ATis' HyperMemory...

[Drayvn]

Ok ill just copy and paste what i said in the comments box.

"I think this idea has stemed from being partnered with Microsoft and IBM and working on that, as i recall the GPU on the Xbox 2 doesnt have any on board memeory and the WHOLE system uses just the RAM, this increases performance as its getting data quicker from the CPU than it having to travel through its own RAM first, but i was expecting something more, because the Xbox 2 is also touted to have pathways which link both GPU and CPU so that the GPU can get data directly from the CPU Memory. Maybe that is going to be the high end part, as they are bringing out their own mobos now for both processors with full PCI-E support, i can see this happening.

Various and numerous times i have said this was going to happen with ATI, it was plainly obvious when u read the specs and diagram from the Xbox 2. ATi have picked up a few pointers.

And hey, i thought it would be another year before we would see anything using the PCI-E stuff, and hey we got one now! Thats good in my books!"

Ok i dont know about u guys, but when i first heard the news about the Xbox 2 and all its specs and stuff, i could see something like this coming, it was so obvious when ATi partnered with IBM, Microsoft and other partners that they would pick up something like this, or unless they asked to do this in the Xbox 2, any what i wanted to say was in the quotes above. Same technology from the Xbox 2 is coming to PC. I just hope they can progress to the stage the the GPU would take information straight from the CPU as it supposedly does on the Xbox 2.

 

[Schadenfroh]

ok.......

its a faster agp aperture (already present on all video cards and motherboards) on a pci express bus.

a very similar feature has been present since AGP 2x, see this

whats the big deal?

 

[Drayvn]

AGP was essentially a one way system

PCI-E is Bi-directional, so it isnt incurring a loss of performance, as the card can process it both ways at the same speed.

AGP could do alot one way, but when it was used in two ways, it was very very poor.

 

[CaiNaM]

Originally posted by: Drayvn
AGP was essentially a one way system

PCI-E is Bi-directional, so it isnt incurring a loss of performance, as the card can process it both ways at the same speed.

AGP could do alot one way, but when it was used in two ways, it was very very poor.

but it wasn't an issue with video gfx, as it's basically a one way stream... "bi-directional" or not, a pci-e video card is no faster than it's agp counterpart.

i do believe tho that's the reason pci-e is important for nv sli - so the cards can exchange data and "load balance".

 

[zephyrprime]

I don't think this has anything in common with the Xbox 2 at all. Hypermemory is a method to transfer data from one memory pool to another. The XBox has zero use for anything like that since it only has 1 memory pool.

AGP is "one way" but it can still read or write in both directions. It's just that it can only read or write at a particular time, not both at the same time.

I agree with Schadenfroh. This hypermemory really is just a rehash of AGP memory. Remember, Intel originally made AGP specifically so that video cards could reduce their memory requirement. This never happened. I don't foresee hypermemory being of any use either until video texture memory is treated like a local cache instead of a dedicated memory and that probably won't happen until at least DX Next (if then).

 

[gururu]

Requires new hardware
'Under previous interconnect standards, the data transfer between the visual processor and the CPU was not fast enough for real-time graphics applications.'


With AGP, graphics cores couldn't access system memory fast enough under high memory loads. Now apparently through PCI-Express, system memory can be accessed at an acceptable rate.

 

[SickBeast]

Current graphics cards don't even fully utililze the bandwidth offered by 4X AGP, let alone 16X PCI-E. If they can figure out a way to make this bandwidth fully utilized, I'm all for it. I'm surprised they haven't done so already.

 

[zephyrprime]

If they can figure out a way to make this bandwidth fully utilized, I'm all for it. I'm surprised they haven't done so already.

There's just no use for it. Inventing something to use up more bandwidth is akin to asking for software that's slower.

 

[SickBeast]

Originally posted by: zephyrprime

If they can figure out a way to make this bandwidth fully utilized, I'm all for it. I'm surprised they haven't done so already.

There's just no use for it. Inventing something to use up more bandwidth is akin to asking for software that's slower.

I have a 128MB graphics card that requires setting the AGP aperture to 256MB to obtain decent Doom 3 performance. More efficient use of AGP bandwidth would surely gain me some performance. 8X AGP allows an enormous amount of data to be transferred from system memory to the graphics card, but only 25% of this bandwidth is currently being used, even on the latest cards like the X800XT.

 

[ViRGE]

Originally posted by: gururu
Requires new hardware
'Under previous interconnect standards, the data transfer between the visual processor and the CPU was not fast enough for real-time graphics applications.'


With AGP, graphics cores couldn't access system memory fast enough under high memory loads. Now apparently through PCI-Express, system memory can be accessed at an acceptable rate.

The cards couldn't access the memory fast enough period, and this wasn't an AGP problem. System memory is too slow for video card use, and always has been. In 1999, we had systems that did 133mhzx64bit memory, so an effective bandwidth rate of ~1GB/sec, while we had video cards that did 333x128bit memory, so an effective bandwidth rate of ~5GB/sec. Now we're up to >30GB/sec in video memory, but only 6.4GB/sec system memory; it's still too slow by nearly an order of magnitude.

 

[zephyrprime]

Originally posted by: SickBeast
I have a 128MB graphics card that requires setting the AGP aperture to 256MB to obtain decent Doom 3 performance. More efficient use of AGP bandwidth would surely gain me some performance. 8X AGP allows an enormous amount of data to be transferred from system memory to the graphics card, but only 25% of this bandwidth is currently being used, even on the latest cards like the X800XT.

