Predict: The Next Generation Console Tech

[AlNom]

Folks, let's try not to get too worked up. There's always the "do not respond" option.

On a completely different track, it's a bit fun seeing Redwood (40nm) specs if you compare it to Xenos (90nm):

Die Size
Xenos 182mm^2 + 80mm^2
Redwood 104mm^2

Memory Bus
Xenos 128-bit GDDR3
Redwood 128-bit GDDR5 or DDR3

ALU's
Xenos 48
Redwood 80

Texture filtering (bilinear)
Xenos 16 tex/clock
Redwood 20 tex/clock

ROPs
Xenos 8 pix/clock,
Redwood 8 pix/clock,

Core Clocks
Xenos 500MHz
Redwood's 650-775MHz

Mem Clocks (Bandwidth)
Xenos 700MHz GDDR3 (22.4GB/s)
Redwood 900MHz DDR3 (28.8GB/s) or 4Gbps GDDR5 (64.0GB/s)



Even if you kept the core clocks the same, it's still packing over 1.66x the shader power, 25% more texture filtering capability, and DX11 compliancy... The 10MB of eDRAM do throw things a bit out of wack in terms of direct die space comparison, but all-considering, Xenos is about 2.5x the amount of silicon.

 

[aaronspink]

Even if you kept the core clocks the same, it's still packing over 8.33x the shader power, 25% more texture filtering capability, and DX11 compliancy... The 10MB of eDRAM do throw things a bit out of wack in terms of direct die space comparison, but all-considering, Xenos is about 2.5x the amount of silicon.

And that's the thing. Getting the performance increase for the next generation consoles is easy because the current generation really are fairly slow. For example, you can currently get a 1GB 128b 5570 for $65 from newegg! It is likely that the integrated CPU/GPU combo coming from AMD will be even faster and will likely sell in the $100 range. The performance jumps possible from current/coming PC hardware are leaps and bounds higher than what both the PS3 and 360 currently offer. Coming up with a BOM for a significant performance leap over the existing designs isn't going to be hard.

The hard part is having all the developer tools and infrastructure in place.

 

[ShadowRunner]

This i agree with entirely. I think its a given that we will see a big graphical leap this gen, it makes little sense not to as its an easy way to give people a reason to upgrade. There is no reason not to go for a significant performance upgrade. I cant believe we will see any wii style two consoles duct taped together aproach like has been suggested in the past, not unless there is some huge innovation that enables them to sell a non-competative(graphically) console in large numbers with a huge profit margin.

 

[Squilliam]

Folks, let's try not to get too worked up. There's always the "do not respond" option.

On a completely different track, it's a bit fun seeing Redwood (40nm) specs if you compare it to Xenos (90nm):

Even if you kept the core clocks the same, it's still packing over 8.33x the shader power, 25% more texture filtering capability, and DX11 compliancy... The 10MB of eDRAM do throw things a bit out of wack in terms of direct die space comparison, but all-considering, Xenos is about 2.5x the amount of silicon.

Isn't the Xenos GPU rated at 240Gflops whilst the Redwood GPU rated at 640 GFlops which would yield only about 4* the realizable shader performance at best case and 2* at worst case depending on relative efficiencies?

In terms of the number of transistors Redwood has aproximately 3* the number that Xenos has.

Which is exactly that SEGA did with its addons and then eventually integrated systems. It doesn't work that well in the context of consoles.

The addons were physical and expensive hardware, what im talking about is relatively minor in comparison. Its a change in the makeup of a silicon chip vs completely new mechanical hardware.

Which means you now have to test every console game on multiple hardware with multiple settings = increased cost for minimal benefit.

Multiplatform developers are already working with 3-4 completely different systems. It would be a minor change and it would only effect some of the code destined for the Xbox 360 SKU, they wouldn't have to retest everything for example and if theres a PC port it synergises with that testing as they are already working on a seperate DX11 path for many games.

If someone else is going to put out a new console it will be a generational change and the current 360 hardware will not be competitive. Look at it this way, the current 360 and PS3 are competitive on a hardware level with current generation bottom feeder CPUs and GPUs.

An iterative strategy is a lot less harmful for the eco-system than throwing everything out and starting afresh every 5-6 years like a typical console cycle. Since the current consoles are likely to hang around like a bad smell for the next five years, by the time they expire they will be so out of date they'll be competing with netbooks.

This way they can start to move towards supporting some of the newer technologies like 3D without throwing out the current userbase and they can have a console which will be able to play next generation games as future consoles will be so far ahead of current consoles in terms of performance they may as well be dropped.

