• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

Difference between dual channel and 128-bit GPU memory?

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

SPL Tech

Member
Joined
Nov 28, 2006
I am trying to understand the correlation between system RAM used in dual channel, and the multiple pathways used on GPUs. So a 128-bit GPU has four RAM modules, two of which are in parallel to increase capacity, right? So each module is 32 bit, and they are DDR modules, so they operate with 64 "effective bits (32 bit x2 transfers per clock cycle)," and with two groups of RAM modules in parallel (two RAM modules in parallel with two more), we get an "effective" transfer of 128 bit. Okay, so how is that any different than standard "dual channel" system RAM? It's not, right? Both dual channel DDR3 and 128-bit GDDR5 cards essentially operate in "dual channel" mode, right?

So then if we extend this scenario to 192-bit GPUs, those operate the same way as triple channel system RAM, no? And 256-bit cards are effectively "quadruple" channel, right?

Last, if my 192-bit cards are effectively "triple channel," why is the effective speed of the RAM 6000 MHz. and the speed in Precision X listed as 3000 MHz? Shouldent the individual RAM modules be 1/3rd the effective speed, so 2000MHz.?
 
Last edited:
One thing about video memory buss compared to system memory bus is that you can have 192 bit triple channel and the information the CPU requests might only be 32 or 64 bits however transfer of the data does not transfer any faster with dual or triple because of the small data set needed sometimes.

Video card bus take more advantage of 256-Bit and 512-Bit bus because of the repeated textures and framing, another thing about Vram that is different it can be written to and read at the same time and that goes over the bus.
 
Last edited:
Okay, but if a 192-bit card is "triple channel," then why does Precision show a memory clock speed that is half the effective speed? Shouldent it be 1/3rd the effective speed?
 
I've never heard of a video cards memory architecture being dubbed as ANY channel. It works a bit different than system ram in that sense
 
So then what is the difference between 128-bit GDDR and dual channel RAM? I thought in both cases the processor has the ability to access two RAM chips at once. tripple channel / 192 bit = access 3 RAM chips at once, quadruple channel/ 256-bit= access 4 RAM chips at once, and so on and so fourth.
 
So then what is the difference between 128-bit GDDR and dual channel RAM? I thought in both cases the processor has the ability to access two RAM chips at once. tripple channel / 192 bit = access 3 RAM chips at once, quadruple channel/ 256-bit= access 4 RAM chips at once, and so on and so fourth.

You have dual channel Ram on the board. This is a feature the Cpu must have on the IMC in order to run. My Phenom II does not have a triple channel controller therefore can only access Random memory in channel one or two.

Your processor is only up to 64 bits. You cannot pass 128 bits of data through the cpu.

The Gpu can be 128bit and the software is written 128bit. That's data bandwidth brother. Just because you have 256bit on the video card does not mean that video card has triple channel memory.

Lets say your using an IGP APU ok. It's using system Ram. If your running a 64 bit OS, then the memory is going to run and access in 64 bits IF the program is written in 64bits. If not in a 64bit environment the processing will happen in 32bit. You still only have dual channels to select from and that's a random access. The cpu can randomly select which channel to read and write to.

So since the IGP isn't 128bit, that doesn't mean anything to the way the GPU processes information.

I guess another way to put it would be.....

Channels are like car lanes.... and the bigger the bit, the wider that lane is to let more traffic through.
 
Your processor is only up to 64 bits. You cannot pass 128 bits of data through the cpu.

No. Dual-channel means the CPU is accessing the memory on two combined 64-bit channels. That's 128 bits. Even old Athlon XP 32-bit CPUs supported dual-channel, and both AMD and Intel's high-end server chips support quad-channel (256-bit). Addressing (32- or 64-bit) is not equivalent to bus bandwidth.

Okay, but if a 192-bit card is "triple channel," then why does Precision show a memory clock speed that is half the effective speed? Shouldent it be 1/3rd the effective speed?

Bus width and frequency are not inversely related. All SDRAM (which DDR is a variant of) use a 64-bit bus. You are correct that 192-bit means it is technically "triple-channel", but it is still DDR, meaning that it can transfer data twice in one clock cycle, therefore its "effective" speed (if it weren't DDR) is twice its physical clock speed.

If they are both operating at 1066 MHz (533 MHz * DDR), the difference there between 192-bit and 256-bit would be that the 256-bit bus can transfer 34112 MB per second, while the 192-bit bus can only transfer 25584 MB per second. Note my math may be a bit off, given tricks recent DDR versions do to increase bandwidth even more, but the ratio is the same.
 
Last edited:
Okay, thanks everyone, got it now. So is there a bus speed between the GDDR chips and the GPU memory controller, or is the "bus" just physical contacts with no clock rate limit? I am wondering if it is possible to get into a senario where you overclock the GDDR so much that you saturate the pathway between the memory controller and the memory, or if that would be impossible because the bus has no clock speed limit like a PCI-E bus does.
 

Whoever made that graphic typo'd a bit... that should be 0-63,64-127 or 1-64,65-128.

I am wondering if it is possible to get into a senario where you overclock the GDDR so much that you saturate the pathway between the memory controller and the memory

Theoretically, yes. Practically, no, because the required voltage increase releases the magic smoke.
 
Last edited:
Back