blah blah blah memory controllers blah blah blah...
I'll add...
256 bit memory controller = 64 bit memory controller x 4
384 bit memory controller = 64 bit memory controller x 6
512 bit memory controller = 64 bit memory controller x 8
1x64 bit memory controller @ 600Mhz (1200 Mhz effective) yields 76,800,000,000 bits per second throughput.
4x64 bit memory controller @ 600Mhz (1200 Mhz effective) yields 307,200,000,000 bits (38.4 GBytes) per second.
4x64 bit memory controller @ 650Mhz (1300 Mhz effective) yields 332,800,000,000 bits (41.6 GBytes) per second.
6x64 bit memory controller @ 600Mhz (1200 Mhz effective) yields 460,800,000,000 bits (57.6 GBytes) per second.
8x64 bit memory controller @ 600Mhz (1200 Mhz effective) yields 614,400,000,000 bits (76.8 GBytes) per second.
In order to support a 256 bit memory system, you need to spend X amount of money on RAM. Assume 256 MB.
In order to support a 384 bit memory system, you need to spend 1.5X the amount of a 256 bit system, but you also get 1.5X the total memory in the process. 1.5 x 256 MB = 384 MB.
In order to support a 512 bit memory sysem, you need to spend 2X the amount of a 256 bit system, but you also get 2X the total memory in the process.
Let's assume you have 16 bit memory modules with 8 bit interfaces and you are using two of these for your memory. You stip your data across these modules 8 bits at a time. Module 0 will hold bits 0 through 7, then bits 16 through 23. Module 1 will hold bits 8 through 15, then bits 24 through 31.
With this in mind, 8 bit memory controller 0 can transfer bits 0 through 7 from memory module 0 and at the same time memory controller 1 can transfer bits 8 through 15 from memory module 1.
Same thing works at a larger scale for 256, 384, and 512 bit memory controllers accessing hundeds of MB of memory.
The other benefit of using wider memory systems is that you don't have to use cutting edge memory in order to get better performance.
For instance: 384 bit memory controller with 384 MB of 400Mhz GDDR3 memory yields the same throughput of a 256 bit memory controller with 256 MB of 600Mhz GDDR3 memory. The cost of 384 MB of 400Mhz GDDR3 could be less than 256 MB of 600Mhz GDDR3.
384 bit controller with 384 MB of 500Mhz GDDR3 will give you some 48GBps, which is much higher than what you get with 600Mhz or 650Mhz GDDR3 and a 256 bit memory controller.
Course, to the GPU maker, adding these extra 64 bit controllers takes up die space and makes PCBs that much more complicated thus increasing costs for the GPU and for the PCB.
I expect wider memory buses sooner than later unless GDDR3/4 memory clocks increase by decent margins in the near future.
Those are my thoughts on memory controllers. Someone will probably shoot this down soon. It's happened before.
Facebook
Twitter