Probably not. The GPU can't tell whether the PCIe slot can't deliver enough power unless there's a voltage drop, and even then I'm not sure a GPU would properly handle that. So it's good to have a powered riser in that case.
-
Save 15% on ALL SFF Network merch, until Dec 31st! Use code SFF2024 at checkout. Click here!
Part Flexible PCIe risers
- Thread starter iFreilicht
- Start date
Probably not. The GPU can't tell whether the PCIe slot can't deliver enough power unless there's a voltage drop, and even then I'm not sure a GPU would properly handle that. So it's good to have a powered riser in that case.
Haha that would be great, wouldn't it? No there's not. That's why the Zaber sentry sandwiches the riser PCB between two parts instead of screwing it in directly. (@ZombiPL can tell you more about that solution)
The GPU has to negotiate with the motherboard on what power it wants to draw, and the motherboard can refuse. Of course, that does nothing to prevent the GPU from just drawing more current anyway (out of spec), and an accepted failure mode is the GPU simply refusing to power up (in spec).
Ah darn. I was hoping to use standoffs in a custom case to give extra support to heavy cards, but doing so would limit riser comparability if I needed to change risers later on.
To be exact, we are "sandwiching" not the riser PCB but PCI-E riser socket. We found so many risers with mounting holes made without keeping any standards, that we had to decide if we will promote only one riser, from one manufacturer or if we will work a little bit longer and we will design a universal solution for almost every pci-e riser there is. We picked 2nd option.
If you think about mounting your riser using those PCB holes, i suggest to either make a special holder for the PCB to strengthen it or think about making some support for your GPU. The riser's PCB isn't the strongest part of it, and if you will pick one from those cheap ~$5-10 USD risers, then if you will put the weight of GPU on it, its PCB might not last long. It isn't something to be affraid of with well constructed risers, where holes are made almost in the line with soldering and plastic pci-e socket is made with special supporting holders for PCB. In situation where PCB is holding socket almost only with soldering, or mounting holes on PCB are far from pci-e socket, then this is a thing to think about. That is why we picked such mounting option in Sentry. We do not really have to think about the type or shape of pcb but only about quality of wiring and soldering in the riser we want to use.
Last edited:
In addition to the HDPLEX riser, Li-Heat make a range of risers with different exit angles.
Very interesting to see actual numbers for the power delivery: 12V, 2.4A.
"35cm must be less than 45W; 50cm must be less than 35W" (source: http://cryinginthenight68.wixsite.com/liheat/products)
So Bus powered GPUs are out of question?
@lordzahl new Moddiy Premium Gold PCIe Riser works well with Bus powered GPUs. I tested it with 1050Ti.
Quality in this case means "signal integrity". How to measure that requires a electronics lab with gear to measure signal integrity of low-voltage differential signals.
High quality results in is high reliability. Low quality risers will not break any components, but might lead to crashes during high loads or in edge cases like the PC waking up from sleep.
In general, whether a riser works or not depends on your environment, and you really have no way of knowing other than testing. If you're lucky, a cheap 5 buck ribbon cable will work without any issues. If you want to be 100% certain that it will work at any time (for example if you're traveling with your PC and don't know how much noise your system will be subject to), get the 3M riser. If you don't mind experimentation and won't move the system often, a cheap riser could work for you as well. You might want to try some out (make sure to order some place where you can send them back for free), but in general the shielded ribbon cables (Li-Heat, Moddiy Premium and similar) seem to work for most people and look quite nice.
just thinking out loud: what about a really simple 'terminator-like' device and a driver? the terminator comes instead of a real pcie device.
e.g. the terminator could be something looping back data on each lane (or across lanes, etc). it must be an active component because it needs to negotiate the pcie version, number of lanes used, TLP length, etc). passive loopback would be too easy
so a pcie device with an fpga would do it. the problem is that they are really expensive so killing the whole idea of 'cheap' testing. altera and xilinx might have some nice boards.
unfortunately a cheap(er) x1 device is not an option because most likely we can not say that if one lane has a good quality then all the lanes have and regarding signal integrity this is absolutely an invalid test for multi-lane risers.
Also, having a NIC with enough bandwidth is really expensive, too. a 2x100G NIC doing a loopback would be pretty close to what I am thinking about. (either having a loopback cable or configured for doing loopback on chip level). the problem is not only the expensive NIC but also the CPU required for generating and capturing this amount of data. trafgen could be something like https://trex-tgn.cisco.com/ or http://dpdk.org/browse/apps/pktgen-dpdk/refs/
another option could be to add this feature to an existing videocard driver and write a tool like the Furmark. Or using Furmark itself if it can generate really high traffic with some settings on pcie. Intel's PCM (performance count monitor) or Microsoft's Xperf might give some information about the occupancy of the lanes but I am unaware of any counters like retransmit or so.
sorry for flushing all these things here.
I think the benchmarks done e.g. with GPUs are usually insufficient because of the used bandwidth. but it would be nice to have our own quality measurement. unfortunately all the things above are only for the given environment and do not show how the riser would behave in an other environment...
just thinking out loud: what about a really simple 'terminator-like' device and a driver? the terminator comes instead of a real pcie device.
e.g. the terminator could be something looping back data on each lane (or across lanes, etc). it must be an active component because it needs to negotiate the pcie version, number of lanes used, TLP length, etc). passive loopback would be too easy
so a pcie device with an fpga would do it. the problem is that they are really expensive so killing the whole idea of 'cheap' testing. altera and xilinx might have some nice boards.
unfortunately a cheap(er) x1 device is not an option because most likely we can not say that if one lane has a good quality then all the lanes have and regarding signal integrity this is absolutely an invalid test for multi-lane risers.
Also, having a NIC with enough bandwidth is really expensive, too. a 2x100G NIC doing a loopback would be pretty close to what I am thinking about. (either having a loopback cable or configured for doing loopback on chip level). the problem is not only the expensive NIC but also the CPU required for generating and capturing this amount of data. trafgen could be something like https://trex-tgn.cisco.com/ or http://dpdk.org/browse/apps/pktgen-dpdk/refs/
another option could be to add this feature to an existing videocard driver and write a tool like the Furmark. Or using Furmark itself if it can generate really high traffic with some settings on pcie. Intel's PCM (performance count monitor) or Microsoft's Xperf might give some information about the occupancy of the lanes but I am unaware of any counters like retransmit or so.
sorry for flushing all these things here.
I think the benchmarks done e.g. with GPUs are usually insufficient because of the used bandwidth. but it would be nice to have our own quality measurement. unfortunately all the things above are only for the given environment and do not show how the riser would behave in an other environment...
Hi sir
Is this the data you want?
Image original
Thank your for the nice product, The attached file is my test results using Infiniband FDR HCA, Mellanox MCX354A-FCBT. You can compare my result with others on the internet.
https://community.mellanox.com/thread/3401 https://community.mellanox.com/thread/1675 http://astronomy.swin.edu.au/staff/wp/ajameson/technical-memos/mellanox-mt27500-connectx3/ https://www.reddit.com/r/HPC/comments/3qb7ey/ib_write_bw_performance/ For the stress test, I performed ib_write_bw command for 80 hours with "run_infinitly" option. # ib_write_bw -F -b --size=8388608 --run_infinitely [server ip]
