How many PCIe Extensions is TOO MANY??
Linus Tech Tips1 HN recommender · recommended in 1 month, 2017–2026
demo · intermediate · evergreen · 2017–2026 on HN
Linus Tech Tips tests whether chaining many PCIe riser extensions causes loss of GPU performance. They keep adding extensions to see how far the signal will go.
HN cites it as empirical evidence that PCIe cables can be chained to roughly three meters before failing.
Up next
2 mo on HNGell-Mann Amnesia and Michio Kaku
Angela Collier
2 HN recommender · 2024–2026
12 mo on HN8-bit computer update
Ben Eater
14 HN recommender · 2017–2023
11 mo on HNLight Years Ahead | The 1969 Apollo Guidance Computer
TNMoC
11 HN recommender · 2021–2026
11 mo on HN“Hello, world” from scratch on a 6502 — Part 1
Ben Eater
11 HN recommender · 2019–2026
10 mo on HNFinding the Missing Memristor - R. Stanley Williams
The Qualcomm Institute
12 HN recommender · 2010–2024
10 mo on HNThe Ultimate Game Boy Talk (33c3)
media.ccc.de
10 HN recommender · 2017–2026
11 mo on HNThe Thirty Million Line Problem
Molly Rocket
11 HN recommender · 2018–2024
9 mo on HNUSENIX ATC '21/OSDI '21 Joint Keynote Address-It's Time for Operating Systems to Rediscover Hardware
USENIX
12 HN recommender · 2021–2026
8 mo on HNJim Keller: Moore's Law, Microprocessors, and First Principles | Lex Fridman Podcast #70
Lex Fridman
8 HN recommender · 2020–2023
7 mo on HN34C3 - The Ultimate Apollo Guidance Computer Talk
media.ccc.de
9 HN recommender · 2018–2025
8 mo on HNJim Keller: Moore’s Law is Not Dead
UC Berkeley EECS
7 HN recommender · 2019–2025
8 mo on HN"Programming a 144-computer chip to minimize power" - Chuck Moore (2013)
Strange Loop Conference
7 HN recommender · 2021–2024
7 comments from Hacker News
I love a good “PCI Challenge”, where someone tries to see how much functionality they can wring out of PCI devices in a weird situation. Linus Tech Tips has a bunch of good ones: - How crazy of a device can you use with a server motherboard’s PCI hot-swap BIOS feature: link - How many PCI extenders can you use with a high end graphics card: ▶ this video
PCIe is more like a packet-switched network than a bus, which is incidentally why things like Thunderbolt (effectively external PCIe) and sillier demonstrations like ▶ this video work. ...and with things like link , you can get even higher latencies.
Someone needs to get this man a speed square, also a longer riser cable [1]. [1]I don't care much for Linus Tech, but ▶ this video
Here is an LTT video about how many PCI-E extensions one can daisychain without problems. The limit is around 3 meters. ▶ this video
LTT has a video where they tried to see how many PCIe riser cables they could string together before it stopped working.[1] They got to several meters. Maybe you could argue that it's worse inside a PC case since there's more EMI, but it seems like your PCIe riser cable would have to be very out of spec before you'd notice anything. [1] ▶ this video
I wonder if that benchmark actually loaded the PCIe bus to any significant degree after the initial benchmark startup, or just updated a couple small parameters on a single scene and thus mainly just tested the local computation on the GPU? You'd want to somehow monitor the PCIe bus error rate - with a marginal signal and lots of errors -> retries, something that loads the bus harder (loading new textures etc) could suffer way more. They do briefly show a different PCIe riser made out of generic ribbon cable [1, 3:27], and say that one failed after chaining only two ~200mm lengths. The quality of the riser cable certainly matters. [1] ▶ this video
It's a shame you can't get a cable to link computers via PCIe (supports cables over 3m[0]). Multiple x16 connections that would lead to some decent bandwidth while being low-latency, without the network overhead or shelling out for high-end switches. You can skip halfway. [0] ▶ this video