raspberry-pi-pcie-devices

There's more comparisons with various Arm hardware and Cyberpunk 2077 over here: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... The DGX Spark and Mac Studio are currently the two best Arm-based platforms for running that game, it seems to like a lot of CPU to feed a decent GPU.

Jeff Geerling (who is on HN every so often) is the site I'd go to: https://www.jeffgeerling.com/blog/2024/radxas-sata-hat-makes... Looks like he got near 100 MB/sec writing (and higher read) on Samba from it (capable of saturating 1 Gbps network) in testing and almost saturating a 2.5Gbps link with another adapter: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... (He's got several other blogs and YouTube videos in/around RPi topics.)

We're also trying to preserve the utility of older cards by getting at least portions of the drivers working on alternate platforms (like arm64), so they can be repurposed for Plex/Jellyfin transcoding, retro gaming, GPU compute, etc.: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss...

The current answer is maybe: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... The PCIe implementation on the 5 is supposedly more complete/less broken than on the CM4, but so far the only person crazy/inspired enough to test hasn't gotten back to this card with their Pi 5 setup.

> the M1 Ultra would surpass it by a wide margin Well, in which tasks? Comparing e.g. geekbench results for the M1 Ultra https://browser.geekbench.com/v5/cpu/14664498 and a decent 5950X (PBO) result https://browser.geekbench.com/v5/cpu/14665931 it looks like a lot of Apple's multicore score is attributable to AES (which uses special fixed-function HW). And with other tests, different CPUs win different benchmarks. 5950X notably wins in clang :) > bring eGPU support to the M1 line Probably not M1. marcan tweeted recently that Apple's PCIe integration has been is broken in the same way as on Broadcom's RPi4 SoC — mapping PCIe BARs as normal memory (which allows unaligned access) doesn't work. I can't find the tweet (deleted??) but here's the RPi thread: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... Basically there's no qui

"Work", in this case, meaning you can get a display to run through one. I've tested DVI, VGA, HDMI, and DisplayPort across four different graphics cards (three AMD 5000-7000 series, and one SM750 card from ASRock Rack. A console (CLI) runs great and is stable, but Xorg and Wayland (well, Weston at least) are a little less stable. And with the MMIO modifications we had to make to work around the Pi SoC's (BCM2711) flaws, things are a bit slow currently. And other SBC SoCs seem to have similar PCIe hardware-level issues (like the RK3588 in the Pine64 SOQuartz, at least). Tons more detail can be found in these two GitHub threads: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... (Both issues are over GitHub's limit, so you have to show hidden items to see the whole

I documented everything in excruciating detail in this GitHub issue [1] — but in summary it was RAID5 created by mdadm (to replicate what the Drivestor 4 Pro does via ADM—it doesn't have ZFS as an option), and bcache set up on top of that with an EXT4 filesystem. [1] https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss...

The Raspberry Pi does not support PCIe I/O BAR space. That's not a hard requirement to run a modern GPU as far as I know, but drivers do kind of assume it's available. In particular, VBIOS stuff often needs it. The message from the Radeon driver is not a hard error, but presumably other things break as a result. The Broadcom SoCs used on Pis still suck, and although this can probably be made to work, someone familiar with this codebase needs to care enough to fix the drivers to make it happen. It's probably quite doable for the open source drivers, but I wouldn't expect Nvidia's blob to work any time soon unless they take the time to fix it. Not all PCIe root ports are equal. Some are, shall we say, less than amenable to devices more complicated than a USB controller. Looking through the discussion here, there is some rather concerning st

There are a few people currently working on NAS boards based on the CM4. A few different approaches in terms of chipsets used and performance targets. I'm trying to track them in this GitHub issue [1] but a couple people have remained more or less anonymous as they don't want to attract attention too early. The single PCIe lane is the biggest limitation if you're looking for raw speed (350 MB/sec is kind of the upper real-world sustained transfer limit), though since the gigabit Ethernet port is on a different interface, you can still expect to get 80-100 MB/sec network transfer speeds. Something like this, with the right case and OMV or other adequate software would be a relatively competitive replacement for low-end NASes. I'm also exploring building a 2.5G NAS with a CM4, but the PCIe bus speed limitation is what kinda hamstrings that.

There are a few addon boards (for $15-30, most shipping after the new year) that take the CM4 that extra mile: https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... I'm most excited about a few with mini PCIe or M.2 slots on the bottom.

Follow https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss... for progress. You can't boot directly from the SSD without a microSD card, but it (or SATA support for boot) may come soon... just a question of priorities for the Pi Engineering team.

It looks like the MCUZone CM4 4G mentioned here [1] is intended for just that purpose! I would be happy to try out a chipset provided it's not too expensive, and (ideally) it has good Linux driver support. Drivers seem to be the dealbreaker for a few cards, if they rely on PCI express features the Pi OS or SoC currently don't support (like BIOS, or sort-of MSI-X support). [1] https://github.com/geerlingguy/raspberry-pi-pcie-devices/iss...

Check out the linked GitHub repo: https://github.com/geerlingguy/raspberry-pi-pcie-devices — for any card I'm actually testing, I have an open issue. For proposals or discussions around a certain topic, there are some ongoing conversations in the Discussions tab. Specifically concerning Infiniband, check out this issue: https://github.com/geerlingguy/raspberry-pi-pcie-devices/dis... - I'd be willing to test, but currently I have a bit of a backlog of other gear I'm getting through testing :)