What makes you say that only 25% of the bandwidth is being used? I think you have a misconception about AGP. There's nothing wrong with AGP and cards can certainly use more than 25% of the bandwidth of AGP. When people say that cards don't use all of AGP's bandwidth, they don't mean cases like yours. They mean cases where the amount of texture memory and frame buffer memory fits entirely onto the card. In cases like this, a limited test with one game have shown that that one game used only about 166MB/s of bandwidth.

In your case, where you're trying to load 256MB of textures, you will will be putting a lot of traffic on the AGP bus. The only thing that could be done to increase traffic utililzation in cases like yours would be if the card maker would implement SMT on their video cards so that memory latencies across the AGP could be masked.

 

[gururu]

Originally posted by: ViRGE

Originally posted by: gururu
Requires new hardware
'Under previous interconnect standards, the data transfer between the visual processor and the CPU was not fast enough for real-time graphics applications.'


With AGP, graphics cores couldn't access system memory fast enough under high memory loads. Now apparently through PCI-Express, system memory can be accessed at an acceptable rate.

The cards couldn't access the memory fast enough period, and this wasn't an AGP problem. System memory is too slow for video card use, and always has been. In 1999, we had systems that did 133mhzx64bit memory, so an effective bandwidth rate of ~1GB/sec, while we had video cards that did 333x128bit memory, so an effective bandwidth rate of ~5GB/sec. Now we're up to >30GB/sec in video memory, but only 6.4GB/sec system memory; it's still too slow by nearly an order of magnitude.

too slow for what?
With PCI Express's big jump in... seems to make sense.
I agree with techreport. There is a lot more bandwidth that can be used now than with AGP, making system memory more useful than ever for tasks not requiring local video memory.

 

[Genx87]

Only good for the integrated graphics arena. If the GPU has to swap across even the PCIE bus you will see a hit in performance.

Maybe ATI should put their money where their mouth is. Put an X800XT out with 64MBs of memory and use this technology

 

[chsh1ca]

Consider that while the top end cards break 30GB/sec bandwidth, the low end cards are still sitting around that 6-10GB/sec bandwidth range -- The X300 is 6.4GB/sec, the X600 is 9.6GB/sec, and the FX5200 is 6.4GB/sec (nice to see NVidia's mid-range being so kind in terms of memory bandwidth ). Both companies sell loads more of their low-end cards than their enthusiast parts, so while this may not help out a card like the X800XT, it will definitely offer something for the lowest end cards.

 

[ViRGE]

Originally posted by: gururu

Originally posted by: ViRGE

Originally posted by: gururu
Requires new hardware
'Under previous interconnect standards, the data transfer between the visual processor and the CPU was not fast enough for real-time graphics applications.'


With AGP, graphics cores couldn't access system memory fast enough under high memory loads. Now apparently through PCI-Express, system memory can be accessed at an acceptable rate.

The cards couldn't access the memory fast enough period, and this wasn't an AGP problem. System memory is too slow for video card use, and always has been. In 1999, we had systems that did 133mhzx64bit memory, so an effective bandwidth rate of ~1GB/sec, while we had video cards that did 333x128bit memory, so an effective bandwidth rate of ~5GB/sec. Now we're up to >30GB/sec in video memory, but only 6.4GB/sec system memory; it's still too slow by nearly an order of magnitude.

too slow for what?
With PCI Express's big jump in... seems to make sense.
I agree with techreport. There is a lot more bandwidth that can be used now than with AGP, making system memory more useful than ever for tasks not requiring local video memory.

It's too slow because now we need more bandwidth. The system/dedicated ratio is still about 1:5, and this isn't taking in to consideration that most of that system bandwidth is needed by the system, a by-product of the continuing disparity between increases in bandwidth and processor speed. <6.4GB/sec won't get you anywhere when you're dealing with 128MB texture sets.

 

[BFG10K]

as i recall the GPU on the Xbox 2 doesnt have any on board memeory and the WHOLE system uses just the RAM, this increases performance as its getting data quicker from the CPU than it having to travel through its own RAM first,

It doesn't make it faster, it makes it slower. VRAM is a lot faster than system RAM and consoles like to share memory like that because of cost reasons, not performance reasons. It's similar to integrated video on PCs.

As for main memory, fastwrites ensure it's not touched for duplicate reasons.

 

[gururu]

Originally posted by: chsh1ca
Consider that while the top end cards break 30GB/sec bandwidth, the low end cards are still sitting around that 6-10GB/sec bandwidth range -- The X300 is 6.4GB/sec, the X600 is 9.6GB/sec, and the FX5200 is 6.4GB/sec (nice to see NVidia's mid-range being so kind in terms of memory bandwidth ). Both companies sell loads more of their low-end cards than their enthusiast parts, so while this may not help out a card like the X800XT, it will definitely offer something for the lowest end cards.

that is a worthwhile point, although I wonder if such a scenario will end up being as ridiculous as a 9600pro with 256mb memory.
I don't know what the reality of this feature will be. I doubt anybody does, since ATI hasn't completely divulged their solution. I guess I am just optimistic about any innovations that may make gaming more enjoyable. Whether they pan out or not is another story.

The other point I would make is this. Those who are skeptical are basing their arguments on existing architectures and bus speeds, without knowing what sort of hardware or software implementations ATI, IBM, and microsoft together have come up with. I think its a sorry trend when people so quickly discount an innovation before even giving the innovator a chance to show there stuff. It's like all those people that so quickly discount SLI.