Raspberry Pi Compute Module 4 Coming Next Year with PCIe/NVMe Support

The Raspberry Pi 4 features a Broadcom BCM2711 processor with Cortex-A72/A53 cores that makes it competitive against other SBC’s based on Rockchip RK3399 or Amlogic S922X/A311D processors. The new processor also happens to have a PCIe interface, but it’s not exposed on the board since the interface is used by VIA VL805 PCIe USB 3.0 controller for the four USB 3.0 ports on the little computer.

At the time, we expected the PCIe interface to be exposed in future Raspberry Pi boards or a Raspberry Pi Compute Module 4. But some people did not feel like waiting and instead hack the Raspberry Pi 4 SBC to access the PCIe interface. That hack is clearly not for everyone though…

“Artist” rendition of a Raspberry Pi Compute Module 4 – Not the actual thing…

But based on a recent interview of Eben Upton with Tom’s Hardware we got pretty much confirmation that the Raspberry Pi Compute Module 4 will be coming next year, and expose PCIe signals through the edge connector.

The Raspberry Pi Compute Module, CM4, we will support NVMe to some degree on that, because of course, it has a PCI Express channel.

…We have a single lane Gen 2 which is used to supply USB 3 on the Raspberry Pi. On the module that would be exposed to the edge connector and we’re likely to support NVMe over that..

PCIe will be exposed, but it looks like support for NVMe storage is still to be determined. Also don’t expect to be able to connect a PCIe graphics card since most PCIe implementations in Arm processors would not support it due to limitations such as maximum addressable memory as is the case for Rockchip RK3399.

Tom’s hardware further asked where a future Raspberry Pi board might have PCIe support via an M.2 socket for example. Based on Eben answer it will most likely not happen:

In the core product, maybe a future Raspberry Pi may pick up something, but it is challenging. It is not without cost, both in terms of the silicon and in terms of the connector, as the connectors are not free. Also providing board area for the connectors, if you look at the Raspberry Pi there’s obviously not room for an M.2 slot”

…What would a 4A look like? We have a decision to make about what we do with the USB 2. The USB 2 on Pi 4 is brought to the power jack, the OTG is brought to the power jack so we’d have a question in our minds about whether we undid that, probably, as we do on other A boards [and] brought it to the right-hand edge as a single connector”

You can watch the full interview in the video below.

Support CNX Software - Donate via PayPal or become a Patron on Patreon
Advertisements
Subscribe
Notify of
guest
25 Comments
oldest
newest most voted
tkaiser
tkaiser
17 days ago

Glad they expose PCIe on the connector and do not waste it for USB3.

Now I really hope someone will design a base board for the CM4 with a PCIe attached SATA controller for 2, 4 or even 5 SATA ports so something like proposed a year ago could become reality: https://forum.khadas.com/t/vims-proposal-vim3-nas-server-variant/4287/5?u=tkaiser

A stackable inexpensive server/NAS design taking care of heat dissipation and powering demands of connected 3.5″ HDDs (stable 5V/12V needed).

back2future
back2future
16 days ago

PCIe 2.0 (2007), x1, 5Gpbs, line code 8b/10b, burst rate 0.5GB/s?

(Low power demand is a plus)

itchy n scratchy
itchy n scratchy
16 days ago

What a catastrophe! after all the ethernet port completely saturates the single pcie lane /irony

back2future
back2future
16 days ago

Thx, just another consumer side perspective: pcie x1 lane is 36 pins, usb 3.2 24 pins (3.0: 9 or up to 11 pins) and hubs (or hub ic) are available (compared to PCIe switches on x16 slots for bifurcation)
But I’ld like to have a PCIe 3.0 x4 on armv9, in-order cpu SoCs, lpDDR5 for rpi4 price 😉 (on that system, 1Gbit ethernet phy would be acceptable) /irony_2020

itchy n scratchy
itchy n scratchy
16 days ago

Sorry that’s above my head… How is pin count connected with consumer perspective? And why does a 4xpcie gen 3 make a difference for a nas type of use case when there is only geth?

Looking forward to being enlightened 😉

back2future
back2future
16 days ago

I’d say it’s a difference soldering and debugging up to 36 (for 1 lane of pcie or external pcie?) compared to 9 joints for non-professionals/consumers and there’s no reason for only that standard gEth with 4 lanes rd/wr 985MB/s. (and/or wifi 5 version for rpi4 rev) If that’s not within Your taste, we need a new common vision 😉 (Interesting, considering mating cycles that vary from a maybe 200-250 for pcie slots, 250 for external pcie x4-x16, 1500 external pcie x1, standard usb2.0-a 1500, mini-usb gt 5000, usb-micro & usb-c gt 10000 and usb3.0-a rated 20000. And count me in… Read more »

willy
willy
16 days ago

No, an ethernet port is only 1 Gbps in each direction, or 25% of the theorical peak BW. In practice it’s always a bit more with small packets due to multiple bus requests needed to fetch ring descriptors but you’re still far below. You can actually plug a dual-port chip such as the I350-AM2 which only needs one lane to saturate 2 ports. IIRC there’s also an I350-AM4 for 4 ports but it will likely be a bit limited with a single lane. But that could make a good inexpensive dual-gig solution.

itchy n scratchy
itchy n scratchy
16 days ago

Sorry now I’m confused…

I was wondering why a single lane of pcie is a bottleneck in a nas for the storage, if on the other side the gig eth is limiting the traffic.

