Lots 'o Embedded DRAM on R500

[Geo]

. . .sez these folks here:

http://www.teamxbox.com/news.php?id=5388

The Graphic Chip

The graphic chip will be based on the the R500. This VPU has been in design at ATI’s Marlborough, Mass. office. It'll be fully compatible with DirectX 9's PS and VS 3.0 and the next version of DirectX: DX10, the same suite of APIs that will be used in Longhorn.

What nobody is telling you and you'll know about this first, here on TeamXbox, is the revolutionary approach of the Xbox 2 to deal with today's biggest problem in graphics chips: memory bandwidth.

The graphic chip will contain not only a graphics rendering core but up embedded DRAM acting as a frame buffer that is big enough to handle an image that is 480i and can be 4 times over sampled and double buffered. Yeah, we all remember Bitboys but this time you can bet this is for real. This solution will finally make possible HDTV visuals with full screen Anti-Aliasing on.

The technology also supports up to 512 MB of external memory on a 256-bit bus. However, current specs plan to use 256 MB RAM, big enough for next-generation visuals which are all about computational power rather than large storage.

 

[Razor04]

The graphic chip will contain not only a graphics rendering core but up embedded DRAM acting as a frame buffer that is big enough to handle an image that is 480i and can be 4 times over sampled and double buffered. Yeah, we all remember Bitboys but this time you can bet this is for real. This solution will finally make possible HDTV visuals with full screen Anti-Aliasing on.

I am confused about this part...is it saying that it is limited to 480i for AA? If so then that definately is not HDTV...that is plain old SDTV. Nearest HDTV resolution would be 1280x720.

 

[mboeller]

Yeah; for sure

>>HDTV-Support with an eDRAM-Framebuffer able to hold an 480i Frame.<<

[lots of rolling eyes]


for me HDTV starts at 1280x720p nothing less; so the R500 would need at least 2(1xFB+ZB) x 4(for FSAA) x 1280x720x32bit = 28,125 MB or 32MB. Only then the eDRAM-Framebuffer would work. If the Xbox2 would support 1920x1080p then the requirement would be 64MB.

But this assumes that the Framebuffer is always only 32bit, which IMHO will not be true for DX-next.

Maybe an better way would be to have eDRAM for the Z-buffer only.

 

[akira888]

mboeller:
I had that idea myself for a new chip (not that I'm a HW guy myself of course). Put a per-pixel (at the lowest level) hierarchical Z buffer onto eDRAM and then do a depth pass first into this buffer, then render normally in the next frame, perserving the driver calls as AFR does. The cost is twice the transform work and an additional frame of latency, but the benefit would be nearly zero overdraw.

The "teamxbox" report is obviously bunk, the 480i (???) remark gives that away very easily.

 

[London Geezer]

The unbelievability of the article has been discussed, and is still being discussed on the console forum

 

[Wunderchu]

Nearest HDTV resolution would be 1280x720.

AFAIK the resolution of 1080i is 960x540 ...

 

[DemoCoder]

it's 540 lines per field, but the actual vertical resolution is still 1080 scanlines.

 

[OpenGL guy]

it's 540 lines per field, but the actual vertical resolution is still 1080 scanlines.

And isn't the width 1920 pixels?

 

[cybamerc]

Nearest HDTV resolution would be 1280x720.

AFAIK the resolution of 1080i is 960x540 ...

The horisontal resolution is 1920 pixels for widescreen and 1440 for 4:3.

And while you can get by with a half height framebuffer it's hardly ideal.

 

[Wunderchu]

The horisontal resolution is 1920 pixels for widescreen and 1440 for 4:3.

And while you can get by with a half height framebuffer it's hardly ideal.

1920x1080 is 1080p resolution, AFAIK ..

 

[cybamerc]

Wunderchu:

> 1920x1080 is 1080p resolution, AFAIK ..

1080p is 1920x1080 in the widescreen format and 1440x1080 in the 4:3 format.

1080i has the same resolution but a frame is simply created by first updating the odd lines of the display and then the even lines.

 

[RussSchultz]

The interlace versions update every other line, twice as fast as the progressive versions. (Two fields per frame, but same frame rate)

http://www.nettable.com/HDTVFAQ.HTM

 

[AzBat]

Wouldn't 16mb of eDRAM be plenty for 480p in 4:3 and 16:9 aspects? I see that being more of a possibility especially if they decide to do backward compatibility. You would get free 4x FSAA with all your older games and those with standard 4:3 TVs would still get a nice enhancement to all games(old and new). I have a feeling Microsoft will be cost conscious on the eDRAM and won't care if Sony uses 32mb of eDRAM. However, if Sony surprises us and goes for more, then I suspect Microsoft will have no choice but to upgrade its specs as well.

