256 vs 384 vs 512

[Nosada]

With the benchmarks released for the new batch of nVidia cards, I've been debating the use of bigger memory buswidth, but I don't feel I have all the facts to form an unbiased opinion on this matter.

Actual bandwidth of the memory-bus is dependant on speed of the memory and width of the bus, that much seems obvious. Lately it seems however, that with 7Ghz memory on most videocards, the buswidth is becoming less and less the deciding factor. An R9 285 doesnt perform worse than a R9 280 despite having a bus a third narrower. Same goes for the new GTX970 compared to an R9 290X or GTX780.

So am missing something obvious? Does buswidth have an advantage at QHD resolutions only, or is it basically a moot point at this time due to no more bandwidth being required for the current generation of games?

 

[rtwjunkie]

As I understand it, the faster the memory, the less of a need for 384 and 512 bandwitch...to a point. Of course I'm sure there are didminishing returns, but as it stands now, 256 seems to be enough at 7Ghz(+) memory speeds. Of course, I'd still be all over a 384 bit if it was released....as I anticipate a GM200/GM210 chip will be.

 

[Frick]

If a wider buswidth would automatically mean more data being shoveled thru it it would matter, but as you say it doesn't.

 

[rtwjunkie]

If a wider buswidth would automatically mean more data being shoveled thru it it would matter, but as you say it doesn't.

Well said!

 

[EarthDog]

Does buswidth have an advantage at QHD resolutions only

For the most part... there you have it! When you pour on the eye candy, specifically AA, you would want more bandwidth and memory.

(you may want to consider putting a more descriptive title on this thread...)

 

[Vario]

Usually its better to buy larger bandwidth cards. Kepler doesn't seem as sensitive to bus size as prior generations.

 

[Vayra86]

The bandwidth dependancy MUST be brought back on GPU's for the near future, as 1440p / 2160p are becoming the norm. This is being pushed right now but the limits of the current architecture are definitely in sight.

Both Nvidia and AMD are making moves now to reduce bandwidth dependancy. As far as we know at this time, Tonga manages to stay on par with a 384bit wide 280x on a 256bit bus. On the Nvidia side of things, we've seen slides where they market Maxwells reduction of bandwidth dependancy, which should decrease by about 25%. That is considerable. The biggest profits are likely offloading data to a bigger cache, and this is a move that Nvidia has already started to implement with driver updates that include Shader Cache. Maxwell builds on that idea by also providing a larger cache and using it differently. Texture streaming is rapidly becoming more efficient and as soon as streaming introduces similar latency to 'storing in memory', bus width is kind of obsolete.

Quite sure this isn't the whole story, but the bottom line is that it's clear that the days of 'bus width above all else' are finally over.

 

[EarthDog]

Ahh, the miracles of compression...

When was it ever bus width over everything else? I have to admit I never bought a GPU on that premise until 4K gaming came about. 256bit is still fine on 5760x1080...

 

[Vario]

Ahh, the miracles of compression...

When was it ever bus width over everything else? I have to admit I never bought a GPU on that premise until 4K gaming came about. 256bit is still fine on 5760x1080...

2006?

 

[EarthDog]

Hah, something like that. Not recently, that is for sure.

 

[Vayra86]

Sure it was. GTX 660ti was limited by its 192 bit bus, one of the major gripes of this card.

Many lower end/midrange cards are definitely limited by bus width / bandwidth, being able to do more with a similar bus is a good way to cut costs. Same reason we saw 256 bit with 7 Ghz mem on 7xx.

 

[EarthDog]

Its a low end/midrange card. Its not meant to play well at over 1080p.

+1 to cutting costs. But I still don't know who shopped by 'buswidth over anything else' except the misinformed. In the vast majority of cases it is much ado about nothing.

 

[rtwjunkie]

Sure it was. GTX 660ti was limited by its 192 bit bus, one of the major gripes of this card.

Many lower end/midrange cards are definitely limited by bus width / bandwidth, being able to do more with a similar bus is a good way to cut costs.

Very true. Basically the 660Ti could have very nearly been a 670 (and depending on model, could surpass the 670 on certain games. Just imagine what kind of a card it could have been without it's bandwidth artificially capped at 192 bit?!

 

[RCoon]

Its a low end/midrange card. Its not meant to play well at over 1080p.

+1 to cutting costs. But I still don't know who shopped by 'buswidth over anything else' except the misinformed. In the vast majority of cases it is much ado about nothing.

In the event manufacturers find bandwidth lacking, they probably won't waste money on a bigger bus. My best bet is they'll just find a better compression technique until they're stuck at an impass (LOL chrome tried to auto-correct that to imp ass) of computational limitation and then increase the bus width.
Not to mention those wider buses seem to get a little warmer...

 

[EarthDog]

In the event manufacturers find bandwidth lacking, they probably won't waste money on a bigger bus. My best bet is they'll just find a better compression technique until they're stuck at an impass (LOL chrome tried to auto-correct that to imp ass) of computational limitation and then increase the bus width.
Not to mention those wider buses seem to get a little warmer...

Its cheaper to put faster ram on a lesser bus too.

Very true. Basically the 660Ti could have very nearly been a 670 (and depending on model, could surpass the 670 on certain games. Just imagine what kind of a card it could have been without it's bandwidth artificially capped at 192 bit?!

A 670... let's jump for joy that it could be a card that already exists.

 

[rtwjunkie]

A 670... let's jump for joy that it could be a card that already exists.

