Maxwell relevance to system memory speed

[A/C]

After skimping over an article over at anandtech about memory speeds on haswell (mostly TLDR), I thought about the RAM speed relevance with nvidia's upcoming Maxwell cards. http://www.anandtech.com/show/7364/memory-scaling-on-haswell/10

 

Nvidia promised unified memory access with Maxwell, so naturally the memory speed and timings would be very relevant to the performance of the card, am I right?

 

According to the article the sweet spot for haswell igpu performance is ram 2400 with cl9 timings (small wonder at that small cas latency). Does anyone here have enough knowledge about the unified memory technology in order to enlighten us about the matter?

[Dravic]

After skimping over an article over at anandtech about memory speeds on haswell (mostly TLDR), I thought about the RAM speed relevance with nvidia's upcoming Maxwell cards. http://www.anandtech.com/show/7364/memory-scaling-on-haswell/10

 

Nvidia promised unified memory access with Maxwell, so naturally the memory speed and timings would be very relevant to the performance of the card, am I right?

 

According to the article the sweet spot for haswell igpu performance is ram 2400 with cl9 timings (small wonder at that small cas latency). Does anyone here have enough knowledge about the unified memory technology in order to enlighten us about the matter?

 

No, GPUs always have GDDR5 nowadays which STOMPS all OVER the DDR3 we have in PCs.

 

I'm not sure here, but I think that when a GPU has some vram, it has to be mirrored to the regular RAM we have in PC so the processor can feed the GPU with the vRAM when needed. Now... I think, I'm not sure - mind you!, that the unified Memory Access will allow seamless connection of CPU in our system to the memory inside GPU and we no longer will need a vram mirror inside system RAM.

 

Also, the CPU will be able to address the memory of GPU (MUUUUCH Faster but the connection through PCIE will be slow and latency will kill most benefits), as well as the GPU will be able to address system RAM seamlessly.

 

But don't quote me on that, that's what I understood using my small GPU knowledge.

 

But overally, GPUs always could and will be able to access system RAM, do you know why? Because that's what happens when the GPU cannot store more in its memory - it gets more vRAM by accessing system memory directly and transfering files before rendering the next frame it needed the memory for. If you lack even freakin 100mb of vram on the GPU all your framerate will freakin DROP.

[A/C]

No, GPUs always have GDDR5 nowadays which STOMPS all OVER the DDR3 we have in PCs.

 

I'm not sure here, but I think that when a GPU has some vram, it has to be mirrored to the regular RAM we have in PC so the processor can feed the GPU with the vRAM when needed. Now... I think, I'm not sure - mind you!, that the unified Memory Access will allow seamless connection of CPU in our system to the memory inside GPU and we no longer will need a vram mirror inside system RAM.

 

Also, the CPU will be able to address the memory of GPU (MUUUUCH Faster but the connection through PCIE will be slow and latency will kill most benefits), as well as the GPU will be able to address system RAM seamlessly.

 

But don't quote me on that, that's what I understood using my small GPU knowledge.

 

Yes, that makes sense. I suppose then this would only work on pci-e 3.0.

 

But overally, GPUs always could and will be able to access system RAM, do you know why? Because that's what happens when the GPU cannot store more in its memory - it gets more vRAM by accessing system memory directly and transfering files before rendering the next frame it needed the memory for. If you lack even freakin 100mb of vram on the GPU all your framerate will freakin DROP.

 

Yes I knew that, that's mostly where framedrops are from, found that out the hard way using 512mb card a while ago.

[JAKEBAB]

 

But overally, GPUs always could and will be able to access system RAM, do you know why? Because that's what happens when the GPU cannot store more in its memory - it gets more vRAM by accessing system memory directly and transfering files before rendering the next frame it needed the memory for. If you lack even freakin 100mb of vram on the GPU all your framerate will freakin DROP.

 

 

so lets use ddr5 for example and lets say its as fast as gddr5 and whatever but goes in the dim slots on your mobo that vram on a graphics card wouldnt really matter because if the graphics card didnt have enough memory it would just pull from the system memory making u have like 64gbs of vram ? 

 

(i know this will like never happen or by the time it does graphics cards will have loads of vram it wont even matter)

[Dravic]

so lets use ddr5 for example and lets say its as fast as gddr5 and whatever but goes in the dim slots on your mobo that vram on a graphics card wouldnt really matter because if the graphics card didnt have enough memory it would just pull from the system memory making u have like 64gbs of vram ? 

 

(i know this will like never happen or by the time it does graphics cards will have loads of vram it wont even matter)

 

The latency, sir. There is a latency if you are not soldering the vram right onto GPU.

[JAKEBAB]

The latency, sir. There is a latency if you are not soldering the vram right onto GPU.

yea i know because gddr5 is ddr3 on steroids and loads of latency but what im saying if there was no latency and stuff would that mean we have loads of vram?

[A/C]

yea i know because gddr5 is ddr3 on steroids and loads of latency but what im saying if there was no latency and stuff would that mean we have loads of vram?

 

well, in a nutshell - yes.