does Memory bandwidth doubles in sli ?

[brandon888]

does Memory bandwidth doubles in sli ? plan to buy second GTX 660 EVGA SC and then wait till maxwell ... as i know 660 in sli will be like 780's performance ... less stable but anyway

so does Memory bandwidth doubles in sli ? or i will stuck with laughable 144Gt/S on two cards , what will limit me badly ? ;/

 

[(sic)Klown12]

No. In SLI both cards have the same information stored in RAM so while you're moving twice as much data, it's just replicated. If it'll be a limit will depend on a number of factors like resolution and what games you play.

 

[psolord]

Memory bandwidth doubles with dual gpu. Memory framebuffer does not.

This is where most people get confused.

 

[brandon888]

Memory bandwidth doubles with dual gpu. Memory framebuffer does not.

This is where most people get confused.

so in reality cards both cards will be limited to 144Gt/s? if so .. then it's a bad idea to sli 660 or 660 TI .... better go for 670/680

 

[Jimzz]

so in reality cards both cards will be limited to 144Gt/s? if so .. then it's a bad idea to sli 660 or 660 TI .... better go for 670/680

Single card is better than dual in majority of cases.

 

[Erenhardt]

short answer: no.
long answer: no, SLI does not multiply memory bandwidth.

 

[BallaTheFeared]

Each card uses their own bandwidth for their own frame, nothing combines you just simply have double the workers (ROPS/TMUs/Shaders/Bandwidth) working on alternating frames.

660 SLI is faster than a single 680.

 

[guskline]

so in reality cards both cards will be limited to 144Gt/s? if so .. then it's a bad idea to sli 660 or 660 TI .... better go for 670/680

brandon888 it all comes down to the amount of money you want to spend.

You can see from my sig below the 2 rigs I run. Due to space limitations my 3rd rig doesn't appear. It's a FX8320 OC'd to 4.3Ghz on an Asus Sabertooth 990FX rev2 with 2 GTX660s in SLI. It's fast! Faster than my single GTX680 in my 8350 rig below but not as fast as my 3770k rig with 2 670s sli.

In my humble opinion, 2 GTX660s in SLI is one of the best "bang for the buck" combos out there for 1920 x 1080 monitors. There is a newegg video link showing the performance of GTX660s in SLI vs a single GTX680. Trust me, from my testing the 660s are significantly faster than the single 680 in a 1920 x 1080 resolution. By the same token I use the 670s sli in the 3770k rig to power my 27" 2560 x 1440 monitor.

 

[lagokc]

Memory bandwidth doubles with dual gpu. Memory framebuffer does not.

This is where most people get confused.

So for alternate frame rendering memory bandwidth doubles but the information is mostly duplicated anyway? For split frame rendering wouldn't memory bandwidth increase because they wouldn't always be pulling the same bits from memory? How about back in the days when SLI stood for scan line interleaving?

 

[omeds]

It kind of doubles, but it doesn't. You are still just as bandwidth constrained as a single GPU, however due to AFR with alternating frames, bandwidth constrained performance is effectively doubled. But, think of it this way, you will still hit bandwidth limitations that will effect performance at the same point as a single GPU.

 

[willomz]

Why does it even matter? Just look at actual results and base your purchase on that.

 

[brandon888]

It kind of doubles, but it doesn't. You are still just as bandwidth constrained as a single GPU, however due to AFR with alternating frames, bandwidth constrained performance is effectively doubled. But, think of it this way, you will still hit bandwidth limitations that will effect performance at the same point as a single GPU.

Then 144 gt/s limit will happen often ... even on 1080P ... :/ sad .... well

 

[boxleitnerb]

Memory bandwidth doubles with dual gpu. Memory framebuffer does not.

This is where most people get confused.

No, because each GPU still has its own memory bus and VRAM. Nothing doubles with SLI actually. But the GPUs are working on alternate frames, effectively doubling the work (fps) output. That is a fine but distinctive difference.

 

[stateofmind]

I don't think it's true. True, the memory does not *doubles*, but it surely does not stay the same, effectively. The fact that the frame is shared is only one part of the deal - the GPU has more uses for the memory and bandwidth than simply copy the frame..

Like Omeds said, in other words, you are less constrained. This is due to the fact that a single GPU has all its cores work on the same memory bandwidth whereas, effectively, two GPUs with half the cores could smartly use each their memory capacity and bandwidth for each frame, resulting, finally, in a better performance.

It is also visible in benchmarks

 

[ShintaiDK]

No, because each GPU still has its own memory bus and VRAM. Nothing doubles with SLI actually. But the GPUs are working on alternate frames, effectively doubling the work (fps) output. That is a fine but distinctive difference.

