PLX Chip On a GPU vs. On a Motherboard - Is There a Difference?

[JThorpe]

Hello All,

So I have a question regarding the PLX chip and GPU's.

Would the PLXchip being located on the GPU (GTX 690/HD 7990)be the same as a PLX chip being located on the Motherboard?

An example would if the HD 7990/GTX 690 is paired with a 1155 (future:1150 - Haswell) MB with one x16 PCIe 3.0, could you experience the same latency if you have 2 HD 7970's/GTX 680's being paired with a 1155 (future:1150) with two x16 PCIe 3.0's utilizing a PLX chip?

So 1155(1150) and a HD 7990/GTX 690 [GPU PLX] vs. 1155(1150) and 2 HD 7970's/GTX 680's [MB PLX].

Thank you in advance.

 

[Jaydip]

PLX Chip On a GPU should have the upper hand.It should have less latency compared to MB.

 

[ViRGE]

PLX Chip On a GPU should have the upper hand.It should have less latency compared to MB.

Perhaps I'm just reading this differently than you, but while you're talking about GPU to GPU latency, I'm reading it as CPU to GPU latency. In the case of the latter, the latency would be identical; the PCIe bridge chip would introduce the same amount of latency regardless of where it is.

 

[Jaydip]

You are right Virge but I thought he was asking about the possible gpu latency.

 

[blackened23]

PLX Chip On a GPU should have the upper hand.It should have less latency compared to MB.

This is not true. The opposite is the case.

PLX in general allows for quad SLI on motherboards, which is the primary purpose of having the PLX chip. Otherwise, PLX generally adds an undesirable amount of latency - if you compare 2X SLI on a PLX motherboard as compared to a motherboard using CPU PCIE x16 (or even double x8) lanes, the latter will have better benchmark scores.

I'm speaking specifically about motherboards using PLX chips, these motherboards are designed for quad SLI generally speaking. PLX on a single card dual GPU solution is a bit of a different situation - the PLX chip is actually necessary for those types of cards, and they won't have the same motherboard related latency issues.

 

[willomz]

Thing is if it's just 2 way SLI/CF then you don't need a PLX chip at all.

 

[Grooveriding]

If you're asking about the chip you see on Z77/67 motherboards to enable tri or quad SLI compared to the chip on a dual-gpu card like a 7990 or 690, then yes I think there is a difference in performance.

The motherboard chip allows additional PCIE lanes between the GPUs across the motherboard bus, but does not create additional PCIE lanes into the CPU through the bus, as that is a limitation of the type of CPU you are using. The chip on board the PCB of a dual GPU card is enabling communication on board the card between the GPUs, which is then sent out from the card across the motherboard's PCIE bus.

The chips are AFAIK different but doing something similar. It's actually an interesting question. If you took a 690, the equal of two 680s on a single card when clocked the same as a 680 SLI setup, would it perform the same or better on a Z77 mb using a single PCIE 3.0 x16 slot compared to two 680s using the two PCIE 3.0 slots at x8 as they perform on a Z77 board. I think essentially both setups have the same bandwidth available to them, don't they ?

If you're talking about the PLX chip on a Z77 board to allow tri or quad SLI, it will perform worse than a board that actually provides 40 true PCIE 3.0 lanes as 16/16/8 or 16/8/8/8 directly into the CPU, as well as adding more latency. The PLX chip just creates additional lanes between the cards on the bus, but they still end up travelling over the same 16 lanes into the CPU.

 

[willomz]

With no PLX and SLI/CF each card gets a x8 connection, simple.

WITH PLX then both cards have access to the x16 but they can't use more than x16 between them. So they can both use the x16 but they'd have to use it alternately.

I presume that with PLX you could still give each card it's own dedicated x8 connection if that was preferable.

PLX adds flexibility in how you use the bandwidth, but it doesn't add any extra bandwidth.

 

[Jaydip]

This is not true. The opposite is the case.

PLX in general allows for quad SLI on motherboards, which is the primary purpose of having the PLX chip. Otherwise, PLX generally adds an undesirable amount of latency - if you compare 2X SLI on a PLX motherboard as compared to a motherboard using CPU PCIE x16 (or even double x8) lanes, the latter will have better benchmark scores.

