Multi GPUs, what is the second GPU really doing?

[nicehat]

One question that has always hounded me surrounds the need for two GPU's to work in tandem vs one alone. The new 295 x2 is a asetek cooled multi GPU with  two Hawaii GPUs onboard, fully enabled and clocked at 1018MHz. It runs slightly better than the 1GHz peak clock of the Radeon R9 290X, and each GPU has its own 4GB bank of GDDR5 memory hanging off of a 512-bit interface. So we are looking at 8 GB of VRAM and a combined memory bandwidth of 1024 bits.

 

Why are we using two GPUs when one with an increased bus, and double the VRAM might be enough? 

 

If I take my 2 7950s and monitor them using afterburner during a game, I will see that one GPU does all the work, while the other does almost nothing. It sits at around 10-30% workload. Is this a waste of a die? Why couldnt one 7950 GPU with double the bus and memory perform as well with less heat than two GPUs in tandem? Or is this the case? What exactly is the second GPU doing? Is it just cheaper to have the workload distributed over 2 dies rather than a bus architecture/memory overhaul?  

 

There really isnt alot of information about how dual GPUs work and how the frame work load is distributed. I just dont get why both 295 x2 GPUs need to be watercooled when one basically does all the work. If this is the case, do both cores really need to be overclocked? What if the master die was overclocked at 1018, and the slave (for the lack of a better term) was clocked lower, since its work load is unequal? Is this second OC necessary? Wouldnt this cut down on heat and maybe remove the need for a second waterblock? If someone knows then enlighten me.

 

 

 

 

 

[Paralectic]

You are answering your own question, we don't know but it works.

[ShadowCaptain]

As far as I know the most common type is AFR, alternate frame rendering, where each card renders a frame and then they are synced up.

 

The utilisation of the second GPU is entirely dependant on the game engine and the cards themselves, some will take advantage of SLI/Crossfire better giving almost equal use, some games will not use the second card at all

[Rashy]

What If the other card is use for PhysX? (Nvidia's)

Edited April 14, 2014 by Rashy

[ShadowCaptain]

What If the other card is use for PhysX? (Nvidia's)

 

Then in PhysX games, all PhysX calculations are offloaded onto the second card, not really worth it in my opinion

[Rashy]

Then in PhysX games, all PhysX calculations are offloaded onto the second card, not really worth it in my opinion

Yeah, dont worry im just asking.

[Tam3n]

You are talking about Vram and bus width, but memory alone is obviously not all that is required to render frames. Memory is only needed to feed the work in the core(s) where all the actual number crunching aka rendering is happening. Faster memory will help delivering resources and instructions to the various processors inside the gpu doing calculations.

 

With a dual gpu you achieve 2x performance in theory, since you simply have two times the processing power of one gpu. You can render different frames on both gpus at the same time in syncronation...

 

In a game that has good multi gpu support you can see 100% gpu utilization on both gpus when v-sync is off.

[Paralectic]

You are talking about Vram and bus width, but memory alone is obviously not all that is required to render frames. Memory is only needed to feed the work in the core(s) where all the actual number crunching aka rendering is happening. Faster memory will help delivering resources and instructions to the various processors inside the gpu doing calculations.

 

With a dual gpu you achieve 2x performance in theory, since you simply have two times the processing power of one gpu. You can render different frames on both gpus at the same time in syncronation...

 

In a game that has good multi gpu support you can see 100% gpu utilization on both gpus when v-sync is off.

I think this is correct.

Take a look at the SLI improved 337.50 drivers by nVIdia for example.

[MG2R]

I find it weird to see asymmetrical core usage. Have you confirmed this behavior on multiple games, or is this one specific game only?

 

In theory, both cores should be --as mentioned before-- doing equal amounts of work. The reason you see double the vRAM is because each core needs its own RAM to perform its job. In the vRAM things like textures and models are stored, neither core can render (parts of) frames without that information. When you see marketing boasting "8 GIGS OF RAM!!!" on a dual GPU card, remember that it's (relatively speaking) still a 4 gig card, you can't multiply the RAM/core with the cores and come up with a number that actually means anything. Same goes for the bus width.

[nicehat]

I find it weird to see asymmetrical core usage. Have you confirmed this behavior on multiple games, or is this one specific game only?

 

In theory, both cores should be --as mentioned before-- doing equal amounts of work. The reason you see double the vRAM is because each core needs its own RAM to perform its job. In the vRAM things like textures and models are stored, neither core can render (parts of) frames without that information. When you see marketing boasting "8 GIGS OF RAM!!!" on a dual GPU card, remember that it's (relatively speaking) still a 4 gig card, you can't multiply the RAM/core with the cores and come up with a number that actually means anything. Same goes for the bus width.

