Part Flexible PCIe risers

iFreilicht

FlexATX Authority
Original poster
Feb 28, 2015
3,243
2,361
freilite.com
A lot of new case designs are using flexible PCIe risers, and for a good reason. They allow for flexibility when placing the GPU, enabling much more efficient layouts.

There are many different risers out there, so I thought it would be helpful to have a thread where we can collect information on those risers and discuss recent developments regarding those risers.

So here are all the ones I have in my private research notes. I'll expand and correct this list until a suitable section on sffwiki.net is made.

Generic riser:



Price: ~3-20$
Shielded: No
Length: ~5-30cm
Connector: Straight
Available: ebay, multiple sellers.

These risers have been used in quite a few builds without causing problems. @Hahutzy specifically reported that he still hasn't had an issue using one of these in Hassium and the Hutzy XS. Other users like @Runamok81 haven't been that lucky and had to wrap their riser in aluminium foil to make it work.

In general, I wouldn't recommend using a single one of those, though you can of course try your luck.


LiHeat:



Price: ~20-50$
Shielded: Yes, copper tape
Length: 5-50cm
Connector: Straight ("D.type"), left-angle ("B.type" - seen in picture), right-angle ("A.type")
Available: Through ebay-user liheat48 or by e-mail.


These risers are pretty much glorified generic risers. They are made up of regular ribbon cables with copper tape in between to prevent crosstalk and metallic tape wrapped around the outside. This makes them quite stiff and thick, more than 2mm.

They perform fairly well, but users on different forums report problems with some of them. These risers are now offered with plastic strain reliefs on both ends which prevent damaging of the solder joints, maybe someone can post a more recent picture.

They are exclusively sold as leftover stock from contracts with larger companies on ebay, so it might take some time until the desired length and connector option becomes available directly. Sometimes they have a few more options when you contact them via mail, and they also make risers of custom length if you can reach a certain MOQ.

LiHeat produced the PW-PCI-E and PW-PCI-E38 risers for LianLi, which were used in the PC-O and PC-Y6 cases. These cables have been discontinued for unknown reason, more info below.

This riser was tested in a chained 100cm configuration and still managed to run Firestrike with minmal performance loss.

LiHeat actually posted drawings of each riser here, this will help you choose between A, B and D type.

LianLi:



Price: ~70-80$
Shielded: Quite probably.
Length: 30cm or 38cm
Connector: Right-angle
Available: Retail and online stores.

LianLi started using a flexible riser of some sort in the PC-C36-Muse a long time ago, but only recently revisited riser-based designs with the PC-O series and the novelty PC-Y6. These cases initially used risers from LiHeat (see above), but those were swapped out later for the PW-PCIE38-1 and PW-PCIE30-1, the supplier of which is unknown. One might suspect that this switch was caused by either supply or quality issues with the LiHeat cables.

Thermaltake:



Price: ~30$
Shielded: Yes.
Length: 22cm
Connector: Straight, very long.
Available: Retail and online stores.


Thermaltake started selling the AC-039-CN1OTN-C1 with the release of the not-so-SFF Core P5 Wall-mount chassis. Supplier and quality are unknown, but it might be made by sintech, see below.

HDPLEX:



Price: 35.5$ (as of 2016-08-03)
Shielded: No.
Length: 15+cm
Connector: Left-angle
Available: Only from HDPLEX directly.


This riser is probably the thinnest one in existence at ~0.14mm. Compared to other risers, it is manufactured as a double-sided FPC (flexible printed circuit). Because it isn't multi-layered, it is not shielded whatsoever. It is normally bundled with their H5 chassis, which uses a similar kind of cable for the Front USB3.0 connections.

The connector on the motherboard side is very different from normal risers, but works perfectly fine and leaves about 2mm of space towards the edge of an ITX mainboard, so no fitment issues arise from this configuration.

A short review can be read here. It was confirmed by HDPLEX that this riser was tested with GTX 750Ti GPUs without a 6pin connector with no problems. While it is advertised as 15cm, that seems to be the absolute minimum length this riser can have. Mine was more like 16.5cm.

3M:



Price: ~90-80$
Shielded: Yes, using twin-axial cables.
Length: 25cm or 50cm
Connector: Straight
Available: digikey

Most famously used in the DAN A4-SFX, this riser seems to be the holy grail in terms of quality. 3M has already confirmed it to be compatible with PCIe 4.0 specs, which is a very impressive feat. Unfortunately, it is one of the most expensive risers to date and is not available with angled connectors.

This riser is also available for 8x and 4x links.

ModDIY/sintech:



Price: ~20-40$
Shielded: Yes, probably using metallic tape.
Length: 5-30cm
Connector: Straight
Available: modDIY (19cm and 30cm), ebay sintech.cn (5+cm), aliexpress (5+cm)

Not a lot to say about this one. A few people have used it and so far I haven't heard of any problems. It seems very similar to the one Thermaltake is using, so maybe sintech is actually producing the Thermaltake riser.

Internally I would suspect it to be similar to the LiHeat riser, though it looks thinner on pictures. Sintech also makes custom risers to order and has a website where a right-angle riser with shielded ribbon cables is shown.