I'm speaking specifically about motherboards using PLX chips, these motherboards are designed for quad SLI generally speaking. PLX on a single card dual GPU solution is a bit of a different situation - the PLX chip is actually necessary for those types of cards, and they won't have the same motherboard related latency issues.

Are you sure about this? because the plx chip on 690 has 48 lanes split 2x16 between two gpus and the remaining 16 with the MB pcie bus.Most normal MBs don't have this amount of bandwidth.

 

[blackened23]

Are you sure about this? because the plx chip on 690 has 48 lanes split 2x16 between two gpus and the remaining 16 with the MB pcie bus.Most normal MBs don't have this amount of bandwidth.

The PLX chip on the 690 merely enables the GPUs to communicate with each other with both GPUs being on a single card. It has no bearing on what bandwidth the actual PCIE x16 slot has, that is determined by the CPU PCIE lanes. Like I said it's a bit of a different situation. Since that communication between GPUs is being done completely on the card itself, it doesn't deal with the performance woes of PLX as it would on a motherboard implementation.

On motherboards using the PLX chip, as someone else said, it gives you more flexibility and more pcie slots for quad SLI - but the latency is higher and will degrade performance appreciably. Think of it this way - the CPU PCI express lanes are always used no matter what. If you add the PLX chip on top of that, it's another layer of latency to deal with.

Now, some quad SLI motherboards will actually disable the PLX chip *unless* you're using tri or quad SLI so that the end-user can get the best benchmark scores and performance - the Asus Maximus V Extreme actually does this. It will disable the PLX chip altogether if you're using 2 or fewer GPUs for better performance, and will only enable PLX with 3 or more GPUs. However, most motherboards do not disable PLX in such situations, and thus have slightly worse performance with single or dual GPUs. I actually used the P8Z77 premium for some time which supports Quad SLI, and the benchmark scores for dual SLI were appreciably worse than dual SLI benchmarks on the P8Z77 deluxe (doesn't use PLX). For instance, off the top of my head, the 3dmark11 score with the premium board was roughly 2500 points shy of what the deluxe gets (all other factors being the same). Needless to say, since i'm not a fan of Quad SLI I ended up selling the P8Z77 premium and kept the P8Z77 Deluxe because of the performance woes related to the PLX chip. The Deluxe is still serving me well, although I imagine i'll end up getting the upcoming P8Z87 Deluxe for haswell very soon 🙂

 

[aigomorla]

ok can someone correct me?

When was the PLX chip ever on nvidia cards?
The Nvidia "PLX" chip if i recall was called the NF200.

^ Remember that guy and some boards costing MORE if people wanted 3xSLI?

PLX i swear was on AMD cards...

Am i behind in tech? Did they move the PLX chip onto nvidia cards?

 

[Jaydip]

The PLX chip on the 690 merely enables the GPUs to communicate with each other with both GPUs being on a single card. It has no bearing on what bandwidth the actual PCIE x16 slot has, that is determined by the CPU PCIE lanes. Like I said it's a bit of a different situation. Since that communication between GPUs is being done completely on the card itself, it doesn't deal with the performance woes of PLX as it would on a motherboard implementation.

On motherboards using the PLX chip, as someone else said, it gives you more flexibility and more pcie slots for quad SLI - but the latency is higher and will degrade performance appreciably. Think of it this way - the CPU PCI express lanes are always used no matter what. If you add the PLX chip on top of that, it's another layer of latency to deal with.

Now, some quad SLI motherboards will actually disable the PLX chip *unless* you're using tri or quad SLI so that the end-user can get the best benchmark scores and performance - the Asus Maximus V Extreme actually does this. It will disable the PLX chip altogether if you're using 2 or fewer GPUs for better performance, and will only enable PLX with 3 or more GPUs. However, most motherboards do not disable PLX in such situations, and thus have far worse performance with single or dual GPUs. I actually used the P8Z77 premium for some time which supports Quad SLI, and the benchmark scores for dual SLI were far far worse than dual SLI benchmarks on the P8Z77 deluxe (doesn't use PLX). Needless to say, since i'm not a fan of Quad SLI I ended up selling the P8Z77 premium and kept the P8Z77 Deluxe because of the performance woes related to the PLX chip. The Deluxe is still serving me well, although I imagine i'll end up getting the upcoming P8Z87 Deluxe for haswell very soon 🙂

If a MB has only 16 lanes for PCIE so if I 680 SLI it can only use 8X|8X combination (even with PLX chip as CPU still sees only 16 lanes and the two gpus can't exceed 8X|8X combo when both are accessed).But with 690 both individual GPUs can still be accessed with 16X speed making it a faster choice.I am not sure we are talking about the same thing though 🙂.