I wasn’t talking about adding more NICs to pcie.

Back in time with my ThinkPad t40p ide/HDD was the bottle neck at about 1/4 the ethernet speed, but nowadays?

Willy
Willy
16 days ago

Then the single PCIe lane is not the limiting factor because as you say the GigE already is. And since a single PCIe lane is more than twice the GigE speed, you can indeed saturate GigE with SATA traffic on the same PCIe lane without saturating the PCIe. In addition, most often you’ll transfer in a single direction, which means that GigE will receive when SATA sends, and conversely, so you have even more room. Ideally, you’d install a PCIe-PCIe bridge, and a SATA controller and a 2.5G NIC. It should remain possible to saturate that NIC with unidirectional SATA… Read more »

itchy n scratchy
itchy n scratchy
16 days ago

It’s a pity there aree no more reasonably priced A20 boards on the market. 1GB Eth and SATA perfect for a mini NAS. When the boards were available there was no proper OS support and now as the OS situation is mature, there are no more dirt cheap boards on the market.

tkaiser
tkaiser
16 days ago

> It’s a pity there aree no more reasonably priced A20 boards on the market.

A20 is 2 x Cortex-A7 at around 1 GHz. No, thanks. I had some hope when quad core Allwinner R40/V40 was available 4 years ago but it seems the Banana clowns are the only ones making R40/V40 boards (overpriced and limited by design flaws).

itchy n scratchy
itchy n scratchy
16 days ago

Of course it’s 2x A7 but to bring an old laptop hdd online that’s ok I guess.
R40 has no one working seriously on at sunxi and post lemaker bananas are a joke.

m][sko
16 days ago

It will be nice if raspberry pi foundation release compute module in same time as normal board

xnc-hardware
xnc-hardware
16 days ago

I like this, Every new feature like this will be carried/copied over to the chinese PI SBCs for us to have 🙂

Jon Smirl
16 days ago

The pricing on Rasp Pi compute modules is too high to allow any significant use of them. I bet this one will be $50. For $50 I can buy embedded motherboards that include x86 CPUs. So why bother with ARM at that price point. These things should be $20.

crossbar
16 days ago

$50? Who said that? The current CM3+/8G is about 30$ (farnell) I’m in no relation with raspF/T but the final CM4 would about 35$ (in very large quantities around 30$) Note: The engineering pre-samples flying around these days before release doesn’t have a real price point.

jshan
jshan
12 days ago

where can i buy $50 x86 compute module. would you leave a link pls ?

Jon Smirl
11 days ago

Here’s a random one…
https://detail.1688.com/offer/567645865493.html

It is much simpler to buy that Atomic Pi Jean-Luc linked.

tkaiser
tkaiser
11 days ago

Funnily an RPi 4 at default clockspeeds is already twice as fast as the J1800 board from your link. 🙂

I would opt for an J1900 at least (twice as much cores than the J1800 and only slightly lower single-threaded performance).

Jon Smirl
11 days ago

It was just the first random board I found under $50. I didn’t look very hard. It is better to spend about $60 for these. $50 gets you bottom of the barrel.

The general point here is that Rasp Pi compute modules are way overpriced. Likely $50 for the module and then you still have to add your own base board which is going to cost at least $10. At $60 x86 is a better value.

J1900 for ~$50
https://detail.1688.com/offer/554571782111.html
https://detail.1688.com/offer/573476003037.html
https://detail.1688.com/offer/603121966409.html
https://detail.1688.com/offer/549046313803.html

tkaiser
tkaiser
11 days ago

> Rasp Pi compute modules are way overpriced. Likely $50 for the module

Just checked it. I get the CM3 Lite for less than €30 (VAT excluded). So I would expect the CM4 in the same range. And of course this is way overpriced but with some specific use cases in mind most probably the least expensive offer.

Willy
Willy
6 days ago

Don’t forget that it will be very hard to cool these boards down. Ideally, sandwich-like heatsinks like those made for DRAM sticks should be used, but they are probably not sufficient to dissipate multiple watts.

Stuart Naylor
1 day ago

There is only you who keeps posting this fictitious $50.
Raspberry are strange in that there products are almost defined by price.

The Pi4-2gb is $35 and likely the Pi4-A will be aimed at the Pi3-A price of $25.
The compute module was always between the two as is more of a commercial product for integrators who buy in qty and likely negotiated.
Buts starts at £23.82 ex Vat likely as the current compute 3 models do.

Stuart Naylor
1 day ago

The Pi4A will be interesting as like Upton bang for $ for me its the best thing they do. Zero was great but that old Pi1 architecture often makes things an uncomfortable squeeze. Upton never seems to build up a product release and he starts talking about a product its often far more complete than he will say where often you just get an announcement its here. Compute model is interesting for some but the Pi4A is going to be interesting for the Pi3A+ as it likely to be in the same position a Pi3B+ is now when you can… Read more »

Advertisements