboard then you're going to need to power the Pi 4 from the GPIO header and use a USB A to USB-C adapter. Support nightmare.

[jamesh]

I think someone said a Model Pi4A has been ruled out?
And a Compute4 is still ????

Who said that?

[Gavinmc42]

I think someone said a Model Pi4A has been ruled out?
And a Compute4 is still ????

Who said that?

The decision was made that there will never be a Pi4A.

https://www.raspberrypi.org/forums/view ... 3&t=250303

Never say never?
A different connector than USB3 for the PCI-e?

[W. H. Heydt]

I think someone said a Model Pi4A has been ruled out?
And a Compute4 is still ????

Who said that?

The decision was made that there will never be a Pi4A.

https://www.raspberrypi.org/forums/view ... 3&t=250303

Never say never?
A different connector than USB3 for the PCI-e?

If you follow the links, it goes back to the announcement blog entry. It does *not* say that a Pi4A has been ruled out. What it does say is that they haven't identified a less-than-$35 way to do it yet.

It is my *personal* opinion that there may never be a Pi4A, and--for similar reasons--there may never be a CM4. I think both will be worked on, but there are significant challenges that may not make either one economically viable. I think a Pi5A and CM5 are more likely, since by the time those would be feasible the RPT will be a lot more comfortable working though the issues around the radical hardware changes that have come with the Pi4B.

What I would like to see is a revision to the Pi2B to give a "Pi2Bv1.4" as basically an underclocked (say, roughly 1GHz) Pi4B without WiFi and--perhaps--only one HDMI connector. I doubt there would be enough market demand to support such a beast, but if such a thing hit the market, I'd get at least 6 of them and probably more like 10.

[red_mamba]

We always listen. If I see stuff on here that I think is interesting and useful, I pass it up to the HW team. It's very rare that they haven't already considered anything I pass on.

James,

I had a plan for quite some time now to make rPi clustering easier for people to learn this sorts of things on a budget.
I'm working on CM3 version at the moment which is OS and can be seen and followed by anyone interested here: https://workspace.circuitmaker.com/proj ... M3-Cluster

The primary idea was to give users cheap way of testing and playing with docker/k8s and growing popularity of microservices.
But I see that 1GB of ram on rPi3 makes it unusable for any docker based OS. I tried balena OS and lack of ram is a big issue.

What I would like to see for CM4 is the ability to buy modules with different RAM sizes. 1, 2 and 4GB for example. Where for clustering 4GB would be the preferred choice and probably at least 8GB of storage, maybe 16Gb for OS.

The next thing that would be great to have is an easy way to route out signals for networking to connected it to a network hub that is integrated on the motherboard. I personaly thing that exposing PCIe signals should be given a serious thought as it gives users high speed access to SoC.
It doesn't matter what you connect to it at the end of the day as long as you have the choice to connect something

Also have you guys considered keeping the board size the same, but making PCB taller and maybe add gigabit LAN and USB chip on there and just route that signals to SODDIM connector. What ever can not be exposed directly can be exposed via board-2-board connectors. Which gives users the abillity to still connect extra monitor and stuff. I've seen SoM boards like this before. I just wouldn't go only b-2b connectors way as that eliminates the possibility of mounting the modules vertically and stack them tightly for clustering

my 51 sents

Obviously I cannot comment on any future CM4 features, but have passed on your comments.

James,

can you also point them to SMARC 2.0 standard

https://sget.org/standards/smarc/

[pallavaggarwal]

If ever Raspberry Pi foundation wants to build a Pi4 computer module, I would suggest the Stamp type form factor as it will allow engineers to build their boards quite easily without the need for an extra connector.

Also, keeping the power supply need to a minimum. The module should take 3.3V or 5V voltage input to power the module and rest needs to be on the module itself.

I am looking forward to a Raspberry Pi 4 Compute Module.

