Live data from Hacker News

Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

jeffgeerling.com

21–30 of 74 posts

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#21

Earlier quoted context omitted.

> If nothing else, there better be some decent PCIe available by then- 8x PCIe 5.0 or better. ARM needs to start doing bandwidth, some day (in places other than server class chips); it's been on USB3.0 for a long long time. Its not up to Arm to attach PCIe lanes but the chip makers who have been competing with x86 which already offer plenty of PCIe lanes. So far, only apple has bothered to make an Arm CPU which can c…

>So far, only apple has bothered to make an Arm CPU which can compete with "PC" hardware. Also Amazon: https://aws.amazon.com/ec2/graviton/

Those are server processors, I was thinking more along the lines of laptop/desktop performance. Other examples of Arm server CPU's are from Ampere and AMD's discontinued Opteron A series.

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#23
post #5

Alas PCIe switches were starting to get affordable before the great chipmaker consolidation of 2015-2019 ("and seemingly overnight the cost of these switches increased three-fold"[1]). Now most every PCIe switch is >$6/lane. Even just the 12 lanes here, on an old PCIe generation, is going to be a $72 chip, perhaps more (article: PEX8619 Broadcom PCIe switch, $129, 16 lanes, $8/lane, no stock. Compare versus Microchip…

> If nothing else, there better be some decent PCIe available by then- 8x PCIe 5.0 or better. ARM needs to start doing bandwidth, some day (in places other than server class chips); it's been on USB3.0 for a long long time. Its not up to Arm to attach PCIe lanes but the chip makers who have been competing with x86 which already offer plenty of PCIe lanes. So far, only apple has bothered to make an Arm CPU which can c…

i thought it was pretty clear i meant the consumer arm ecosystem in general but thanks for chiming in & clarifying. yup.

usb4's ability to transport a variety of protocols across the switched network seems perfect. reviews like the Plugable tb4 hub show video and data intermixing nicely[1]. the complexity seems scary but in my mind it feels utterly & completely justified, and it builds atop the world we have & makes it better, quietly fixes the underpinnings, rather than insisting on re-invention. thunderbolt hasn't been flawless but it's been a niche product, a luxury product, and we'll quickly gain the experience to make it as pleasant bulletproof & flawless as one might hope for. already, reviews like the one i cited don't come with caveats or speed bumps: they show a working, satisfying consumer experience. there's a lot of afraidness of thunderbolt & usb4, but i for one think we ought to dial down our resistances & afraidness.

that said, i also think the door is totally open to new interconnects emerging. i'd love to see cxl or gen-z (recently folded into cxl) or opencapi or rapidio or whatever become consumerized.

[1] https://www.anandtech.com/show/16964/plugable-tbt4hub3c-thun...

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#24

Please forgive my lack of familiarity with PCI express hardware, but I have a question: All those PCI express devices are all sharing the single v2 lane? Is the PCI express bus a network?

All I/O is a network these days.

* USB is a network.

* PCIe is a network.

* SATA is a network.

* Infinity Fabric / Hypertransport (AMD socket-to-socket communications) is a network.

* UltraPath Interconnect (Intel socket-to-socket communications) is a network.

-----------

People think their internet connection is Layer4 (TCP) -> Layer3 (IP) -> Layer2 (Ethernet) -> Layer1 (Wires), but that's wrong.

Its really Layer4 (TCP) -> Layer3 (IP) -> Layer2 (Ethernet) -> Layer 3 (PCIe) -> Layer 2 (PCIe frames) -> Layer 1 (PCIe wires) -> reinterpret PCIe on the Ethernet-controller -> Layer1 (Ethernet wires).

There's a bunch of "networks" in-between our networks that we use. That's why OSI model exists, to understand each different network and each of the different rules behind them all.

When a good network is abstracted away, you can ignore all the details.

EDIT: It might not even be Layer2(Ethernet) -> Layer3 (PCIe)... on a NUMA-system it could be Layer2(Ethernet) -> Layer2 (NUMA / MESI messages to CPU#2) -> Layer2 (Southbridge communications) -> Layer3(PCIe)...

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#25

Earlier quoted context omitted.

I can see someone making a very powerful open source WiFi router out of this. The Pi can run passively cooled unlike most Intel or AMD CPUs. You wouldn't be doing any DPI on the network traffic without slowing things down, but this thing with some networking hardware cards could become very interesting to many people. The cost is too high because it's a niche product, of course, but such a board would be able to repl…

How many networking cards? I know we tend to think of internal buses as infinitely-fast, but a single PCIe 2.x lane is saturated at 500MB/s, which begins to look a little skinny once we're considering 10GE or even Wifi 6.

A single NVMe drive will saturate that.

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#26
post #5

Alas PCIe switches were starting to get affordable before the great chipmaker consolidation of 2015-2019 ("and seemingly overnight the cost of these switches increased three-fold"[1]). Now most every PCIe switch is >$6/lane. Even just the 12 lanes here, on an old PCIe generation, is going to be a $72 chip, perhaps more (article: PEX8619 Broadcom PCIe switch, $129, 16 lanes, $8/lane, no stock. Compare versus Microchip…

> It feels insane to hope for but I'm really hoping that by the time 2025 rolls around we see some modest affordable arm chips that have Thunderbolt built in. Perhaps even multiple Thunderbolt. If nothing else, there better be some decent PCIe available by then- 8x PCIe 5.0 or better. In the shorter term, I'm hoping the RK3588 actually comes out, is as good as promised, and has a reasonable price. [1] 4 lanes of PCIe…

I don't think the 4x Cortex-A76 + 4x Cortex-A55 is in the same league as the Apple M1, but still quite a bit faster than the Pi4's 4x Cortex-A72.

IMO the node used is more important than the core design. Rockchip claims they will be using an 8nm node for the rk3588. And I think the rk3399 and Pi4 are using 28nm nodes right now. So if this chip materializes it could be much faster (compared to Linux oriented ARM boards of today).

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#27
Things that are on my wishlist:

- Compute-Module should have a unified interface, that is not about to change for the next 3 generations of compute modules

- Someone create a "notebook" case having a board with mount for exactly this interface

- Case should have a decent keyboard and display without being to big

- How about having GPIO pins hidden

I would love to have an "upgradable" Notebook with arm architecture...

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#28

Once you get to this size and price, there’s no point in using a Pi…there are much more powerful options available.

I can see someone making a very powerful open source WiFi router out of this. The Pi can run passively cooled unlike most Intel or AMD CPUs. You wouldn't be doing any DPI on the network traffic without slowing things down, but this thing with some networking hardware cards could become very interesting to many people. The cost is too high because it's a niche product, of course, but such a board would be able to repl…

A 4 lane PCIe switch would allow for a much cheaper board and still be plenty to saturate the available bandwidth. Maybe 2x 2.5Gbe with one hooked to a switch, plus wifi.

Around this price point there are already ARM boards (and passively cooled x86 machines) available much better suited to this kind of use case, like the Honeycomb LX2K.

Re: Mini-ITX Seaberry Adds 11 PCIe Slots to a Raspberry Pi

#29

We're like one or two hardware generations away from having a VME backplane like environment where we'll be able to do some truly nifty and nasty things, at home and hobby levels. The politics of getting a standard out that isn't owned by (and therefore twisted to favor) a couple of big companies is challenging, but it's happened before, right? (crickets)

VME?
Post reply on HN