I built a special PCIe card to test GPUs on the Pi
jeffgeerling.com
jeffgeerling.com
Given his progress and dedication, I fully expect a stack of A100's "graciously sent by _someone_" by this time in 2024 ;) To compare with a 32xRP2040 UHD VGA cluster in a separate post. Or something! This Dude is killing it in this niche.
That seems like the right choice. Specs exist for a reason. If you don't follow them, there is no reason your device should work. To adapt the world to work around your mistakes is insanity.
That said, there are often latent bugs in these situations that this sort of research/experimentation can uncover. Manufacturers would do well to invest the time in understanding why it doesn't work because those reasons can often lead to problems that are hard to catch when it is working.
"the Pi has problems with cache coherency on the PCI Express bus beyond 32-bits. And many (well, all nowadays) of the drivers expect that to function."
Yeah, that's a problem.
"there were issues with the BAR (Base Address Register) space allocated on the Pi's OS."
From the GitHub link:
# The default BAR address space available on the CM4 may be too small to allow # some devices to initialize correctly. To avoid 'failed to assign memory'
These don't look like implementation quirks, they are just flat-out errors.
Where are you getting your data?
I didn't mean to say anything against Linux nor imply that current Windows supports more hardware than current Linux.
I'm saying that one of the reasons Windows is so ubiquitous (on desktops) today is that in the 80s and 90s (during the "desktop wars") Windows computers supported a much wider range of hardware. Especially more than Mac OS who was their main competitor. Users could plug in ISA cards, PCI cards, serial printers, and a bunch of other things and most of the time it would work. The majority of users chose the OS that supported the expensive hardware they already had, and the rest is history.
Similarly: I agree with the statement "One could even argue that being tolerant of shitty hardware is one of the reasons that allowed Linux to become popular" and think it puts Linux in a position to continue to move forward in to the same sort of success that Windows had in the "desktop wars".
That big shot of the reviewer's face. The giant red arrow. The bright background. The breathless title with ALL CAPS. That's like 4/5 for clickbait videos. All it's missing is the face showing some ridiculously exaggerated emotion.
Don't let those thumbnails mislead you. geerlingguy is an active member of the open source community and he produces some absolute top rate instructional and informative videos.
And regarding clickbait, I wrote a bit on that topic over in the Ampere Dev Community site[1], but in summary:
> I decided to focus the intro bit on ‘Could this be the next Mac Pro’? (needs to be slightly clickbaity to get people to actually click. All of us engineering types SAY that we don’t like clickbait, but for 99% of us, if you gave us the clickbait version and the non-clickbait version in your YouTube feed… we’d click the clickbait version 100% of the time
[1] https://community.amperecomputing.com/t/review-of-the-ampere...
I, also, strongly dislike this (thumbnail) presentation but I suspect it must be necessary as it is very widely adopted.
Veratisium did a video on it https://www.youtube.com/watch?v=S2xHZPH5Sng
[1] https://store.timebeat.app/products/ocp-tap-timecard-mini-pl...
And it includes the $100 LEA-M8F GPS module... And the $150 SiTIME SiT5721 oxco over controlled oscillator. Excellent parts.
I was going to ask if the Pi had decent ptp support to actually be useful speaking ethernet, but you've already answered that (yes), https://www.jeffgeerling.com/blog/2022/ptp-and-ieee-1588-har...
I do kind of wish it had a PCIe ethernet controller onboard for the host too, also with ptp tied to the host. Sending that fine timing data across serial & PCIe bused to the host & to the hosts card feels unideal, like a source of some jitter, but maybe it's not really an issue.
Ideally computers would support consuming a PLL reference via 1GBASE-T, as that already has a need for a locked clock for TX and RX near-side cross-talk compensation, and CERN demonstrated usage of this for clock carrier distribution in their GPS-shielding tunnels. OFC integrated with PTP to accurately sync time; AFAIK they did for ns-accurate event time stamp assignment in some distributed sensor networks.
Why do you think corporations and everyone here push the MIT license so hard and in general dislike copyleft licenses, even weak ones like the EPL? It always boils down to "I want to use and modify your work for free at my workplace", though people will usually try to say this very indirectly, usually with some rant on "commercialization".