 

hmm Just to clarify, if I had a 295 x2, it would have all 8 GB available for video buffer whereas 2 290Xs would only have 4 GB. That and the increased bus (does having 2 512 bus lanes make it additive for memory transfer?) makes a single multi GPU desirable (espec for multi monitors) over 2 single GPUs. As you say, the core usage issue is weird. Ive looked into AMD supported games like Tomb Raider, Bioshock, Hitman Absolution, and non supported games like Skyrim, Total War Rome II (these un supported games do alot worse). I hope @LinusTech will be able to clear some of this clout either here, in a techquicke or on the WAN show. 

[Unlocker9000]

When you run multiply graphics cards, the idea is that the gpu's share the work. So the first graphics card draws the first frame, the second card draws the second frame, the first card draws the third frame and so on.

So the do the same work (drawing frames) and this requires that both gpu's has vram to load the same textures (as @MG2R explained).

Some games are badly optimised, so they don't really run well on multiply graphics cards, and it might even perform worse than on one graphics card. And it's often the graphics card vendors who optimises their drivers, so the badly optimised games run better on multiply graphics cards.

[ivhyn]

hmm Just to clarify, if I had a 295 x2, it would have all 8 GB available for video buffer whereas 2 290Xs would only have 4 GB.

Nope, both solutions only have 4GB Vram, the bits on the framebuffer have to be mirrored in order to work with two GPUs, even when they're on one pcb. 

[MG2R]

hmm Just to clarify, if I had a 295 x2, it would have all 8 GB available for video buffer whereas 2 290Xs would only have 4 GB.

You're looking at this the wrong way. A dual GPU graphics card is EXACTLY like two single GPU cards in SLI/Xfire... Except that both GPU's are on one PCB. They still have exactly the same wiring as two separate PCB's, with the only difference being that the SLI connector now runs embedded on the PCB, instead of going externally via a cable or (flex) bridge.

 

(for completeness: their are some optimizations in terms of power delivery circuitry that can be done when both GPU's are on one PCB, but other then that everything stays the same)

 

So, in terms of memory and memory bandwidth, a 295 X2 is the same as two 290X's in Xfire.

[nicehat]

 

Nope, both solutions only have 4GB Vram, the bits on the framebuffer have to be mirrored in order to work with two GPUs, even when they're on one pcb. 

 

You're looking at this the wrong way. A dual GPU graphics card is EXACTLY like two single GPU cards in SLI/Xfire... Except that both GPU's are on one PCB. They still have exactly the same wiring as two separate PCB's, with the only difference being that the SLI connector now runs embedded on the PCB, instead of going externally via a cable or (flex) bridge.

 

(for completeness: their are some optimizations in terms of power delivery circuitry that can be done when both GPU's are on one PCB, but other then that everything stays the same)

 

So, in terms of memory and memory bandwidth, a 295 X2 is the same as two 290X's in Xfire.

 

You are right, I got it mixed up. So, what;s the point in getting a 295x2 over 2 290Xs if heat noise, and space arnt variables? Or are those the only things you pay for? For example, the Titan Z  puts more than a 2x premium over a single Titan and basically gives you a better form factor with slight power optimizations as you say? 

[MG2R]

You are right, I got it mixed up. So, what;s the point in getting a 295x2 over 2 290Xs if heat noise, and space arnt variables? Or are those the only things you pay for? For example, the Titan Z  puts more than a 2x premium over a single Titan and basically gives you a better form factor with slight power optimizations as you say? 

 

Dual GPU graphics cards usually have better power optimizations, resulting in lower power use and thus lower heat output. They also come in smaller form factor. If all of those things are of no issue for you, then there really isn't any reason to go with one of them and you'll probably be better (cheaper) of buying two single GPU cards.

 

That is, of course, unless you want to run quad SLI/Xfire and you don't have enough PCIe slots.

[ivhyn]

You are right, I got it mixed up. So, what;s the point in getting a 295x2 over 2 290Xs if heat noise, and space arnt variables? Or are those the only things you pay for? For example, the Titan Z  puts more than a 2x premium over a single Titan and basically gives you a better form factor with slight power optimizations as you say? 

That's the question now. You could put a dual GPU in an ITX case while giving it great performance, that shouldn't be a concern for Titan Z buyers though because it's wider than dual slot and I'm not quite sure if it will fit into an ITX rig. 

I never saw the point in a dual GPU but you surely could build a great mining rig with six 295X2

[kamelion]

I think the game the OP is playing isn't optimised for crossfire/SLI. When I play an optimised game on my SLI system , both cards are working equally, sharing the load.

[The_Evenger]

wait wait wait wait, what?  did you say that your second card is sitting at 10-30% usage when gaming, are you sure that you have crossfire enabled, are you running a game without duel gpu support, my second readeon HD 6950 sits at 95-98% usage whenever im gaming...