 

[Blazkowicz]

Sure : counting 48 vs 400 is way too dramatic, if you consider Redwood is 80 ALU instead (Vec5) it's closer to reality (after that I don't know too much the differences in ALU, scheduling etc.)

 

[AlNom]

Oh right.. frickin PR nonsense. ugh. Makes more sense now. Fixed the post.

In terms of the number of transistors Redwood has aproximately 3* the number that Xenos has.

hm... 627M for Redwood. Wavey speculated on Xenos + ROPs (or rather subtract eDRAM) being 232M+20M. So roughly 2.5x the number. hm....


Humph, so just after 4 years, we've got a part that could have been a shrink after 2.5 nodes (excluding eDRAM of course). (also dx11 features not-withstanding)

 

[liolio]

Wouldn't they use GF28nm bulk? They don't need SOI and they certainly don't need the added expense. Also their Xenon CPU doesn't use/need SOI either.

I didn't think that much about it was mostly a try to "relax" the discussion that was going
If I take tit seriously may be non SOI process may not be a good choice to run three PPC cores @3.2GHz. I clearly don't know.

If they did indeed go for such a beast technically then would they also include OOOE for Xenon if it remains a part of the future chip to net more effective performance for code which doesn't suit the VLIW model?

When I put this together I include a three cores PPC for the sake of backward compatibility. Put fixing Xenon dark spots (LHS, etc.) and adding some cheap form of OoO (say as in Cortex A9 which is also a 2 issue processor) could be a nice bonus overall.

Lastly their technology is dictated by both their camera technology and the supporting CPU/GPU technology. If they have to wait for 4-5 years for the camera technology to catch up to where they want it to be to release an effective Natal replacement then won't they have to wait until then to release their console which fully leverages this technology? I don't see them wanting to release Natal 2 any time except on release of next generation hardware.

Makes sense but a failure of Natal and the success of competing motion "thingy" may accelerate their decision to release altogether Natal2 and their next gen system.

So if they are moving that direction why not slap Redwood or Juniper onto the side of the Xenon processor and start moving towards that technology today so they aren't caught with their pants down technologically by more advanced consoles in 2012/2013. There's a quirky little timespan between 2012-2014 where they need to be prepared to respond to competitor actions and not updating hardware leaves them vulnerable, especially when considering 3D.

It would also make sense but my prediction was more of a excessive geekiness show out than anything. I don't expect them to be able to design such a chip by themselves, AMD/ATI would be their best bet as they could reuse AMD/ATI existing VLIW architecture/ISA/Compiler/etc. Still even if they get the chip up having the matching software working reasonnably on time would be close to impossible if Intel difficulties with Larrabee are any clue. Honestly the odds are really low and as you say a "fusion like" chip including an custom ATI GPU makes more sense.

Still I would like to see such a chip a bit of a Cell2 under disguise.

EDIT

By the way Aaron I've two questions for you
1) What a "reasonable" system would look like for you?
2) What a "risky" system would look like for you? Say you were given some freedom where would you head?

 

[Squilliam]

Oh right.. frickin PR nonsense. ugh. Makes more sense now. Fixed the post.



hm... 627M for Redwood. Wavey speculated on Xenos + ROPs (or rather subtract eDRAM) being 232M+20M. So roughly 2.5x the number. hm....


Humph, so just after 4 years, we've got a part that could have been a shrink after 2.5 nodes (excluding eDRAM of course). (also dx11 features not-withstanding)

The flop count is 2.5* the difference btw at least when comparing rated -> rated. Thats about in line with the transistor count especially when considering the Redwood GPU is carrying a lot of fixed function transistors for functionality outside of rendering.

[

 

[MfA]

AMD/ATI would be their best bet as they could reuse AMD/ATI existing VLIW architecture/ISA/Compiler/etc. Still even if they get the chip up having the matching software working reasonnably on time would be close to impossible if Intel difficulties with Larrabee are any clue.

Actually it would be trivial, if they simply bolt on a instruction/data cache/decoder on each VLIW and leave the basic SIMD structure in place ... in the short term they could simply run the old software which runs the same program on each SC. Instant DX11 compatible graphics pipeline.

Obviously not an optimal solution, but you'd be up and running instantly with a pretty good solution.

 

[Mobius1aic]

Folks, let's try not to get too worked up. There's always the "do not respond" option.

