Predict: The Next Generation Console Tech

[Proelite]

Questions for the wise older timers.

Would it makes sense to go for a small SOC, like around 100-200 mm2, but have 2-4 of them in your console?

That's one way to go, or to not go around SOC performance constraints .

 

[Ninjaprime]

Questions for the wise older timers.

Would it makes sense to go for a small SOC, like around 100-200 mm2, but have 2-4 of them in your console?

That's one way to go, or to not go around SOC performance constraints .

No. The purpose of using a SoC at all is to reduce the number of chips you need to use/make. Copying a SoC just means you have more chips and redundant hardware. Its lose/lose. It does make sense to have a non-SoC setup if your SoC is going to be large though, as 2 chips that are 200mm^2 are cheaper than one big SoC thats 400mm^2.

 

[Proelite]

No. The purpose of using a SoC at all is to reduce the number of chips you need to use/make. Copying a SoC just means you have more chips and redundant hardware. Its lose/lose. It does make sense to have a non-SoC setup if your SoC is going to be large though, as 2 chips that are 200mm^2 are cheaper than one big SoC thats 400mm^2.

Long term cost reduction, and power efficiency seems to be the advantages that MS would pay for up front. I doubt that a big SOC would be anywhere as expensive as Xenon + Xenos in 2005.

 

[Ninjaprime]

Long term cost reduction, and power efficiency seems to be the advantages that MS would pay for up front. I doubt that a big SOC would be anywhere as expensive as Xenon + Xenos in 2005.

If you're talking a 400mm^2 die, it would most certainly cost more than Xenon and Xenos combined.

 

[Proelite]

If you're talking a 400mm^2 die, it would most certainly cost more than Xenon and Xenos combined.

That's interesting.

Assuming the die size is 300mm^2 instead, how would you divide the die space to different components of the soc.

In addition, in order to boost performance, can the chips on the SOC run at higher clock compared to bigger discrete versions. I.E. 4.0+ ghz for CPU, ~1.0ghz for the GPU, etc.

 

[3dilettante]

Going with an SOC means sacrificing speed for the various components. The CPU and GPU are going to have to live under the same cooler and fight over the same footprint for power and ground.
Aside from possible efficiencies by eliminating a hop off-chip for GPGPU, the SOC would be more about saving money. Until code exists that really needs fast communication between CPU and GPU, both sides would see a decrease.

In the case of the Xbox 360, the GPU and CPU were not merged into one design for several nodes, by which time their current clocks could comfortably fit within the standard TDP for a single chip.

If the design is an SOC at the outset, one factor may be that additional resources were invested in designing the system bus unit that emulated the external link between the GPU and CPU for the 360. That's engineering effort and die area Microsoft would have liked not spending money on.

 

[Proelite]

Going with an SOC means sacrificing speed for the various components. The CPU and GPU are going to have to live under the same cooler and fight over the same footprint for power and ground.
Aside from possible efficiencies by eliminating a hop off-chip for GPGPU, the SOC would be more about saving money. Until code exists that really needs fast communication between CPU and GPU, both sides would see a decrease.

In the case of the Xbox 360, the GPU and CPU were not merged into one design for several nodes, by which time their current clocks could comfortably fit within the standard TDP for a single chip.

If the design is an SOC at the outset, one factor may be that additional resources were invested in designing the system bus unit that emulated the external link between the GPU and CPU for the 360. That's engineering effort and die area Microsoft would have liked not spending money on.

To me that sounds like the business people at MS won back a year or more ago when the decision for SOC was made, if such a decision was made. One has to remember that the Wii was still doing pretty well then, and MS was extremely onboard with the whole casual idea. We'll have to wait until next year to see concrete information about the Loop, but so far it sounds pretty disappointing from a hardcore gamer's point of view. Maybe Epic will come to the rescue again? MSG is extremely fickle and shortsighted when it comes to their overall direction.

 

[AlNom]

I must have missed the memo confirming SoC... huh

 

[3dilettante]

A decision such as whether or not to have an SOC or separate components is a pretty fundamental one that wouldn't be easy to change once past the early design specification stage.