Adexelec:

Price: ~80$
Shielded: Yes and no.
Length: Varies:
Connector: Straight, Left-angle
Available: Only from manufacturer.

This company made the risers for the Orthrus Prototype, in which they seem to work great. It seems like their risers are only available made to order.
 
Last edited:

rokabeka

network packet manipulator
Jul 9, 2016
248
268
As of now I do not think anybody spends time on verifying the performance of those risers appropriately. Or at least I could not find a page with convincing results.
A test where they just run a 750ti with a 'small loss' does not tell anything about the riser without checking the utilization of pcie channel. it only defines that in the given config with the given game the riser was satisfactory.
I think somebody should come up with a good test metholodology to really verify those risers. Based on the tests, e.g. classes could be defined, etc. iirc cpu-z shows at least the pcie bandwidth used, probably that can be a good point to start with. so if there is a pcie-bandwidth hungry game (or 3dmark test) then at least that should be the reference. the riser is a component with its features and parameters. depending on the job it will be used for it might or might not be important. E.g. bitcoin mining rather has bottlenecks everywhere else (GPU+CPU) but on PCI-E bandwidth. but in case of e.g. a PCI-E SSD or a NIC or a ramdisk it can be important.

I completely understand that as of now it is not that important because many cards are happy with just x8, too and they will not realize any loss between x16 risers. this is probably a bit similar to the quality of HDMI-cables which unfortunately sometimes leads to the land of magic :D
 
  • Like
Reactions: SeñorDonut

iFreilicht

FlexATX Authority
Original poster
Feb 28, 2015
3,243
2,361
freilite.com
That is very true, we should find some way of testing and evaluating risers consistently.

The problem with GPU-Z is that the bandwidth displayed is only bound to the current link the GPU is communicating over, so as long as you get to PCIe 3.0 x16, you won't see any difference there. So that tool is pretty much worthless for benchmarking.

But yeah, ideally we'd have a benchmark that is stressing the link as much as possible. Maybe something like this already exists, but if it doesn't, it shouldn't be impossible to program.
 

K888D

SFF Guru
Lazer3D
Feb 23, 2016
1,483
2,970
www.lazer3d.com
You would also need some way of evaluating how well the cables shield from interference as this seems to be an important factor for these cables.

Bends in the cable also seem to effect the performance, perhaps when the wire is bent the cross section is reduced and signal bandwidth reduces or resistance increases or something like along those lines that will reduce performance.

Is there a way you can measure signal drop from one end of the cable to the other? Not in a data sense, but measuring current or something like that?

The ability of the wires to transmit in general may correlate with its PCIe 3.0 x16 performance.

Kind of like hdmi cables, an expensive hdmi cable won't give a better picture than a cheap one, but it is more likely to hold its signal strength over longer distances. I'm not sure how you would measure that though.
 

rokabeka

network packet manipulator
Jul 9, 2016
248
268
The problem with GPU-Z is that the bandwidth displayed is only bound to the current link the GPU is communicating over, so as long as you get to PCIe 3.0 x16, you won't see any difference there. So that tool is pretty much worthless for benchmarking.

yes, you are right, I messed things up, sorry. It was GPU-Z I have seen and remembered as it had PCI-E statistics, too. maybe from version 0.8.1 it has 'bus interface load'. I need to verify.

edit:
in case of a riser of a given quality probably it worths a try to tune PCI-E settings where possible. I know that we can not really play with the frequency but configuring TLP could give different results, not necessarily performance degradation by reducing TLP size.
 
Last edited:

QinX

Master of Cramming
kees
Mar 2, 2015
541
374
The major issue with the PCIe risers is not related to power but to signal quality. The best way to test that would be the test the differential signals using the PCIe eye test.

One thing I would be interested in knowing. Going from PCIe 1x tot 16x magnifies the amount of differential pairs by a factor of 16. Does that make a bad PCIe cable 16 times as likely to fail?

If so, wouldn't using a PCIe 3.0 4x riser be inherently more reliable then a 16x riser? Even though the potential performance loss is 5%-ish
Does losing 5% in performance and gaining more reliability mean anything to users? Or would they be pissed because of said 5% performance loss even if the cable works and is potentially much cheaper.

This seems like a good starting point for the eye pattern test, this kind of testing doesn't seem to be easily accessible to do.
http://cdn.teledynelecroy.com/files/whitepapers/designcon2013_slides_pcie_gen3_seminar.pdf
 

EdZ

Virtual Realist
May 11, 2015
1,578
2,107
AFAIK, PCI-E does not do any sort of link tuning. Either you have a good link, or a link that will have errors (interference, impedance mismatch, etc), but you would never have a link with 'degraded' speeds.
 

