PCIE compatibility question

[dentedcapacitor]

From a previous thread, I need to clarify PCIE compatibility to fully diagnose a problem.

 

Will a PCIE2.0x16 graphics card work in PCIE3.0x8 slot?

 

Does a PCIE3.0x16 graphics card still utilize all 16 lanes even when it is used in a PCIE2.0x16 slot or it only utilizes 8 lanes?

[Godlygamer23]

1.) Yes, it'll work just fine. PCIe is backwards and forward compatible with all generations.

 

2.) All 16 lanes are still used.

[Enderman]

1) yes it will work

 

2) lanes is just a name for bandwidth

a GPU will use as much bandwidth as it needs

usually it uses about 8 PCIe2.0 lanes (same as 4 PCIe 3.0 lanes) but nvidia gpus need at least 8x to work

when you go down to 4x 2.0 lanes you can start to get a performance bottleneck

 

this is why you can have multiple GPUs running at 8x and there will be no performance decrease

[Godlygamer23]

1) yes it will work

 

2) lanes is just a name for bandwidth

a GPU will use as much bandwidth as it needs

usually it uses about 8 PCIe2.0 lanes (same as 4 PCIe 3.0 lanes) but nvidia gpus need at least 8x to work

when you go down to 4x 2.0 lanes you can start to get a performance bottleneck

You do not need x8 for NVIDIA cards to work. You need x8 for SLI to work. Lanes are not the same as bandwidth. Lanes are pathways of communication. If lanes were just a name for bandwidth, could a PCIe 3.0 x16 slot be faster than a PCIe 2.0 x16 slot? Same number of lanes.

[Enderman]

You do not need x8 for NVIDIA cards to work. You need x8 for SLI to work. Lanes are not the same as bandwidth. Lanes are pathways of communication. If lanes were just a name for bandwidth, could a PCIe 3.0 x16 slot be faster than a PCIe 2.0 x16 slot? Same number of lanes.

pretty sure if you had all your lanes except 4 used by other expansion cards the GPU would not work, since its hardcoded into the drivers

 

and the "3.0" part of the name indicated third generation which is double the bandwidth of 2.0

thats why 3.0 8x = 2.0 16x bandwidth

its still measuring the bandwidth, it just that the amount of bandwidth depends which PCIe version you are talking about 

[Godlygamer23]

pretty sure if you had all your lanes except 4 used by other expansion cards the GPU would not work, since its hardcoded into the drivers

 

and the "3.0" part of the name indicated third generation which is double the bandwidth of 2.0

thats why 3.0 8x = 2.0 16x bandwidth

its still measuring the bandwidth, it just that the amount of bandwidth depends which PCIe version you are talking about 

Lanes do not directly correlate to bandwidth. NVIDIA cards work with x4 just fine - they need x8 for SLI to work.

[Enderman]

Lanes do not directly correlate to bandwidth. NVIDIA cards work with x4 just fine - they need x8 for SLI to work.

yeah they do

 

PCI Express link performance^[21][22] PCI Express

version Line code Transfer rate^[a] Bandwidth Per lane^[a]              In a ×16 (16-lane) slot^[a]

1.0 8b/10b          2.5 GT/s          2 Gbit/s (250 MB/s)                 32 Gbit/s (4 GB/s)

2.0 8b/10b          5 GT/s             4 Gbit/s (500 MB/s)                 64 Gbit/s (8 GB/s)

3.0 128b/130b     8 GT/s             7.877 Gbit/s (984.6 MB/s)      126.032 Gbit/s (15.754 GB/s)

4.0 128b/130b   16 GT/s           15.754 Gbit/s (1969.2 MB/s)     252.064 Gbit/s (31.508 GB/s)

[Godlygamer23]

yeah they do

 

PCI Express link performance^[21][22] PCI Express

version Line code Transfer rate^[a] Bandwidth Per lane^[a]              In a ×16 (16-lane) slot^[a]

1.0 8b/10b          2.5 GT/s          2 Gbit/s (250 MB/s)                 32 Gbit/s (4 GB/s)

2.0 8b/10b          5 GT/s             4 Gbit/s (500 MB/s)                 64 Gbit/s (8 GB/s)

3.0 128b/130b     8 GT/s             7.877 Gbit/s (984.6 MB/s)      126.032 Gbit/s (15.754 GB/s)

4.0 128b/130b   16 GT/s           15.754 Gbit/s (1969.2 MB/s)     252.064 Gbit/s (31.508 GB/s)

I know that - but the number of lanes do not indicate the amount of bandwidth available.

[Glenwing]

pretty sure if you had all your lanes except 4 used by other expansion cards the GPU would not work, since its hardcoded into the drivers

 

and the "3.0" part of the name indicated third generation which is double the bandwidth of 2.0

thats why 3.0 8x = 2.0 16x bandwidth

its still measuring the bandwidth, it just that the amount of bandwidth depends which PCIe version you are talking about 

 

I've personally run and benchmarked both a GTX 780 Ti and an R9 290 in a PCIe 2.0 x1 slot, so no... there are no requirements for lanes for graphics cards. There's no technical limitation for multi-GPU either, the x4 SLI requirement from NVIDIA is completely artificial and can be hacked around.