 

[blackened23]

If a MB has only 16 lanes for PCIE so if I 680 SLI it can only use 8X|8X combination (even with PLX chip as CPU still sees only 16 lanes and the two gpus can't exceed 8X|8X combo when both are accessed).But with 690 both individual GPUs can still be accessed with 16X speed making it a faster choice.I am not sure we are talking about the same thing though 🙂.

The data isn't retrieved from the 690 in that manner, it isn't the same as 680 SLI. The data goes across 1 PCI express x16 lane. This is why I stated that it's a different situation - single card dual GPUs are treated differently.

 

[notty22]

ok can someone correct me?

When was the PLX chip ever on nvidia cards?
The Nvidia "PLX" chip if i recall was called the NF200.

^ Remember that guy and some boards costing MORE if people wanted 3xSLI?

PLX i swear was on AMD cards...

Am i behind in tech? Did they move the PLX chip onto nvidia cards?

NF200 is a pci-e 2.0 bridge chip, so it's retired. Nvidia uses a 3.0 bridge chip now on the gtx 690

 

[willomz]

If a MB has only 16 lanes for PCIE so if I 680 SLI it can only use 8X|8X combination (even with PLX chip as CPU still sees only 16 lanes and the two gpus can't exceed 8X|8X combo when both are accessed).But with 690 both individual GPUs can still be accessed with 16X speed making it a faster choice.I am not sure we are talking about the same thing though 🙂.

As you say "when both are accessed", however if you have PLX and one card isn't using the bandwidth, then the other card can use the full x16.

 

[Jaydip]

As you say "when both are accessed", however if you have PLX and one card isn't using the bandwidth, then the other card can use the full x16.

Yup

 

[willomz]

So why in that case do you think a 690 would be faster?

 

[Jaydip]

The data isn't retrieved from the 690 in that manner, it isn't the same as 680 SLI. The data goes across 1 PCI express x16 lane. This is why I stated that it's a different situation - single card dual GPUs are treated differently.

But then Why 690 has 48 lanes? isn't that one of the reason to make it really close to 680 SLI?

 

[Jaydip]

So why in that case do you think a 690 would be faster?

Because if you stress both the GPUs enough 16X would be faster technically.

 

[willomz]

The 690 can still only use 16 lanes at any one time.

 

[Jaydip]

The 690 can still only use 16 lanes at any one time.

IIRC 690 had dedicated 16X channel to each gpu.I may be wrong and will check.

 

[blackened23]

IIRC 690 had dedicated 16X channel to each gpu.I may be wrong and will check.

A single PCI Express card cannot circumvent the PCI express cpu lanes. The PLX chip on the 690 allows communication between the two GPUs on the card, nothing more nothing less with nearly minimal latency. It then (obviously) travels across 1 PCI Express slot. The data interchange across the PCI express bus with a 690 is different than it is with 2x GTX 680s.

 

[Jaydip]

A single PCI Express card cannot circumvent the PCI express cpu lanes. The PLX chip on the 690 allows communication between the two GPUs on the card, nothing more nothing less with nearly minimal latency. It then (obviously) travels across 1 PCI Express slot. The data interchange across the PCI express bus with a 690 is different than it is with 2x GTX 680s.

You are correct.I didn't consider that 690 still sits on a single PCIE slot so it can't access 32 lanes.

 

[Aikouka]

Here's what came to mind when I read this topic:

• I'm ignoring GPUs set aside for things like physics.
• In a multi-GPU scenario, all GPUs share the same information.
• Frames are transmitted to the display card through the SLI "bus."
• In a multi-card setup, is data is sent to each card individually?
• In a single-card, multi-GPU setup, is data is only sent once to the card, and the card splits the data?

Arguably, those last two points could spell out a performance advantage for a single-card, multi-GPU setup given there's only a single transmission from the source.

 

[willomz]

Kinda irrelevant if they are working on different frames.