rokabeka

network packet manipulator
Jul 9, 2016
248
268
AFAIK, PCI-E does not do any sort of link tuning. Either you have a good link, or a link that will have errors (interference, impedance mismatch, etc), but you would never have a link with 'degraded' speeds.
yes.
as they say 'pci-e is not a bus' it is rather a network. and there is retransmission. what I can immagine is that depending on the environmental noise and the signal being transmitted it can make it or can not. the more you need to retransmit your packets the slower the connection is. I myself have never tested nor experimented pci-e channels. iirc there were some tests showing small degradation with pci-e risers. (I do not mean using a x16 card with x8 riser, I mean riser vs no riser, same number of channels). I have no explanation to that degradation. but if we can come up with a simple and reproducable test then either we can say that 'buy the cheapest available' or 'there is a difference, e.g. ....'.
this also makes me think about the first pcie slot vs the last one (e.g. on an x99 mobo with a 40 lane cpu). is it possible that even if they both have 16 lanes the one closer to the CPU is any better?
internet is huge, probably these questions are already been answered, I just could not find them :)

edit:
I am assuming it works as BirdofPrey says, just took too much time for me to finish
 

jeshikat

Jessica. Wayward SFF.n Founder
Silver Supporter
Feb 22, 2015
4,969
4,780
I don't think there's much issue with performance necessarily. Most of the issues I see with these flexible extenders is reliability.

The hardest part is it can be very inconsistent. One person won't be able to get one of the nicer shielded ones to work at all while someone else with similar hardware but with a cheap $5 eBay extender instead won't have any issues at all.
 

iFreilicht

FlexATX Authority
Original poster
Feb 28, 2015
3,243
2,361
freilite.com
This seems like a good starting point for the eye pattern test, this kind of testing doesn't seem to be easily accessible to do.
http://cdn.teledynelecroy.com/files/whitepapers/designcon2013_slides_pcie_gen3_seminar.pdf

"For Signal Integrity Engineers". Not accessible at all.

If so, wouldn't using a PCIe 3.0 4x riser be inherently more reliable then a 16x riser? Even though the potential performance loss is 5%-ish
Does losing 5% in performance and gaining more reliability mean anything to users? Or would they be pissed because of said 5% performance loss even if the cable works and is potentially much cheaper.

It depends on the user. If they are looking into a riser-based design anyway, they might prefer reliability to performance, but knowing that you lose performance always hurts.

But if the user is just a potential customer for a new build and can choose between a non-riser and a riser-based design that loses him 5% performance, he won't give a damn about the added reliability.

but if we can come up with a simple and reproducable test then either we can say that 'buy the cheapest available' or 'there is a difference, e.g. ....'.

Writing a Program that just sends as much data to the GPU and receives it back should be enough. That way you've got symmetric load and can just check how fast you get your data back, right?

I don't think there's much issue with performance necessarily. Most of the issues I see with these flexible extenders is reliability.

For that you pretty much have to run a system for a few days on full load and just see how often it BSODs, I guess. You need a lot of time for that to become representative, though.
 

rokabeka

network packet manipulator
Jul 9, 2016
248
268
Writing a Program that just sends as much data to the GPU and receives it back should be enough. That way you've got symmetric load and can just check how fast you get your data back, right?
yes. bandwidth+latency(min/max/avg)
edit:
oh. we might want to differentiate between v2.0 and v3.0 as they are different in both speed and encoding. or we might not :)
 
Last edited:

jeshikat

Jessica. Wayward SFF.n Founder
Silver Supporter
Feb 22, 2015
4,969
4,780
For that you pretty much have to run a system for a few days on full load and just see how often it BSODs, I guess. You need a lot of time for that to become representative, though.

The problem with that is it's still not a guarantee the extender is reliable. I don't think there's any good way to really test reliability without both lots of time and lots of different hardware to validate with.
 

iFreilicht

FlexATX Authority
Original poster
Feb 28, 2015
3,243
2,361
freilite.com
oh. we might want to differentiate between v2.0 and v3.0 as they are different in both speed and encoding. or we might not :)

It should always be checked whether the card is running in 2.0 or 3.0 when the test is being made, but it should jump to 3.0 quickly once there's enough load on there. Still, for robustness, this should at least be logged.

The problem with that is it's still not a guarantee the extender is reliable. I don't think there's any good way to really test reliability without both lots of time and lots of different hardware to validate with.

That's where the link QinX posted comes in, but you need a related degree and a good amount of expensive gear to make those tests, I would imagine.
 

jeshikat

Jessica. Wayward SFF.n Founder
Silver Supporter
Feb 22, 2015
4,969
4,780

EdZ

Virtual Realist
May 11, 2015
1,578
2,107
Even doing an eye-pattern test will only tell you about the electrical characteristics of the riser itself. It will tell you nothing about any issues that may occur due to interference.
 

QinX

Master of Cramming
kees
Mar 2, 2015
541
374
Even doing an eye-pattern test will only tell you about the electrical characteristics of the riser itself. It will tell you nothing about any issues that may occur due to interference.

Yeah, for that you would either need to subject it to real world tests or have it be running in a test chamber that can spit out a lot of noise.
 

iFreilicht

FlexATX Authority
Original poster
Feb 28, 2015
3,243
2,361
freilite.com
So you need an adjustable EMI source as well, so a wide-band antenna of sorts with a waveform generator and then test the riser with the ludicrously expensive signal analyser while using that as a source of interference. That sounds like a very complicated setup, I'd suspect it would take a whole day to test a single riser extensively.