On a completely different track, it's a bit fun seeing Redwood (40nm) specs if you compare it to Xenos (90nm):

Die Size
Xenos 182mm^2 + 80mm^2
Redwood 104mm^2

Memory Bus
Xenos 128-bit GDDR3
Redwood 128-bit GDDR5 or DDR3

ALU's
Xenos 48
Redwood 80

Texture filtering (bilinear)
Xenos 16 tex/clock
Redwood 20 tex/clock

ROPs
Xenos 8 pix/clock,
Redwood 8 pix/clock,

Core Clocks
Xenos 500MHz
Redwood's 650-775MHz

Mem Clocks (Bandwidth)
Xenos 700MHz GDDR3 (22.4GB/s)
Redwood 900MHz DDR3 (28.8GB/s) or 4Gbps GDDR5 (64.0GB/s)



Even if you kept the core clocks the same, it's still packing over 1.66x the shader power, 25% more texture filtering capability, and DX11 compliancy... The 10MB of eDRAM do throw things a bit out of wack in terms of direct die space comparison, but all-considering, Xenos is about 2.5x the amount of silicon.

I have a Redwood based 5570 w/ 1 GB DDR3 in my low profile system. Sure it is a little ass kicker, though a bit pricey. Stands up to one hell of an overclock without volt modding: 840 MHz core / 1100 MHz memory. A next gen MS system with it wouldn't be bad at all with 1 GB of GDDR5 total system memory on 128 bit bus with GPU. Get it down to 32 nm or lower and substantial speed increase could be made. Also is not Radeon 5xxx architecture capable of 2 pixels per ROP unit per clock (or effectively something like that) as well as some form of very low cost AA? Handles higher resolutions better than the G94M / Geforce 9800M GS (basically it's a mobile 9600GT, 256 bit interface, 64 shaders, 28 TMUs, 16 ROPs) in my laptop, though it's clocked much slower (530 MHz core/ 1250 shaders/ 850 memory) than the 5570.

As for Apple, I can't see them making a console or at least a games centric console. I see them making a full fledged computer, perhaps a scaled up Mac Mini with dedicated graphics made for better gaming and multimedia capabilities. The iPhone's popularity rested on the fact that it was primarily a communications device with many a great feature, gaming being one of them, with a new form of interface (touchscreen, motion), in a neat mobile factor. A basic console or even media machine I don't think would make all that much sense for them. Though thinking about this brings to mind MfA's "PC based console thread" from a month ago. Hell I even designed a PC based Steam centric console in MS Paint. Valve is known for games primarily though, Apple is not.

 

[TheWretched]

I recently bought a new laptop with a ATI 5650 and a Core i3 330M (2.13Ghz). Runs like hell^^ Just Cause 2 plays swell, even with nearly everything maxed out at "HD Ready" resolution (1368x768). And it runs cool, even maxing out heat (Prime95 with 4 threads and Furmark), it never gets above 71°C, and the keyboard still is cool to the touch.

And all that was in a package for 699€ (keeping in mind this system has 500GB HDD, 4GB RAM, 1GB VRAM, a display, Wifi, card readers, eSata... ...). I guess a console could easily take such a system and incorporate it into a sub 399€ console. But the step up from a 360 or PS3 would be quite small.

 

[Acert93]

Folks, let's try not to get too worked up. There's always the "do not respond" option.

On a completely different track, it's a bit fun seeing Redwood (40nm) specs if you compare it to Xenos (90nm):

Die Size
Xenos 182mm^2 + 80mm^2
Redwood 104mm^2

Memory Bus
Xenos 128-bit GDDR3
Redwood 128-bit GDDR5 or DDR3

ALU's
Xenos 48
Redwood 80

Texture filtering (bilinear)
Xenos 16 tex/clock
Redwood 20 tex/clock

ROPs
Xenos 8 pix/clock,
Redwood 8 pix/clock,

Core Clocks
Xenos 500MHz
Redwood's 650-775MHz

Mem Clocks (Bandwidth)
Xenos 700MHz GDDR3 (22.4GB/s)
Redwood 900MHz DDR3 (28.8GB/s) or 4Gbps GDDR5 (64.0GB/s)



Even if you kept the core clocks the same, it's still packing over 1.66x the shader power, 25% more texture filtering capability, and DX11 compliancy... The 10MB of eDRAM do throw things a bit out of wack in terms of direct die space comparison, but all-considering, Xenos is about 2.5x the amount of silicon.

And that's the thing. Getting the performance increase for the next generation consoles is easy because the current generation really are fairly slow. For example, you can currently get a 1GB 128b 5570 for $65 from newegg! It is likely that the integrated CPU/GPU combo coming from AMD will be even faster and will likely sell in the $100 range. The performance jumps possible from current/coming PC hardware are leaps and bounds higher than what both the PS3 and 360 currently offer. Coming up with a BOM for a significant performance leap over the existing designs isn't going to be hard.

The hard part is having all the developer tools and infrastructure in place.