Epic would have no real incentive to ask for a change.
An SOC doesn't really look all that different to the code, and there's no hard value for not having X extra MHz from total nobody really knows.

 

[Proelite]

I must have missed the memo confirming SoC... huh

Just some rumors from a site called SemiAccurate.
I've heard they've been pretty reliable in the past, but I won't put money on them.

 

[liolio]

Just some rumors from a site called SemiAccurate.
I've heard they've been pretty reliable in the past, but I won't put money on them.

I would not call it reliable but out of the rumours sites I believe they are the more reliable.
I would like to hear something from the ARstechnica mole though

 

[Rangers]

I'm pretty sure we're going to get a good leap. Epic seems to be MS's closest developer ally and they stress in every recent interview that they feel next gen needs to be a large visual jump to rope people in. So I can almost think they speak as a proxy for MS.

For example here's a timely tease from Mike Capps

Do you want to engage with the console manufacturers before their new systems come out? Do you want to influence them on hardware specifications?

That’s absolutely our plan. I can’t say much more than that. Okay, let’s say, a year ago that was our plan, and I can’t tell you whether we’ve done it or not yet.

What do you want from the next generation of systems?

I think it’s very important that a gamer sees an Xbox Next or PlayStation Next and can clearly see the tech is not possible on current consoles.

http://www.develop-online.net/features/1462/Epic-Games-next-gen-manifesto

 

[Ninjaprime]

To kind of expand on what I said earlier, here is a overly simplified explaination of the problem with a large die: Draw a circle. Put 20 random dots on it. Those dots are defects on your wafer. Lets say at 200mm^2 you can put 100 dies on a wafer, so at 400mm^2 you can put 50 on the same wafer. The wafer price is constant. The defects are constant.

Now, those 20 defects are on 20 dies on your 200mm^2 part, and you get 80 dies and 80% yields. On the 400mm^2 part, you get the same 20 defects, but unless you are really lucky and two defects are on the same die, you get 20 bad dies and only 30 good ones, so you get 60% yields. If you are paying say, $5,000 per wafer, your 200mm^2 chips are $62.50 each, and your 400mm^2 chips are $166.67 each, or $41.67 more than if you had 2 dies that were 200mm^2.

Now take into consideration that a newer process is going to have a higher defect rate, you might only get 40% yields with a big chip, and the gap just gets larger.

 

[Ninjaprime]

That's interesting.

Assuming the die size is 300mm^2 instead, how would you divide the die space to different components of the soc.

In addition, in order to boost performance, can the chips on the SOC run at higher clock compared to bigger discrete versions. I.E. 4.0+ ghz for CPU, ~1.0ghz for the GPU, etc.

If MS has a SoC as the rumors say, I would guess that its much more GPU than CPU this time around. I would take an offhand guess of a 20/75/5 die share or so. 28nm probably gives you 3.5-4 billion transistors at 300mm^2, so maybe a custom Power7 derivative quadcore @ ~500 million transistors, or if they want to go exotic and heavy multicore again, a BlueGene/Q/Wirespeed derivative with 8 cores. A GCN based AMD GPU with ~3 billion transistors would probably be about right with 32 CUs(2048shaders)/128 TMU/32 ROPS or so. It would have to be underclocked relative to desktop parts though. The 5% would be the chipset functions and what not.

 

[Proelite]

I'm pretty sure we're going to get a good leap. Epic seems to be MS's closest developer ally and they stress in every recent interview that they feel next gen needs to be a large visual jump to rope people in. So I can almost think they speak as a proxy for MS.

