Difference in determining Memory Bandwidth for Ram (lets say DDR3) and Vram (GDDR5)?

bigman3213

Honorable
Dec 14, 2012
11
0
10,510
Okay, so I am learning more and more about Memory clocks/Bandwidth/timings/etcetera.

I am learning mainly about Ram right now. But I have learned that:
bandwidth = DDR clock rate x bits transferred per clock cycle / 8
Memory uses 64 bits transferred per clock cycle, at least up to DDR3 to my knowledge.

So it's not too difficult to figure out that 1333 x 64 = 10664 (which is usually how some vendors mark their product memory). 1866MHz would be 14928, 2133MHz would be 17064 and so we get the numbers (in order of what I posted): 1333MHz = 10.6GB/s, 1866MHz = 14.9 GB/s, and 2133MHz = 17GB/s Bandwidth.

And I thoroughly understand how the timings work and what not.

So my question is, how is GDDR5 determined? Rather, how is the Video RAM determined?
And how? I know this is different, because when I go look at vendor specs of a GPU, I would see lets say 6610MHz (effective) and the memory bandwidth would end up being 105.76 GB/s. Which, if you use the equation given, that would end up being (DDR clock rate x 8) x 2.

So once more, Why and How does this happen?

Thanks ahead for responses!
 
Solution
System RAM is 64bits per memory channel, so 128 bits for most desktop platforms. That 17GiB/s is the theoretical bandwidth per channel, I think. Haswell gets about 20GiB/s actual bandwidth across both channels (for DDR3-1600, two channels), so I'm not sure why it's about half of what you'd expect. Possibly it writes and reads back the data, or something.

memory-bandwidth.png


*note the 3930k is LGA2011, with quad-channel memory.

VRAM depends on the GPU; the memory bus can be anywhere from about 64 bits to 384 bits, depending on which GPU you have.
System RAM is 64bits per memory channel, so 128 bits for most desktop platforms. That 17GiB/s is the theoretical bandwidth per channel, I think. Haswell gets about 20GiB/s actual bandwidth across both channels (for DDR3-1600, two channels), so I'm not sure why it's about half of what you'd expect. Possibly it writes and reads back the data, or something.

memory-bandwidth.png


*note the 3930k is LGA2011, with quad-channel memory.

VRAM depends on the GPU; the memory bus can be anywhere from about 64 bits to 384 bits, depending on which GPU you have.
 
Solution

bigman3213

Honorable
Dec 14, 2012
11
0
10,510

So for the A10 (lets say) Maximum is 17, but because of what the cpu is allowing it, it's only pushing 14.19 GB/s?

So Bus Width is the determining factor for data transferred per clock cycle?