Well, to keep things on pace where I am leaning for MS right now is AMD/ATI. I think the GPU will be whatever is in the pipes @ 230-300mm^2. I think they will go with 2GB of memory and maybe a Fusion style CPU (4-6 cores, smaller shared cache, and Llano/Vector guts on the CPU). MS will push OpenCL and other frameworks. There will be physical storage, but probably tiered (Flash, Small HDD, Large HDD) and the focus will be on connectivity and intuitive interfaces. They will target "good" hardware but will go with what the market has developed and not go crazy.

I wish they would go crazy though. The big hurdle will be ensuring all the XBLA stuff continues to work and ensuring major titles work out of the gate. i.e. Pitching a platform evolution instead of a new platform. MS has their fangs into their online customers, the key will be keeping it that way. With large digital libraries a lot of people won't want to lose their current library of XBL material. Icing on the cake would be, "Hey, AF and MSAA for everyone." We can wish--a lot of current gen games would look better with those added.

 

[Squilliam]

Well, to keep things on pace where I am leaning for MS right now is AMD/ATI. I think the GPU will be whatever is in the pipes @ 230-300mm^2. I think they will go with 2GB of memory and maybe a Fusion style CPU (4-6 cores, smaller shared cache, and Llano/Vector guts on the CPU).

Llano is still using the Phenom based architecture right? Wouldn't it make more sense for a console targeting ~75-125W to use a fusion architecture based upon Bulldozer modules with the two integer cores sharing the one floating point module if you've got the handy GPU right next door and its a low latency round trip to leverage the vast capabilities of the GPU in this respect? Then they just need 2-4 Bulldozer modules and they can round out the remaining die area with the GPU + memory controllers + onboard flash and any other junk they may wish to cram in there.

 

[Acert93]

We can talk Bulldozer when AMD gets working samples into production Big ifs on its performance and availability on the roadmap.

 

[Squilliam]

The future is fusion... believe!

Or

It only does everything!

With catch phrases like these, who needs a working prototype?

At least we ought to know it should be faster than Phenom II, Right?!

 

[Rangers]

No sense talking about these low end options that wont be in next gen consoles.

 

[AlNom]

Perspective.It was just a fun exercise to see where we're at in terms of silicon tech, die space, power, and performance after ~4 years. sigh, anyways. If you'd like to contribute, feel free by comparing current silicon area costs. Really now.

 

[corduroygt]

I'm still wondering why would we want consoles with x86 CPU's with all that wasted silicon area for decoding legacy instructions, not to mention BC would be a problem. Just having more small PPC cores with some improvements like OOE (plus SPU's for PS4) should be more than enough for both consoles and focus should be on the GPU side.

 

[Squilliam]

I'm still wondering why would we want consoles with x86 CPU's with all that wasted silicon area for decoding legacy instructions, not to mention BC would be a problem. Just having more small PPC cores with some improvements like OOE (plus SPU's for PS4) should be more than enough for both consoles and focus should be on the GPU side.

Because the area dedicated to legacy X86 as a proportion of total die area is declining with time and X86 probably has some of the greatest investment in terms of power/watt and overall performance/mm^2 as well as excellent power management features which could save 2000+KW/h over the lifetime of the console. In addition to this, X86s strengths best compliment unified GPUs relatively speaking compared to the POWER range of processors.

 

[liolio]

Perspective.It was just a fun exercise to see where we're at in terms of silicon tech, die space, power, and performance after ~4 years. sigh, anyways. If you'd like to contribute, feel free by comparing current silicon area costs. Really now.

I find the comparison pretty interesting, nowadays it's possible to put together a system that would beat the crap out of a ps360 within the silicon area the main xenos die took back in time, say a athlon X2 stuck to a redwood derivative. If we take the daughter die in account we can stuck a juniper with athlon X2.
For future system I hope that manufacturer will be pretty reasonable systems are likely to hit the road late and as I'm not to buy a system costing more than 200$/€ it could be a while before I get my hand on something new...
EDIT
I say an X2 but an Intel i3 would better and its die size is 81mm². By watching at the fermi disaster in perf per mm² and watt Intel reworked larrabee could prove a bothering competitor in not that distant future. I read on a website that there are rumours about Intel launching a product whom code name is "moise" in Q2 2011. It seems compute related it's unclear if it's related to SCC or Larrabee. It's a possiblity that Intel may end with competitive hybrid part sooner than expected and thus that manufacturers consider Intel offering. A reasonable estimation could be an i3 stuck to twelve larrabee cores for a package ~200mm² @32nm. If larrabee cores are clocked somewhere between 1.5 and 2 GHz for what it's worse the peak FLOPS figure would in the league as AMD/ATI offering.