^^ This.

In SLI/CF, its simply 2 purely independent GPUs. Nothing is shared besides telling who is rendering what frame and then send the finished frame to the GPU with the output connected.

 

[moonbogg]

The only thing that doubles perfectly with SLI is the cost.

 

[boxleitnerb]

^haha, perfect! I would not have thought of that

 

[stateofmind]

ShintaiDK, I don't think so. Yes, it's two separate GPUs, but what about the mechanics of the SLI operation?

1. Lets take an example: The mobile GPUs GT 650M and 765M (the ones I checked)
They are basically the same, only the 765M has twice the CUDA cores. The 650M even has a bit lower maximal core clocks. However, the 765M outperformed by the 650M SLI in most cases, some of them are significant.

True, drivers might have a word here, but still.

2. Also, you can see that the GTX 765M does not have double the 650M performance

3. Nor the GTX 650 Ti vs GTX 650 in own Anandtech review:
http://www.anandtech.com/show/6359/the-nvidia-geforce-gtx-650-ti-review/13

You can check all the games. The difference does not equal x2 minus the clocks differences and it cannot be explained by merely looking at the specs. How come?

and a lot of other reviews support it.

 

[stateofmind]

and the AA performance is another indicator too:
http://forums.anandtech.com/showthread.php?t=2217589

 

[ShintaiDK]

ShintaiDK, I don't think so. Yes, it's two separate GPUs, but what about the mechanics of the SLI operation?

1. Lets take an example: The mobile GPUs GT 650M and 765M (the ones I checked)
They are basically the same, only the 765M has twice the CUDA cores. The 650M even has a bit lower maximal core clocks. However, the 765M outperformed by the 650M SLI in most cases, some of them are significant.

True, drivers might have a word here, but still.

2. Also, you can see that the GTX 765M does not have double the 650M performance

3. Nor the GTX 650 Ti vs GTX 650 in own Anandtech review:
http://www.anandtech.com/show/6359/the-nvidia-geforce-gtx-650-ti-review/13

You can check all the games. The difference does not equal x2 minus the clocks differences and it cannot be explained by merely looking at the specs. How come?

and a lot of other reviews support it.

Cuda cores alone does not dictate performance. Memory bandwidth, core clockspeed of TMUs, ROPs etc is all part of it.

There is nothing magic, no fairy dust in drivers or SLI/CF etc. Its basic frame altering rendering. The only driver issue is to get them synced without (too much) microstuttering.

 

[stateofmind]

that was my point. It was an answer to the statement that the memory bandwidth is the same. I say it is - effectively - not.

If you have the same cores count, same memory bandwidth, same TMUs count and lower clocks, how come the SLI is much faster in most cases?

 

[ShintaiDK]

that was my point. It was an answer to the statement that the memory bandwidth is the same. I say it is - effectively - not.

If you have the same cores count, same memory bandwidth, same TMUs count and lower clocks, how come the SLI is much faster in most cases?

SLI would not be faster if you took a single card, halved all the speeds and used 2 in SLI vs 1 original card. The cards in SLI would be slower.

 

[bystander36]

that was my point. It was an answer to the statement that the memory bandwidth is the same. I say it is - effectively - not.

If you have the same cores count, same memory bandwidth, same TMUs count and lower clocks, how come the SLI is much faster in most cases?

As said above a few times. Each card runs at the same speed they normally run. SLI or crossfire increase performance, by having each of the two cards work on their own frame. They take turns doing every other frame.

The reason perfect scaling does not occur is because the CPU has to prep those frames, and might not be able to keep up.

Whether or not you consider the bandwidth doubled depends on how you look at it.
- For each frame, the bandwidth remains the same.
- For every two frames, the bandwidth is doubled. (Though it still takes the same time to send the information through for each frame. It is just theoretically possible to send two frames at the same time.)

 

[Arkaign]

SLI would not be faster if you took a single card, halved all the speeds and used 2 in SLI vs 1 original card. The cards in SLI would be slower.

This.

The various penalties involved in SLI will always make it slower than a single card with exactly 2x or more raw performance of the single card in an SLI config.

However, SLI allows for putting two cheapish cards together to get good performance, or for putting two strong cards together to get impossible single-card performance. All still with the drawbacks of latency/etc.

 

[stateofmind]

But that's not what I said.

Again, only examples:

1. Benchmarks. Two GTX 650 Ti are faster than one GTX 680.

2. One GTX 680 is not twice as fast as GTX 650. One GTX 650 Ti is not twice the performance of a GTX 650 (mines the clock differences).

and more such example. How do you explain that?