Pallav Aggarwal
https://pallavaggarwal.in/learn-about-raspberry-pi/

[krazykyran]

Thumbs up to the SMARC format.
I work for a company that uses the CM3 and CM3+ modules using our own board design (that I had a big part in).
I appreciate the low cost of the Pi when I started on this endeavor, however there are features I'd really like to add that I just can't.
PCIe is one of them. On SOC ethernet is another. Multiple (>2) UART ports is another.
I am excited about the Pi 4, and can't wait for the compute module format. I know it will never work in a SODIMM format, there just arent enough pins. Also as the Pi 4 is not backwards compatible, it would make sense to use a completely different connector.
SMARC already supports these extra features.
and .... you can buy SMARC boards to use it now (just google SMARC boards and lots of suppliers show up).

[nomdediot]

Haven t read all the thread but just to add some pieces :
- industry will need a solution when cm3 will not be available any more
- size of sodimm200 module is great for many small designs where the available space is limited
- current compute module has NC pins and also optional SDx pins

Maybe 2 new compute modules can be created to adress multiple usage :
- first version with the current sodimm200 connector and pcie lanes added on NC pins, because the current gpios available are enough for many usages but pcie is cool to enlarge possibilities
- another version with a larger connector to have more peripheral available (spi, i2c etc). I also agree that SMARC format is great for large applications

[mc_hof]

As I already pointed out in a previous post, SMARC is too big in many applications and migration of existing products from CM3 (SODIMM200) to SMARC would be impossible. Also, there is no need to have a larger form factor. Here is why:

As I said, it would be perfectly possible to have another 60 contacts while keeping the same mechanical form factor if they would go for a 260 contact DDR4 SODIMM socket like https://www.te.com/usa-en/product-2309413-1.html.

If you arrange the 60 additional signals intelligently on the SODIMM260, you could minimize the effort to migrate existing designs from CM3 (SODIMM200) to CM4 (SODIMM260). Proposal: I would add a pair of additional signals every 6 ... 7 contact pairs, so that the end of the existing traces only have to be moved by a mm or less, after having replaced the SODIMM200 by a SODIMM260 connector. Any medium experienced PCB designer can reattach the existing signals in a matter of some hours.

The invest to do this is really low. And I think that there will be enough customers for both versions (SODIMM200 and SODIMM260) to justify two versions.

[jamesh]

The invest to do this is really low. And I think that there will be enough customers for both versions (SODIMM200 and SODIMM260) to justify two versions.

Almost certainly not.

Anyway, horse has bolted on this one.

[nomdediot]

mc_hof, I think you do not consider all the costs this can introduce. Cm are for the industry. Changing the pcb is not a long job with a similar connector.

But then
- stock of previous version is lost
- testing tools are to be reworked
- new reference is to be managed
- maintenance and repair of previous product become impossible, customer disagree with that
.....

Anyway we agree I think two versions of CM can exist. One with sodimm200 and another one, sodimm260is nice too.

[mc_hof]

mc_hof, I think you do not consider all the costs this can introduce. Cm are for the industry. Changing the pcb is not a long job with a similar connector.

But then
- stock of previous version is lost
- testing tools are to be reworked
- new reference is to be managed
- maintenance and repair of previous product become impossible, customer disagree with that
.....

Anyway we agree I think two versions of CM can exist. One with sodimm200 and another one, sodimm260is nice too.

I completely agree. Maybe I din't point out clearly enough, that reworking an existing base PCB is only justified if the application really needs the additional features of a CM4 (PCIe, additional serial interfaces, ...). With a SODIMM260 module Raspberry could generate additional revenue since it could be used in applications where the designer otherwise would have to choose a different module with the desired features.

[jamesh]

Just so people don't keep writing stuff that will have no impact, I think I can say that a future CM4 form factor and connection system is already decided on. Cannot tell you any more than that, but suggestions made now are unlikely to be acted upon.

[aBUGSworstnightmare]

Just so people don't keep writing stuff that will have no impact, I think I can say that a future CM4 form factor and connection system is already decided on. Cannot tell you any more than that, but suggestions made now are unlikely to be acted upon.

Can this be read like: YES! There will be a CM4 compute module?