 

As for the OP, there is no such thing as PCIe incompatibility. Any PCIe device of any generation of any size will work in any slot of any generation of any size, provided that you can physically fit the card into the slot (sadly most x1 slots are blocked off so larger cards can't be inserted, for example).

[Enderman]

benchmarked both a GTX 780 Ti and an R9 290 in a PCIe 2.0 x1 slot

cool, didnt know that

 

I know that - but the number of lanes do not indicate the amount of bandwidth available.

yeah it does...look at the "bandwidth per lane" part

 

if its 1. its number of lanes*2Gbps

if its 2.0 its twice that much

if its 3.0 its 4x that much

if its 4.0 its 8x that much

it increases exponentially each generation

[Godlygamer23]

cool, didnt know that

 

yeah it does...look at the "bandwidth per lane" part

 

if its 1. its number of lanes*2Gbps

if its 2.0 its twice that much

if its 3.0 its 4x that much

if its 4.0 its 8x that much

it increases exponentially each generation

You're missing what I'm saying. The number of lanes present do not tell you how much bandwidth is physically available. If you had no specs, the number of lanes would be useless.

[Enderman]

You're missing what I'm saying. The number of lanes present do not tell you how much bandwidth is physically available. If you had no specs, the number of lanes would be useless.

well thats because each version is different...

as long as you know the PCIe version then you know the bandwidth depending on the number of lanes

 

but if you dont need to know the exact values you can just divide or multiply the number of lanes by 2 to find how the bandwith compares to a different version

 

for example you want your GPU to have 8x 2.0 speed, you can just divide by 2 and now you have 4x 3.0 speed

its that simple

the numbers and Gbps dont really matter when you can compare to different lane and version values with simple math

[Glenwing]

cool, didnt know that

 

yeah it does...look at the "bandwidth per lane" part

 

if its 1. its number of lanes*2Gbps

if its 2.0 its twice that much

if its 3.0 its 4x that much

if its 4.0 its 8x that much

it increases exponentially each generation

 

Right, as he said: lane count does not directly correlate with bandwidth. There is more at work. You can get 4GBps of bandwidth with 4 lanes, or 8 lanes, or 16 lanes. There are different ways of configuring for bandwidth and the number of lanes is one of the factors.

[Glenwing]

as long as you know the PCIe version then you know the bandwidth depending on the number of lanes

 

This is basically like saying CPU frequency directly correlates with performance, because as long as you know the architecture you can derive the performance... The fact that you need to throw in that "as long as" statement means that in fact it doesn't directly correlate.

[Enderman]

Right, as he said: lane count does not directly correlate with bandwidth. There is more at work. You can get 4GBps of bandwidth with 4 lanes, or 8 lanes, or 16 lanes. There are different ways of configuring for bandwidth and the number of lanes is one of the factors.

well, the number of lanes is directly proportional to the bandwidth within the PCIe version

and converting to other versions is a matter of multiplying or dividing by 2

 

so 4x is ALWAYS twice the bandwidth of 2x, and half the bandwidth of 8x, and a quarter of 16x, doesnt matter what version

[Glenwing]

well, the number of lanes is directly proportional to the bandwidth within the PCIe version

and converting to other versions is a matter of multiplying or dividing by 2

 

so 4x is ALWAYS twice the bandwidth of 2x, and half the bandwidth of 8x, and a quarter of 16x, doesnt matter what version

 

Yes they are related, the point is that the relationship isn't direct, meaning the number of lanes alone does not give you the bandwidth, so lanes by itself is not another word for bandwidth.

 

If you listed x4, x8, and x16 in a column, and then 2GBps, 4GBps, 8GBps, 16GBps etc. in another column, you could not draw lines directly between them. You would need a third column inbetween with the options PCIe Gen 1, 2, and 3. Then you would draw lines from the lane count to the PCIe version, and from there to the bandwidth. Hence, an indirect relationship. You can't draw conclusions about bandwidth straight from lane count, you need to travel through another consideration first.

[Enderman]

Yes they are related, the point is that the relationship isn't direct, meaning the number of lanes alone does not give you the bandwidth, so lanes by itself is not another word for bandwidth.

 

If you listed x4, x8, and x16 in a column, and then 2GBps, 4GBps, 8GBps, 16GBps etc. in another column, you could not draw lines directly between them. You would need a third column inbetween with the options PCIe Gen 1, 2, and 3. Then you would draw lines from the lane count to the PCIe version, and from there to the bandwidth. Hence, an indirect relationship. You can't draw conclusions about bandwidth straight from lane count, you need to travel through another consideration first.

true, but the software that tells you the lane count also tells you the PCIe version...so you get both at once