Predict: The Next Generation Console Tech

[french toast]

I guess you know that a AMD A8-385O, so 4 OoO cores and 400sp @600MHz has a TDP of 100W?
GDDR5 memory controller consume more juice. It' already a 225mm2 chip?
Then 6GB of GDDR5 seems complicated on a 256GB bus may be doable but still that's a lot of memory to place on the mobo (12 I think).
I get it that its a "top end design" but I don't think it doable, would be too big too hot.

EDIT
As for the article I believe that Onlive is not a threat to console as the infrastructure is not ready. May be it can take off in some "newly developed country". It like the Onlive concept, I'm subscribing at the time but will stop soon as quality is not good enough, the experience is unstable (most likely because my connection) it goes from good enough for quiet some genre to "deal breaking". I don't how the thing work on a great connection but I suspect that latencies are still too high for many competitive mp games. Overall it can become an establish actor and set its own niche but that's it.
On the other hand I believe that there is room from client-server relationship but I won't go there as there is a proper thread for this

As for the upgradable console, it's a bad idea imho. Introducing new significant peripherals that potentially split your user base is already a iffy move, it's unclear the third party editors support you will receive, the device has to be successful.

Your right about the heat, thats why i suggested a custom water cooling solution..i know its not cheap as it is, but it would be possible to design an efficient cheaper solution, there were already rumours of such a thing for the 360..so i know its possible.

The power at the wall is a big hurdle, if they managed to get it on a mature 28nm High-K gate first process, and with both the gpu+cpu custom built designs, Maybe chopping down some of the TMU's and ROP's i think it would be feasable.

EDIT; I think new & effiecient 32 ROPS and 64 TMU's would be sufficient IMHO.

And the ram upgrade would be free for every console, so fragmentation would be non existant, you could also say hold a core back and then upgrade the ram and enable the lost core for every console for a mid life kicker, you could also enable a small over clock to the gpu and cpu at the same time.

 

[liolio]

Those 100W aren't measured as if they were thermally relevant (or even battery relevant) consumption values, it's just a fixed ceiling for power consumption making it a bit irrelevant in practical terms.

I don't get your point, that's what TDP is.

And it isn't really fair to use a 7 month-old chip which was constantly delayed and already came pretty late by itself.
A more "just" comparison to the state-of-the-art will be in a month or so, when Trinity numbers come out.

Well I believe the answer is yes. It's a pretty successful piece of silicon. for budget gamer and through dual graphics it offers performances that Intel can't touch. I don't think there are any problem with power consumption of the chip, bulldozer on the other hand... (and let see how trinity will fair... )
It's using 32 SOI HKMG process. It seems that there is a revision already (3870k).

Trinity is using the same process, so I don't expect miracle. I expect marginal differences if anything it will save AMD some money as I suspect the chip to be tinier than Llano. they are using a new cpu architecture that as a lot to prove imho.

EDIT
multi quote incident, for some reason split one block in two.
EDIT
more terrible than expected quoting incident, for some reason a part of your first sentence was cut.
EDIT
And how dare you say TDP is irrelevant it serves as a guideline for OEM? You think that they are to fine tune the cooling to the Watt in precision? Neither you think it's a good idea to have just a bit of overhead just in case? Come on, AMD for its APU have 3 TDP I think 100, 65 and a last one I don't remember, on a really practical case all the products in a given TDP will use the same grade of cooling solutions.

 

[liolio]

Do you have a source for that figure? I just read a review of an A8-3850, and the entire system pulled 90 watts at load.

That the TDP AMD gives it doesn't have to be that precise. 10% is a marginal difference.
The SoC describe is x2 the CPU clocked faster, x4 the GPU, and gddr5 memory controller. Do you expect the difference to be in the 10% ballpark?

 

[french toast]

I don't get your point, that's what TDP is.
Well I believe the answer is yes. It's a pretty successful piece of silicon. for budget gamer and through dual graphics it offers performances that Intel can't touch. I don't think there are any problem with power consumption of the chip, bulldozer on the other hand... (and let see how trinity will fair... )
It's using 32 SOI HKMG process. It seems tht there is a revision already (3870k).
Trinity is using the same process, so I don't expect miracle. I expect marginal differences if anything it will save AMD some money as I suspect the chip to be tinier than Llano. they are using a new cpu architecture that as a lot to prove imho.

From what i can gather, Trinity will be 50%faster graphics and 25% cpu..on that same process (mobile)
Also bear in mind custom built power pc cores are far more efficient than x86..especially the Pentium 4 class BD.

And i also said 28nm high K mature process, Hell if we are talking end of 2013 it might even be 20nm...