LOL! No, what I was saying was Nvidia artificially crippled 670's basically to make 660Ti's. And depending on the maker, and how they tuned them or replaced components, it still held it's own against the 670. So, I wasn't saying it was almost a card that already existed, it was created and intentionally crippled, with the primary factor being to cut the bandwidth. That kept them from losing 670 sales with a card issued to meet a lower price point.

It was quite a good card in it's day only a couple years ago. 660's were midrange...the Ti was aimed at lower high end.

 

[EarthDog]

I understood what you said/meant... I was having some fun with it was all.

 

[Aquinus]

The size of the memory data bus should match the capabilities of the GPU. There is no point to pair a really wide memory bus with a GPU that can't take advantage of it. Also keep in mind that when you have a data bus that wide, you need to pull that many bits out of memory as well and to use memory optimally you need to be able to get data into the GPU's shader cache as quickly as you read it out of memory. So making the data bus wide isn't just doubling the number of transistors to do it, keep that in mind too because there are costs to widening any bus in an integrated circuit with respect to latency and circuit complexity.

 

[Vayra86]

Its a low end/midrange card. Its not meant to play well at over 1080p.

+1 to cutting costs. But I still don't know who shopped by 'buswidth over anything else' except the misinformed. In the vast majority of cases it is much ado about nothing.

Oh yeah they did. That is what AMD enthusiasts said to Nvidia enthusiasts when comparing HD7000 to Kepler. And they are right too, HD's play better with high resolutions and lots of AA, from the 7870 onwards, compared to 600 series Nvidia. There is more in the market than absolute high end cards you know For Kepler the limited buswidth was an issue on the GK106 and GK104 until they refreshed it and memory overclocks were almost 1:1 performance increases.

 

[EarthDog]

I know.. but it still doesn't matter!!! People were all too hung up on that for no reason. Its simple people. Look at the performance of the card at your resolution. Is it good enough? Get it. Don't sweat these cheesy details (unless you are rocking 4K - which is new to the landscape which goes right back to my point... who does that is misinformed on what is important and for most its a non issue).

 

[Vayra86]

Back in 2012 for people who were pushing 1440p, that 256bit bus on 670 was holding them back from doing so without stutter and 7970 performed a lot better on similar price points.

Something similar is happening today with current gen GPU's and 4k.

Sorry but I really can't agree it's a non-issue. Bandwidth is a hot topic every gen and for 900 series it shows that the actual energy savings are made by advances in the way they treat and use the bus width We also know from the microstutter chit-chat that simple FPS tables don't say jack shit about actual gaming smoothness, and memory plays a major role in this.

 

[rtwjunkie]

I understood what you said/meant... I was having some fun with it was all.

Yeah, I knew, to a degree, hence the "LOL" (and I chuckled at my computer). I actually I did think it was funny, but felt the need to explain better for everyone's benefit, too.

 

[EarthDog]

Back in 2012 for people who were pushing 1440p, that 256bit bus on 670 was holding them back from doing so without stutter and 7970 performed a lot better on similar price points.

Something similar is happening today with current gen GPU's and 4k.

Sorry but I really can't agree it's a non-issue. Bandwidth is a hot topic every gen and for 900 series it shows that the actual energy savings are made by advances in the way they treat and use the bus width We also know from the microstutter chit-chat that simple FPS tables don't say jack shit about actual gaming smoothness, and memory plays a major role in this.

I'll also disagree with some points...

1. It wasn't the bus so much as it was the 2GB vram on those cards. The 4GB 670's did just fine there (for my uses) (and so does the 770 4gb). I do not recall an article that actually pinned down the problem to the bus...but then again, the interwebs are huge and I may have missed it... and my 'butt dyno' could be off too, LOL!
2. Energy savings... cool. Sorry, I am caffeine free this morning. Can you associate that point with what we are talking about? Sorry..
3. Correct FPS tables do not show the whole story. Then again, I do not recall any frametime reviews/studies/articles on bandwidth. Link me so I may learn.

 

[rtwjunkie]

Overall, I'm going to have to agree with EarthDog, in that with the new GPU's, and most instances of 7 series Kepler, bandwidth is not so much an issue now. If the memory is fast enough, coupled with the compression, 256-bit is enough. The primary thing holding cards back from higher resolutions was the amount of VRAM. Now with Maxwell, we've got a minimum of 4GB VRAM at 7000Ghz. I think the future is here!

That being said, I'd still opt a wider bus if a model came out with it.

 

[Vayra86]

1. Bus width has a direct impact on the amount of texture streaming that is possible. Less VRAM on the card will mean more texture streaming (and swapping) is necessary to maintain a smooth gaming experience. So the 4GB cards handled that resolution better. A 2GB card with a larger bus would have handled it just as well (or would at least show better smoothness). But slapping an additional 2GB VRAM on it is cheaper than adjusting bus width. The 7970 proves this point by combining 3 GB with a 384 wide bus and offering the best performance at the time for high res/high AA.

2. Not directly associated, just pointing out that bus width is an important factor in GPU design and also in required power. You could connect this to overclockability though. Smaller bus = less power = more OC headroom for GPU/mem.

3. Frame delivery depends on the entire rendering pipeline. This ties in to 1. You can have 60 fps average with stutters and a shit gaming experience.

I do agree that for the average gamer/buyer this is a worthless discussion and in 95% of the cases GPU/bus width are balanced. But this is a tech site. We whine about that last 2% of performance. So that's what I do