Tommy McClain

 

[nobie]

Remember, the R500 and whatever chip is in X-Box 2 wont be exactly the same. It doesn't mean R500 will use embedded ram.

 

[Wunderchu]

ahh ... I think I now understand what you ppl are talking about WRT 1080i
.. with 1080i you still get 1080 discrete lines of resolution, displayed in an interlaced fashion...




WRT the eDRAM topic ... here is an excerpt from the ATI Hardware IRC chat held at Rage3D on Dec. 3 , 2003 :

<Ratchet> Will ATI start to use embedded / on-chip memory anytime soon?
<ATI_Toronto> We could tell you but then we would have to kill you. done.
&lt;SirEric> We do have on package memory products now. done.
<ATI_Toronto> last question...
&lt;SirEric> (note: our chips have LOTS of onchip memory already. done)

source: ( http://www.rage3d.com/board/showthread.php?s=3651af711870f15b4dc8d8388147dcb5&amp;threadid=33728806 )

(related threads at Rage3D: http://www.rage3d.com/board/showthread.php?s=3651af711870f15b4dc8d8388147dcb5&amp;threadid=33728821

http://www.rage3d.com/board/showthread.php?s=3651af711870f15b4dc8d8388147dcb5&amp;threadid=33728805

http://www.rage3d.com/board/showthread.php?s=3651af711870f15b4dc8d8388147dcb5&amp;threadid=33728886 )

 

[KimB]

They were probably talking about caches. Modern GPU's require lots of caches for optimal operation.

eDRAM will take up die space, and so will only become good to use in a PC part after the core has outstripped the external memory interface, despite the use of many different types of memory bandwidth optimizations.

Currently GPU's seem to be doing rather well without eDRAM, so it may be a few years yet before we see it in a PC part.

Remember that a console part, if it is to use eDRAM (or something similar, like the 1T-SRAM in the Gamecube), it doesn't need nearly as much as a PC part would, as the output resolution is much smaller.

eDRAM in a PC part would probably best be used in a low-end part, actually, where with proper design you might be able to get decent perormance out of, say, a 64-bit external memory interface.

 

[Tagrineth]

What nobody is telling you and you'll know about this first, here on TeamXbox, is the revolutionary approach of the Xbox 2 to deal with today's biggest problem in graphics chips: memory bandwidth.

The graphic chip will contain not only a graphics rendering core but up embedded DRAM acting as a frame buffer that is big enough to handle an image that is 480i and can be 4 times over sampled and double buffered. Yeah, we all remember Bitboys but this time you can bet this is for real. This solution will finally make possible HDTV visuals with full screen Anti-Aliasing on.

Revolutionary.

Hell of a revolution... I mean, WOW, PS2 has only been doing that exact same thing for what... four years, not counting pre-production tests?

 

[cybamerc]

Chalnoth:

> Remember that a console part, if it is to use eDRAM (or something
> similar, like the 1T-SRAM in the Gamecube), it doesn't need nearly as
> much as a PC part would, as the output resolution is much smaller.

The next gen consoles will support HDTV resolutions so they'll need as much framebuffer space as a PC.

 

[Gubbi]

I don't think the primary purpose of on die RAM in future GPUs is to boost bandwidth. Rather it is to reduce latency for dependent texture fetches.

Bandwidth is easy, just throw pins at the problem and scale frequency. Latency is limited by a nasty physical constant, c.

Given next gen. GPU bandwidth of 50GB. You can plow through a 1920 x 1080 x 4 bytes buffer 6000 times every second. Using 5 of these (Framebuffer+ 4 x Z for FSAA, discounting compression) you get 1200 iterations per second or 20 times every frame (@ 60 frames/second).

So what is needed is enough RAM to hold your texture working set on die. Note that this is not all data belonging to active (visible) textures only those texels that actually contribute to visible pixels needs to be cached.

I'm guessing something like the texture cache on GameCube's Flipper, only larger ~8-16 instead of 1MB.

Cheers
Gubbi

 

[Dio]

Bandwidth is easy, just throw pins at the problem and scale frequency. Latency is limited by a nasty physical constant, c.

Historically, the main reason for using embedded DRAM has always been bandwidth. Before R300 many people said a 256-bit memory bus was crazy, (frankly, it still does feel crazy to me) and I'm not quite sure if the world is ready for 2000-pin chips... and scaling memory frequency isn't particularly easy (memory frequencies are scaling significantly slower than core clock frequencies both in CPU's and VPU's).

I'm not sure c really comes into one more than the other.