How important is GPU memory vs. memory bandwidth?

[Presidium]

I am comparing two AMD GPUs: the R9 270 and R9 270X. I am looking at the two side-by-side in the specs sheet, and I am only noticing two differences between the GPUs:

• 
  - GPU Clock speed (270X's is slightly higher)
  - Memory amount (2GB in 270, 4GB in 270X)

I thought more memory resulted in higher memory bandwidth, but both of them have 179.2. Would someone mind clearing this up for me? Is throughput the same as bandwidth, and what does the 270X really have over the 270 performance-wise? And what does the slightly higher clock speed for the 270X really mean?

Thanks.

 

[pasow]

bandwidth is a relation of the clocking speed, bus width, and memory type. the amount of it is inconsequential to its bandwidth so long as its quantity is a power of its bus width. otherwise you end up with these weird performance differences like Nvidia is currently in litigation over on its 970.

generally the hierarchy is, memory type > bus width > clocking speed.


edit...
it looks like the R9 270x is around 10% faster than the R9 270 with the same quantity of memory.

 

[Gam3r01]

None of them serve any relative performance difference in relation to architecture.
The bus width means very little outside of avoiding bottlenecks, the amount of memory only matters when increasing resolution, and clock speeds would be the most meaningful out of the 3.

I do not recommend getting the larger memory amount on such a midrange card, it serves you no benefit.

 

[Presidium]

bandwidth is a relation of the clocking speed, bus width, and memory type. the amount of it is inconsequential to its bandwidth so long as its quantity is a power of its bus width. otherwise you end up with these weird performance differences like Nvidia is currently in litigation over on its 970.

generally the hierarchy is, memory type > bus width > clocking speed.


edit...
it looks like the R9 270x is around 10% faster than the R9 270 with the same quantity of memory.

So is there a calculation that you can do on the GPU's specs (like memory amount X clock speed X bus width, or something like that) so get a general sense of the amount of throughput you get?

Basically, I'm wondering how "throughput" is calculated. If a card with DDR3 RAM has the same throughput as a card with GDDR5 RAM, is there a difference beside being able to say that the second GPU has "better" RAM?

 

[RobCrezz]

bandwidth is a relation of the clocking speed, bus width, and memory type. the amount of it is inconsequential to its bandwidth so long as its quantity is a power of its bus width. otherwise you end up with these weird performance differences like Nvidia is currently in litigation over on its 970.

generally the hierarchy is, memory type > bus width > clocking speed.


edit...
it looks like the R9 270x is around 10% faster than the R9 270 with the same quantity of memory.

So is there a calculation that you can do on the GPU's specs (like memory amount X clock speed X bus width, or something like that) so get a general sense of the amount of throughput you get?

Basically, I'm wondering how "throughput" is calculated. If a card with DDR3 RAM has the same throughput as a card with GDDR5 RAM, is there a difference beside being able to say that the second GPU has "better" RAM?

Well the (hypothetical) GDDR5 card would have a much smaller memory bus than the DDR3 card if they had similar bandwidth, so the GDDR5 card would likely use significantly less power.

Also, you have to consider compression these days. Nvidia manage to avoid using large power hungry memory busses on modern cards due to some clever colour compression, which effectively increases the "bandwidth" by using the GPU power to reduce the amount of memory bandwidth used.

 

[Presidium]

bandwidth is a relation of the clocking speed, bus width, and memory type. the amount of it is inconsequential to its bandwidth so long as its quantity is a power of its bus width. otherwise you end up with these weird performance differences like Nvidia is currently in litigation over on its 970.

generally the hierarchy is, memory type > bus width > clocking speed.


edit...
it looks like the R9 270x is around 10% faster than the R9 270 with the same quantity of memory.

So is there a calculation that you can do on the GPU's specs (like memory amount X clock speed X bus width, or something like that) so get a general sense of the amount of throughput you get?

Basically, I'm wondering how "throughput" is calculated. If a card with DDR3 RAM has the same throughput as a card with GDDR5 RAM, is there a difference beside being able to say that the second GPU has "better" RAM?

Well the (hypothetical) GDDR5 card would have a much smaller memory bus than the DDR3 card if they had similar bandwidth, so the GDDR5 card would likely use significantly less power.

Also, you have to consider compression these days. Nvidia manage to avoid using large power hungry memory busses on modern cards due to some clever colour compression, which effectively increases the "bandwidth" by using the GPU power to reduce the amount of memory bandwidth used.

Thanks for the reply. This makes sense to me.

One last thing I'm still confused on, though: where does throughput fit into all this?