[fruitoftheloom]

Just so people don't keep writing stuff that will have no impact, I think I can say that a future CM4 form factor and connection system is already decided on. Cannot tell you any more than that, but suggestions made now are unlikely to be acted upon.

Can this be read like: YES! There will be a CM4 compute module?

Very unlikely to be a definite YES..

The CM2 never materialised even though it was "expected"

[Pansy]

Just so people don't keep writing stuff that will have no impact, I think I can say that a future CM4 form factor and connection system is already decided on. Cannot tell you any more than that, but suggestions made now are unlikely to be acted upon.

Thank you

[MysticWhiteDragon]

Just so people don't keep writing stuff that will have no impact, I think I can say that a future CM4 form factor and connection system is already decided on. Cannot tell you any more than that, but suggestions made now are unlikely to be acted upon.

That's great to hear that a CM4 may be possible.

Not enough pins for connections? How about 5093? Use this socket form for a CM.
https://www.anandtech.com/show/14182/ha ... ckage-ever

[jamesh]

Just so people don't keep writing stuff that will have no impact, I think I can say that a future CM4 form factor and connection system is already decided on. Cannot tell you any more than that, but suggestions made now are unlikely to be acted upon.

That's great to hear that a CM4 may be possible.

Not enough pins for connections? How about 5093? Use this socket form for a CM.
https://www.anandtech.com/show/14182/ha ... ckage-ever

Too big. Much too big.

[trejan]

Not enough pins for connections? How about 5093? Use this socket form for a CM.
https://www.anandtech.com/show/14182/ha ... ckage-ever

FCBGA5903 isn't a socket...

[6by9]

Not enough pins for connections? How about 5093? Use this socket form for a CM.
https://www.anandtech.com/show/14182/ha ... ckage-ever

TDP for the 56-core version is 400W.

Wow! Who needs central heating?!

[crossbar]

TDP for the 56-core version is 400W.

We can do it differently
It may be off-topic ... but here is an open (!) source celerity design:

Approximately 500 RiscV cores (RV32IM, 1 GHz, 5-stage pipeline, in-oder, single issue, 4 kB IMem 1rw, 4 kB DMem 1rw) in a mesh many-core array (5 mm x 5 mm). Communication via partitioned global address space (PGAS)

Surprise, surprise: 500 cores only 2 (!) Watt (TSMC 16nm FFC)
(src: https://content.riscv.org/wp-content/up ... Taylor.pdf , slice 30)

Details: http://www.opencelerity.org
High-resolution layout from Celerity: http://bjump.org/manycore/celerity_chip_plot.jpg (ap. 50MByte)

mic_s
.

[tim292stro]

I'm for a DDR4 slot unit. I want to build a board for a multi-drive raid enclosure with enough RAM to do 1GB/4TB of HDD storage - either with 4x SATA-3 or 4x SATA-3 over USB 3.1, and three 1Gbe NICs. Dedicated PHY 1Gbe for management, and two 1Gbe over USB 3.1 for NAS traffic (8021.1aq SPD). Dual 12V inputs to run each shelf.

I had figured on doing this with Rev4 Model B 4GB units, but CM4 would be better in my opinion (due to available emmc). In a 6U case with 9-shelf boards in the front and 3 shelf boards in the back, one could cram 48x 3.5" drives, leaving space for two Mini-ITX server-grade systems in side-by-side, and 2x 1500Watt 12V power supplies.

Think of having each shelf with 4x drives being its own NAS - so 12x NAS'es in one enclosure - independent of the two Mini-ITX host motherboards. Would make a heck of a home front end - Linux KVM/QEMU running Open VSwitch, booting service VMs from Gluster on ZFS partitions hosting: Firewall, VPN, IDS/IPS, VoIP, Email, Web-hosting, etc... with redundancy and fail-over

If anyone wants to beat me to this idea and produce it - I'm in the market for 4x of those cases with NAS shelfs... That works out to 48x CM4s, plus one Rev4 Model B for the KVM LCD tray running an V.2 Onion Hidden Service for tunneling IPMI without exposing ports.