For example here's a timely tease from Mike Capps
http://www.develop-online.net/features/1462/Epic-Games-next-gen-manifesto

The first thing that Mike Capps said doesn't seem to be too optimistic too me. Seems to me like he is leaving an escape clause for himself in the future. "well we did try our best"

If MS has a SoC as the rumors say, I would guess that its much more GPU than CPU this time around. I would take an offhand guess of a 20/75/5 die share or so. 28nm probably gives you 3.5-4 billion transistors at 300mm^2, so maybe a custom Power7 derivative quadcore @ ~500 million transistors, or if they want to go exotic and heavy multicore again, a BlueGene/Q/Wirespeed derivative with 8 cores. A GCN based AMD GPU with ~3 billion transistors would probably be about right with 32 CUs(2048shaders)/128 TMU/32 ROPS or so. It would have to be underclocked relative to desktop parts though. The 5% would be the chipset functions and what not.

Does the GPU transistors include Edram?

 

[Rangers]

The first thing that Mike Capps said doesn't seem to be too optimistic too me. Seems to me like he is leaving an escape clause for himself in the future. "well we did try our best"

Seems to me more like a hint "we were successful getting our desires adopted". Maybe it means more RAM like last time 4GB --->8GB

 

[Acert93]

Seems to me more like a hint "we were successful getting our desires adopted". Maybe it means more RAM like last time 4GB --->8GB

Maybe if we keep setting the bar at 8 we will get at least 4GB. If we end up with 2GB with a 2013/2014 console that stretches into 2022 (!!!) I will post a big

One basic arguement is this: Sure, load times suck. It doesn't seem you are going to get away from an initial load time. But more memory means that if you start in a small gamespace and stream in you could start quickly. But lets assume you have a 2GB active memory for the game level. You can be actively loading the next section so when you get there there is no load time (or at least minimal). So while some people will note how optical/HDD are slow and 4GB will take forever to write to ("too much memory to fill") a large memory pool is one part of the puzzle to solving load times. Not to mention moving up to 1080p with 4xMSAA and multiple buffers means the framebuffer is going to increase a nice bit over these sub-HD 1xAA games. Ditto better diffuse textures, normal maps, etc. I would image the OS is going to grow a LOT as a lot more tools, including ingame video streaming, are going to be added to the API.

I guess, optimistically, memory has grown at about 8x every gen. The N64 had 4MB, the PS2 had 32MB and the Xbox 64MB, the 360 and PS3 512MB. Unless someone makes a compelling move to less/much, much faster memory or invest in a scratchpad or something I don't see why a lot of cheap DRAM would be avoided. Heck, if MS launches with 2GB and Sony with 4GB that right there is going to make a different in development ease. An entire generation of developers whining about memory constraints. And it would be a bigger hit than "just" 2x Memory because on the 2GB machine you need to (-) the OS and (-) the framebuffer. Developers will end up with well over 100% more memory.

So while Al keeps trying to sell me on 2GB I am not having any of it So 8GB it is! Of XDR4 on a toasted bagel in a flying saucer

 

[Rangers]

This is just going to show my tech idiocy here, but what about going with a PC-like setup of system RAM and then VRAM?

DDR3 is so cheap it's RIDICULOUS. $35 for 8GB! $70 for 16GB!

I never really even thought of it before, but in PC's do they load the textures into main RAM, then stream them to the VRAM when needed? So would this allow for levels basically streaming the assets in frequently without pop in? I mean I think DDR3 has ~36GB/s of bandwidth. That's massively faster than any HDD.

So what about a design of 8GB of DDR3+2GB of VRAM? Possible? Viable? Or would such a design underperform compared to, say, a 4GB of unified GDDR5 architecture?

What if the cost tradeoffs are, 16GB of DDR3+4GB VRAM vs 8GB of unified GDDR5? Which would perform better? Where would EDRAM fit in all this?

 

[AlphaWolf]

You can't really look at current GDDR5 costs and compare them to what they would be if MS or Sony start buying by the truckload.

I think there are more cost effective ways to improve streaming than adding buckets of DDR3.

Also, It was my impression that developers generally prefer a unified memory pool.

 

[wco81]

"I think it’s very important that a gamer sees an Xbox Next or PlayStation Next and can clearly see the tech is not possible on current consoles."

How about tech which isn't possible on current PCs?