Is GPU Memory Bottlenecking really a thing?

[Evann]

So everyone is talking about awesome HBM 2.0 and its insane speed, but is it really necessary? 

 

i never heared someone complaining about its "slow GDDR5" - so why is slow memory bandwith a problem?

 

Right now, on recent game titles, is it really a thing?

 

Do you "feel" inside a game if you have a higher memory speed? And if - how?

 

How do you notice something like "Memory Speed Bottlenecking"?

 

 

[LeapFrogMasterRace]

Memory bottlenecking is definitely a thing kind of complicated to explain I would look around google if you are interested.

 

I found a decent forum post  http://www.techpowerup.com/forums/threads/article-just-how-important-is-gpu-memory-bandwidth.209053/

[tbake0155]

I have an r9 290x.  It has 4 GB gddr5 memory and a 512-bit memory bus.  The memory is at 1300 MHz, which yields 332 GB/s of memory bandwidth.

 

The memory will overclock to 1600 MHz stable enough to benchmark.  There are definite increases in performance all the way up to 1600 from 1300.  At 1600 MHz the total bandwidth is around 400 GB/s.

 

It seems like even the old hawaii chip would perform better with faster ram.

[Evann]

2 minutes ago, tbake0155 said:

I have an r9 290x.  It has 4 GB gddr5 memory and a 512-bit memory bus.  The memory is at 1300 MHz, which yields 332 GB/s of memory bandwidth.

 

The memory will overclock to 1600 MHz stable enough to benchmark.  There are definite increases in performance all the way up to 1600 from 1300.  At 1600 MHz the total bandwidth is around 400 GB/s.

 

It seems like even the old hawaii chip would perform better with faster ram.

what do you mean with "performance increase"? --> more FPS?

[tbake0155]

@Evann

 

I meant specifically the score in firestrike, which I'm sure translates to FPS as well.  I can't say it's a massive increase but there's a noticeable gain from overclocking the ram.

 

Give me a few minutes I'll do some comparisons in firestrike and post them.

[Malaki Titi]

Quite possible.

However with majority of the gpus like the 960 and 390, both are choking to death before it could make full use of all their vram.

[hex4]

Historically nVIDIA's Fermi GTX 480 / 580 1.5GB cards are perfect examples of memory starved cards, very powerful GPU with not enough Vram in some circumstances.

[Evann]

16 minutes ago, hex4 said:

Historically nVIDIA's Fermi GTX 480 / 580 1.5GB cards are perfect examples of memory starved cards, very powerful GPU with not enough Vram in some circumstances.

sry of the title was missleading, i am talking about memory bandwith.

[hex4]

Just now, Evann said:

sry of the title was missleading, i am talking about memory bandwith.

Oh.

 

I shall now leave

[tbake0155]

Maybe not as drastic as I thought but here are the results.  This is with stock core clock of 1040 MHz. I might run a few more with the core overclocked to see if it has a bigger effect.

 

The memory clocks I tested are 700 MHz, 1000 MHz, 1300 MHz, 1600 MHz and 1650 MHz.  1700 MHz is not stable. 1600 MHz probably won't be stable with the core overclocked.

 

The scores do scale with the memory clock.

[Evann]

2 minutes ago, tbake0155 said:

Maybe not as drastic as I thought but here are the results.  This is with stock core clock of 1040 MHz. I might run a few more with the core overclocked to see if it has a bigger effect.

 

The memory clocks I tested are 700 MHz, 1000 MHz, 1300 MHz, 1600 MHz and 1650 MHz.  1700 MHz is not stable. 1600 MHz probably won't be stable with the core overclocked.

 

The scores do scale with the memory clock.

Thanks! 

 

So how do you think is this going to look like with HBM 2.0?^^

[tbake0155]

I have no idea!  But I definitely think faster memory is a good thing, for GPUs and for CPUs as well.

[Cyracus]

From my understanding memory amount effects what res and quality you can get while memory bandwidth effects your fps, insofar as the GPU is capable of using, having more than that has little effects, but not having enough is bad. Probably not perfect, but close enough, as games get more intensive and GPUs get stronger you'll need a combination of the two

[sgloux3470]

Memory bandwidth plays a role for higher resolutions.

 

This is most obvious when you look at relative performance difference.  If you compare the 980 to say the 390 at 1080p and then again at 4K, you will see that they are much more closely matched at 4K.  However if you compare that 390 to say the Fury X the relative performance difference will remain fairly consistent. (Same deal comparing 980 to say 970)

 

This is because Maxwell has memory bandwidth limitations at 4K.  

 

HBM so far hasn't been a huge game changer in real world benchmarks as far as I've seen.