GTX 1060 - 192-bit Memory Bus Width Impact

[FrankV]

Quote

Whereas you’d expect half of a GP104 to ship with a 128-bit memory bus, NVIDIA has defied expectations by giving GP106 a larger 192-bit memory bus, giving the chip 50% more memory bandwidth per CUDA core, all things held equal. 

The above quote is from this article: http://www.anandtech.com/show/10474/nvidia-announces-geforce-gtx-1060-july-19

 

My question is what is the impact of the Memory Bus Width. How does that impact things? 

 

In case it's relevant, I currently have a AND 7950 and am strongly considering upgrading to the GTX 1060 in the coming months (not at release, and it depends on reviews and performance, etc)

[xAcid9]

We'll see. 

[BoldarBlood]

I believe in the simplest, non-technical terms; It's how wide the highway is carrying the data from the GPU to the vram and back.  The more data that can flow back and forth, the more the card can draw.  But there must be a balance.  Giving the 1060 a 512-bit memory bus would be useless because the GPU couldn't process that much data.  Giving a GTX 1080 a 64bit memory bus would leave the GPU starved for data that had been loaded to vram.

 

Memory speed is also in the equation... the 1080 actually has a reduced bus width to 256bit over the 980ti's 384bit but because the memory and architecture is faster, the actually throughput of the memory is almost the same 320GB/s vs 337GB/s.

 

or something like that... 

[Imakuni]

16 minutes ago, FrankV said:

The above quote is from this article: http://www.anandtech.com/show/10474/nvidia-announces-geforce-gtx-1060-july-19

 

My question is what is the impact of the Memory Bus Width. How does that impact things? 

 

In case it's relevant, I currently have a AND 7950 and am strongly considering upgrading to the GTX 1060 in the coming months (not at release, and it depends on reviews and performance, etc)

It's just a number. In theory, more is "better". But it's been a long while since we could look at specs and say "this matters, and the card is bad / good because of it".

 

It's not as if we could do anything about it. If we could pick between the same card with a bigger / smaller bus, yeah, maybe the question would make sense. But we can't, so don't sweat about it.

[LAwLz]

2 hours ago, FrankV said:

My question is what is the impact of the Memory Bus Width. How does that impact things? 

It will affect the memory bandwidth of the card (how fast things ca be read/written to the VRAM).

In order to calculate the bandwidth you take the memory frequency multiplied by the bus width.

 

 

I will use my 7850 as an example:
 

Spoiler

 

The 7850 has a 256 bit bus and a memory clock of 1200MHz.

 

 

First we need to change from the actual clock rate to the "effective clock rate" on my card. Since my card uses GRRD5 we should multiply the 1200MHz by 4, since GDDR5 can send 4 signals per clock cycle. So my card has an "effective" clock rate of: 1200 * 4 = 4800MHz.
 

Now let's just multiply the two numbers (width and clock rate) together. Since we didn't convert the MHz to Hz, the number we get will be in megabits.

256 * 4800 = 1228800

 

1228800 megabits are 153.6 gigabytes.

 

My card has a memory bandwidth of 153.6 gigabytes per second.

 

 

Here is a formula:

(((X * 4) * Y) / 8) / 1000 = Z

 

X = memory clock rate in MHz (not effective)

Y = memory bus width

Z = Total memory bandwidth in Gigabytes per second.

 

 

So as you can probably tell, the bus width is directly related to the speed at which the card can read/write to the VRAM. Basically, higher number is faster.

You can't just look at the memory bus width to understand how well/poorly a card will perform though. For example the the 1060 might have a narrower bus than my 7850, but it is clocked much higher which means that it still ends up having more bandwidth (192GB/s vs 153.6GB/s).

 

Your 7950 has an effective clock rate of 5GHz, and a 384 bit wide bus. That translates to 240GB/s memory bandwidth. That does not necessarily mean it will be faster than the 1060 though. Memory is just one part of the card, and if the other parts are faster then the end result might be a faster card. It's kind of like how adding an SSD to your computer might not increase your FPS.

 

Another thing to take into consideration is compression. I am pretty sure Nvidia is ahead of AMD when it comes to color compression and the 7950 is quite a bit older than the 980, so let's say that the 1060 will be able to compress things an average of 40% better than your 7950 (AMD themselves actually claimed 40% higher memory bandwidth efficiency between GCN 1.1 and 1.2).

So at 40% better compression, the 192GB/s bandwidth will be as fast as 270GB/s would be on your 7950. That means that despite having lower memory bandwidth, the effective bandwidth is still 12.5% higher.

 

 

The Too Long Didn't Read is this: You should not pay that much attention to the memory bus width because:

1) A card with a low memory bus width might still have a higher transfer speed than a card with a wider bus.

2) Memory bandwidth (the #GB/s number) might not be comparable between two different cards. 100GB/s might be able to compete with 140GB/s on a different card.

3) Much faster memory might not matter if the bottleneck is somewhere else on the card (like the GPU core).