Next-gen tech? Rambus targets 1TB/s memory for 2010

[phat]

Regarding the 50GB/s for PS3, that's the figure for the dual-channel XDR system that the PS3 was anticipated to have before Sony announced Nvidia as their GPU provider. Because Nvidia doesn't talk XDR, they dropped one XDR bank for the GDDR that the GPU could interface with.

Rambus systems have typically been dual-channel or more. (That was the case with PCs and PS2. Not sure about N64.)

 

[Shifty Geezer]

For 1080p+4xMSAA at 16bit HDR,
16bit HDR (64bit) + Z (32bit) = 12 bytes
1920 * 1080 * 12 * 4 = 99532800
it requires 94MB eDRAM.

Doesn't sound good on paper there! Whereas 94MB of FB in 2+GB (we hope!) RAM with 500GB/s RAM isn't much of a RAM cost and the BW would be there to support it. If you don't have a full TB of BW, more like 500 GB/s, overdraw and the like would have a significant impact on BW shared with big and luscious textures etc.

Again, I don't think we'll have a non-BW restricted console next-gen!

 

[Fafalada]

they can integrate a 8 MB large scratchpad buffer which can be used for tiling or simply for local storage for GPGPU for next to no area cost, even if they use SRAM!

As long as you're talking about "give potential to devs" - I'd rather change it to DRam, make it bigger, and above all, general purpose instead of some crippling specialization like in most other machines to date.

The issue really isn't with tiling or whatever other extra work is involved in using it effectively, it's in the limited range of applications that give any sort of tangible advantage. Manually controlled specialized memory banks suck.

And while on subject on tiling - someone should be writting papers and making SDKs for efficient spatial scene partitioning for rendering layers(as opposed to screen space partitioning) like, right now. There's plenty of local genAoses around here, get cracking, the future awaits.

 

[V3]

And while on subject on tiling - someone should be writting papers and making SDKs for efficient spatial scene partitioning for rendering layers(as opposed to screen space partitioning) like, right now. There's plenty of local genAoses around here, get cracking, the future awaits.

Efficient in what way ? Equal distribution of workload or you have something else on your mind ?

 

[MfA]

And while on subject on tiling - someone should be writting papers and making SDKs for efficient spatial scene partitioning for rendering layers(as opposed to screen space partitioning) like, right now. There's plenty of local genAoses around here, get cracking, the future awaits.

What would be the point exactly? Even with sort last that's not how you would want to distribute work across multiple GPUs (it's worst case as far as occlusion and locality of reference goes ... which are already poor enough in sort last without extra help).

PS. Fafalada, I don't care about your pain ... I'm sorry. Manually controlled specialized memory banks may very well suck, but their cost is low ... and their potential for performance increase is big enough for having it only being useful with cripplingly specialization is still an overall win. Hell, NVIDIA didn't put that local memory in the G80 shaders for shits and giggles (although I'm sure some CUDA and their compiler developers curse them for it).

 

[Megadrive1988]

I noticed in their slide they put 5 GB/sec for PS2.

it's 3.2 GB/sec. but then I realized (duh) they were going in increments of 10x.


I was also surprised by a comment here that the 50 GB/sec for PS3 wasn't because they (roughly) combined the bandwidth of CPU to Rambus and GPU to GDDR3 but were showing that PS3 earlier in its development would've had 50 GB BW Rambus.

 

[Shifty Geezer]

When the RAM was announced, there was speculation whether Sony would go with lesser RAM (256MB XDR) and up the bandwidth as a result, based on the choice of DIMMS (right term?) that Sony went with. The choice was 25 GB/s or 50 GB/s, and Sony went the former. 50GB/s is the current performance of XDR though for anyone who'll use it that high.

 

[phat]

When the RAM was announced, there was speculation whether Sony would go with lesser RAM (256MB XDR) and up the bandwidth as a result, based on the choice of DIMMS (right term?) that Sony went with. The choice was 25 GB/s or 50 GB/s, and Sony went the former. 50GB/s is the current performance of XDR though for anyone who'll use it that high.

With point-to-point memory modules like XDR, the correlation is more memory, more bandwidth since to get 512 MB of XDR would have required two 256 MB XDR modules, each with its own 25 GB/s channel, for a system total of 50 GB/s.

Or something like that. It's been a while.

 

[Quaz51]

the slide is an "equivalence" for a system with one bus unified memory
the N64 have one bus unified memory and it's 500MB/s
if the PS2 had one bus unified memory with no eDRAM it would be probably ~5GB/s
if the PS3 had one bus unified memory it would be probably ~50GB/s
...

