Previous | Next --- Slide 42 of 79
Back to Lecture Thumbnails
neonachronism

Are latency and bandwidth totally orthogonal, or is there some cross-influence? It seems that if latency is lowered, you should be able to make more requests in a given period of time, increasing bandwidth.

On a similar note, if you need n loads, can you make n requests immediately, or do you need to make the sequentially?

tcm

Latency and bandwidth are related in some cases; in particular, when there isn't enough available bandwidth, the contention starts to increase latency. (This becomes especially bad as you approach saturation of the bandwidth.) This is a well-known effect in networking, for example.

Regarding whether latency affects bandwidth, the short answer is not really, unless you are making certain assumptions about how system works. Those assumptions are likely to lead to incorrect intuitions about how memory performance works. So I would recommend that you think of latency as not having a direct impact on bandwidth (in general).

Regarding whether loads need to happen sequentially or whether they can occur in parallel, in modern processors, loads are "non-blocking", which means that until you actually use the load result, you can usually continue executing. Hence it is possible to overlap multiple simultaneous load misses, and modern processors are designed to take advantage of this.

lusiliang93

When I looked for information of NVIDIA GTX 1080, there are two terminology to describe the data transfer rate. The one is memory bandwidth with the unit of GB/s(For 1080,it is 320GB/s). The other is memory speed with the unit of Gbps(For 1080, it is 10Gbps). Is there any specific reason why we need two terminology to evaluate the performance? Thanks!

JJ

I think memory bandwidth is the speed you transfer data from/to memory, memory speed can be operations in the memory for example memset etc. @lusiliang93