ATI Develops HyperMemory Technology to Reduce PC Costs

[trinibwoy]

No your right in a few months both major platforms will be moving to ddr 2 ram. Do you want to see the prices of a 1 gig dimm for that ?

Not to go too off topic but I strongly disagree. There are no signs that the AMD platform is moving to ddr2 'in a few months'. Although we may have differing definitions of 'few'. The change to ddr2 will involve changes to the on die memory controller and with the current price delta between ddr1 and ddr2 and the presently non-existent performance advantage of ddr2 there is no incentive for amd to make that move anytime soon.

 

[Richard]

Hmm is this yet another step towards longhorn's WGF's standards (specifically, onboard memory virtualization)?

Hmm I was a bit surprised the X300 had so much on board memory to begin with but I wasn't aware a complex (?) software implementation was needed for a ram-less PCI-Express board, unless this press release just demonstrates a company's new direction (for the low end).

So, if I understood correctly, I think this is a great move. A ram-less X300 (X200 anyone) to compete with Intel's XG3 and even a 1/2 ram X600 (X500?) for the AIBs would be a great idea. Most importantly, it would require little or no additional development from ATi (aside from the driver).

I do agree with most here that for high end parts, this is almost useless. It could be nice to try out some higher quality mode for which our card doesn't have enough onboard ram but I wouldn't expect performance to be any better than with "regular" PCI-Express memory trashing.

 

[suryad]

As for the Dell comment that they are not producing high end machines that post very high benchie numbers right out of the box I tend to disagree with that somewhat since I am using a Dell Lappie that can play Doom 3 just fine.

Also I think this Hypermemory thing is sorta useful if you have a really high end machine and a graphics card with only 256 mb of memory and are trying to play Doom 3 at the Ultra setting. Then to offset the performance hit of not having 512 mb or greater of vdo memory will probably be where the Hypermemory technology will shine.

As for Tim Sweeney's comment about 1 GB vdo cards...I say where the card at? I got the $$ ready to go!!! Personally I think hardware is going to be advancing at quiet an incredible pace for at least the next 3-4 years to say the least. Seems to me that hardware developes are starting to think outside the box like for example with this Hypermemory thing. I think the logic is sound but for SURE as most of you have said, using this technique on AGP is a bad idea.

 

[trinibwoy]

... the average pc user who is responsible for the vast majority of the profits in the industry.

No they are not. They dont upgrade and thats why the"industry" is in a long long slump.

You contradict yourself. First you say that they are not responsible. Yet you say that their upgrade habits are responsible for the slump? Which is it?

Last time i chked they just started to go to highend systems. They will have uber rigs in 2005.

Yep, but will 1.5GB of PC4000 be the widely accepted definition of high end in 2005? Or will it still be enthusiast overkill?

Just what would qualify for high end (high end being enthusiast)? Doom3 on ultra runs best on 512mb cards. For the highest IQ you will need 512mbs starting now.

As far as I know none of the big titles next year require 512MB of video ram. Only MMORPG's with huge amounts of content may be the exception to that rule. Doom3 on ultra has no discernable difference to Doom3 on high from nearly all accounts so it's not really a good example of near future expectations for vram capacity.

 

[Schuey]

Perhaps I'm missing something here, but isn't the memory bandwidth critically important to high rez gaming and for enabling features like FSAA. I fail to see how ram that is running ten times slower than that on my video card is going to be of any use to the VPU on my X800.
I guess this would eliminate all those stupid 256 meg low-end cards, but it won't do much for high-end gaming.

 

[MfA]

I dont think virtualized graphics memory was a 3DLab's first BTW. It was done by DEC before them if my memory serves me correct.

 

[Sxotty]

So did they actually do anything, or are they just releasing PR that is meaningless. Well I mean do they have a patent or does this just mean that people will have a newer worse graphics card?

 

[DeanoC]

Perhaps I'm missing something here, but isn't the memory bandwidth critically important to high rez gaming and for enabling features like FSAA. I fail to see how ram that is running ten times slower than that on my video card is going to be of any use to the VPU on my X800.
I guess this would eliminate all those stupid 256 meg low-end cards, but it won't do much for high-end gaming.

Because the majority of GPU RAM is wasted holding things that aren't visible, like the top mip level of a texture thats a mile away. By having a slower but still quite fast pool, you can start treating the main GPU RAM as a cache. Even if the cache prediction algorithms go wrong, the worst is slightly lower fill-rate for a frame.