 

[Fafalada]

What would be the point exactly? Even with sort last that's not how you would want to distribute work across multiple GPUs (it's worst case as far as occlusion and locality of reference goes ...

You can composite the scene first and do the pixel pass deferred. And accelerating attribute passes this way would have little issues with locality I imagine.
And in some future when we move beyond 2d displays, most of the occlusion tech will inevitably fly out the window anyway

Manually controlled specialized memory banks may very well suck, but their cost is low ... and their potential for performance increase is big enough for having it only being useful with cripplingly specialization is still an overall win.

I do understand the reasoning, but I'm not any more fond of it.
Mind you I'm not against manual controlled banks as such (I like Cell after all), but yea, specialization doesn't tickle my fancy.

 

[MfA]

You can composite the scene first and do the pixel pass deferred. And accelerating attribute passes this way would have little issues with locality I imagine.

There still has to be an advantage ... you'd be plain better off not sorting at all and simply distributing objects round robin style.

If you have memory to burn for sort last and have your hard set on also doing sort first you could do back to front and approximate tile sorting first (because the tile sorting is only approximate you still need multiple full framebuffers and a final compositing step).

Either way, better than having depth sorted layers rendered before compositing them.

 

[TheChefO]

Is this memory tech still on track to come out in 2010?

 

[Aeoniss]

I wonder if Sony will go with XDR again.. I mean it would appear to be the logical choice. Their partnership has definitely produced interesting results.

But I'm sure Yellowstone\XDR memory variants are rather expensive..

 

[Megadrive1988]

I would be extremely pleased if PS4 used 1 TB/sec Rambus memory.

The PS3 used the low-end of what Rambus Yellowstone/XDR was capable of (26 GB/sec) since first-gen XDR1 was meant to be able to reach 104 GB/sec bandwidth using more channels (correct me if I'm mistaken). I had hoped PS3 would at least used 52 GB/sec, but that was not to be.

BTW XDR2 is capable of 200+ GB/sec.

Going from 26 GB/sec to 1024 GB or 1 TB/sec would be much greater than the leap (nearly 40x) from PS2's 3.2 GB/sec Rambus RDRAM to PS3's 26/GB/sec, which was only an 8x leap.

With 1 TB/sec bandwidth, there would be no need for expensive, space-wasting EDRAM. The PS3 is pretty bandwidth starved IMO but 1 TB/sec should be enough for next-gen CELL CPU and a highend DX11 or DX12 Nvidia GPU (or Larrabee).

 

[Aeoniss]

How is it bandwidth starved? 25GB\s for Cell and 24GB\s(?) for RSX is on par with the competing console and good given the timeframe and dates they were working on tech.

 

[EasyRaider]

The TB/s figure relies on an extremely wide memory interface. I can't see more than 256 GB/s (or thereabouts) external bandwith being realistic for a console. And that would (probably) be in total, for CPU and graphics.

 

[Jugix]

Assuming PS4 is 'only' a modest(*) upgrade to PS3 hardware, what would insane memory bandwidth bring us in terms of visuals?

What barriers would it lift away?

(*) modest = Low-midend GPU in 2011 and Cell with 2x the performance.

 

[Panajev2001a]

Assuming PS4 is 'only' a modest(*) upgrade to PS3 hardware, what would insane memory bandwidth bring us in terms of visuals?

What barriers would it lift away?

(*) modest = Low-midend GPU in 2011 and Cell with 2x the performance.

It could allow them to use a UMA configuration for a CELL CPU upgrade (not revolutionary, not a huge leap) + customized DX11 nVIDIA GPU set-up (which might simplify programmers' life a bit).

 

[TheChefO]

It could allow them to use a UMA configuration for a CELL CPU upgrade (not revolutionary, not a huge leap) + customized DX11 nVIDIA GPU set-up (which might simplify programmers' life a bit).

Question for you game programmers out there:

Would it be more beneficial to have significantly faster ram, or significantly more ram to work with?

 

[T.B.]

Would it be more beneficial to have significantly faster ram, or significantly more ram to work with?

It's a balancing act. More RAM is only going to help you if you can actually use it, otherwise it will just reduce the load on mass storage a bit. So you need enough speed to be able to run through a sizable chunk of it every frame. On the other hand, if you have tons of bandwidth but little space, there are only so many times you want to run over the same data.

Think textures: More storage means bigger/more textures, but unless you have the bandwidth, you cannot use them without stalling your texture units.

Basically it's a triangle of memory bandwidth, memory size and processing power. Increasing one without the others only makes sense to a degree.