Combine with true virtual paging ala 3DLabs and Gamecube Flipper and your graphics visuals will improve

Also if you need smaller pools on the actual card you could start fitting really special RAM to highend boards. With things like free MSAA and free alpha blending...

 

[Pete]

So, if I understood correctly, I think this is a great move. A ram-less X300 (X200 anyone) to compete with Intel's XG3 and even a 1/2 ram X600 (X500?) for the AIBs would be a great idea. Most importantly, it would require little or no additional development from ATi (aside from the driver).

This was my first thought, too. This is the first real-world execution of the "GPU socket" that Rendition/Micron were (considering) aiming for. I'm curious how much the PCIe latency will affect scores, but PC3200 should be good enough for 8x6 and maybe 10x7 gaming, no?

 

[AndrewM]

I think a lot of people here have missed the point, or dont quite understand what virtual memory will enable. Deano explained it quite well.

This PR looks like "hey, we have VM but you wont understand it, so we make up silly names for it".

 

[cho]

16MB on-card and xxxMB "on-board"

 

[Bjorn]

By having a slower but still quite fast pool, you can start treating the main GPU RAM as a cache. Even if the cache prediction algorithms go wrong, the worst is slightly lower fill-rate for a frame.

Can you really consider the PCI Express/System RAM to be "quite fast" ?

It's of course shared with the CPU also. And the P4 f.e seems to be rather bandwidth hungry.

 

[tcchiu]

Supposed futuremark score of the core?

Intel i915 is abot 1500 (3dmark03), if the core didn't outperform IGP too much, i can hardly see the market.

 

[Ailuros]

Also, this seems to be in preparation for WGF (or whatever they're calling it these days).

Bingo

 

[Lezmaka]

I think a lot of people here have missed the point, or dont quite understand what virtual memory will enable. Deano explained it quite well.

Sounds even more likely. The Inq has an "article" about nvidia having an "answer" to ATI's hypermemory. And unless something's changed, I'm pretty sure drivers for Longhorn will require virtualization of video memory..

 

[pc999]

Perhaps I'm missing something here, but isn't the memory bandwidth critically important to high rez gaming and for enabling features like FSAA. I fail to see how ram that is running ten times slower than that on my video card is going to be of any use to the VPU on my X800.
I guess this would eliminate all those stupid 256 meg low-end cards, but it won't do much for high-end gaming.

Because the majority of GPU RAM is wasted holding things that aren't visible, like the top mip level of a texture thats a mile away. By having a slower but still quite fast pool, you can start treating the main GPU RAM as a cache. Even if the cache prediction algorithms go wrong, the worst is slightly lower fill-rate for a frame.

Combine with true virtual paging ala 3DLabs and Gamecube Flipper and your graphics visuals will improve

Also if you need smaller pools on the actual card you could start fitting really special RAM to highend boards. With things like free MSAA and free alpha blending...

GameCube already have that! Why still the games low rez in textures? So virtual memory will be a completely reality in next gen (from ATI at least) ?

 

[AlNom]

GameCube already have that! Why still the games low rez in textures? So virtual memory will be a completely reality in next gen (from ATI at least) ?

I'd say it's doing pretty well with it's 24MB vid memory and it's 3MB 1T-SRAM versus PS2's 4MB eDram and 32mb Main mem.

 

[pc999]

Yes, you may be right

 

[jolle]

HyperMemory uses intelligent memory allocation algorithms to optimize the use of available local memory and ensure critical components are placed in fast local memory when required.

sounds alittle bit like the virtual memory thing 3dlabs does

Think it does too..
Which I understand both Carmack and Sweeny has been pushing for on gaming cards, as it does release developers from the iron set limits of the onboard RAM on video cards..
AGP texturing isnt a viable solution to rely on if you cram more textures into a game then can fit on the board..

If its anyhing like 3Dlabs VM then the onboard ram will be like a Cache mem, and the system RAM storage, when needed.
With some good management it could work fine, 4Gb/sec bidirectional means it can feed about 68Mb worth of textures per Frame @ 60FPS doesnt it? from System RAM to GPU, in a ideal situation.
Just aslong as the most critical bits remains onboard where the bandwidth is tons higher..

This isnt ONLY meant for the lowend is it?
INQ rumoured something about Nvidia announcing a similar tech within soon, if its WFG stuff it just might..
Spose its a good thing in Longhorn what with the GPU use and that, when Onboard RAM runs out..

 

[Rolf N]

I find it rather frivolous to tack a crappy marketing name and a press release on something that's well understood and in use for decades as "virtual memory".

Tons of Kudos to ATI for implementing something a 286 can do