What i have suggested is certainly possible, and although would be a challanging design, microsft has a very sustainable income and profitable xbox division to prop it up, has had the best part of 7 years to design it and will likely be the last of the mohicans, a swan song, we will never see the traditional console like this again.

With Sony and Ninty on the ropes, Microsoft has the chance to stick the knife in..and i think they will.

 

[liolio]

From what i can gather, Trinity will be 50%faster graphics and 25% cpu..on that same process (mobile)
Also bear in mind custom built power pc cores are far more efficient than x86..especially the Pentium 4 class BD.

First I take it that it poses no problem to your logic to state that BD modules used in Trinity offer P4 level of performances and so that it will perform 25% better than the Star cores " state of arts K10 cores" found in Llano. And BD doesn't offer p4 level of performances even if its single tread perfs are way too low.
What you gathered is wrong, or incomplete.

AMD claims the 17W Trinity should offer similar aggregate CPU/GPU performance to existing Llano notebook APUs at ~35W. The standard voltage notebook Trinity APU will offer a 25% increase in CPU performance and a 50% increase in GPU performance over the A-series Llano APUs available today. Finally the desktop Trinity will be 15% faster on the CPU side and 25% faster on the GPU. Although AMD didn't disclose details, it's likely that these numbers are comparing a two-module Piledriver based Trinity to a quad-core Llano.

First it's a claim from AMD. In the class of hardware we're interested in the numbered "claimed" by AMD are 15% and 25%. To put things into perspective we have:
2 BD modules vs 4 star cores
80 VILW5 units vs 80 VLIW4 units + other architectural refinements.
For what I've read about BD so far and 69xx card I expect those numbers to turn out true on anything but a hand few of selected applications / benchmarks.

By the way the TDP are already announced, 100, 65, 35 and 17 Watts. No magic here

And i also said 28nm high K mature process, Hell if we are talking end of 2013 it might even be 20nm...

1) I'm not sure if this process will offer any significant improvement over what TSMC offers.
2) The SoC would be using SOI HKMG process which is a 32nm process.
3) on a 22nm process it would be more practical, still it's impossible for a launch in 2013 (they need to tape the thing out, test the design, run production, build enough stock, etc. => impossible)

What i have suggested is certainly possible, and although would be a challenging design, microsft has a very sustainable income and profitable xbox division to prop it up, has had the best part of 7 years to design it and will likely be the last of the mohicans, a swan song, we will never see the traditional console like this again.

Certainly being the important word, llano is 220mm2 Trinity will be in the ballpark, you want two time as much cpu cores, x5 the Gpu cores, if you look at a llano die shot you will found out that you might really well it the optical reticule limits which +550mm2. In the best case you end up with a fermi size level ship, with the matched low yield, +200 Watts power consumption.

With Sony and Ninty on the ropes, Microsoft has the chance to stick the knife in..and i think they will.

Well they will give it a try that's what companies competing on a market do, it might not take what you think it needs to do so and whether they succeed or not is a completely different matter.

 

[french toast]

First I take it that it poses no problem to your logic to state that BD modules used in Trinity offer P4 level of performances and so that it will perform 25% better than the Star cores " state of arts K10 cores" found in Llano. And BD doesn't offer p4 level of performances even if its single tread perfs are way too low.
What you gathered is wrong, or incomplete.

Lol of course i know the newer BD core is faster than Pentium 4, i was just having a sly dig at the low IPC and poor effeciency of BD...along with the scathing reviews on launch i think it compares very well right now actually.

Why would it be made on 32nm?

 

[AlphaWolf]

That the TDP AMD gives it doesn't have to be that precise. 10% is a marginal difference.
The SoC describe is x2 the CPU clocked faster, x4 the GPU, and gddr5 memory controller. Do you expect the difference to be in the 10% ballpark?

The ENTIRE SYSTEM pulled 90 watts from the wall.

<edit> checking some more reviews, the 90 was on the low side. Anand dragged 126 from the whole system at load.

 

[Deleted member 13524]

I don't get your point, that's what TDP is.

In the case of AMD, it's not.

TDP stands for Thermal Design Power and is "the maximum amount of power the cooling system in a computer is required to dissipate."
This means the term TDP was originally used to calculate the constant thermal power output of a chip during steady-state operation, at a given room temperature (usually 25º).
In the case of CPUs, they usually used stressing benchmarks to calculate the TDP as average heat output and consumption values, which is how Intel does.


During a thermally-irrelevant period (say, an instant or a couple of seconds), Intel's latest CPUs may actually surpass the TDP in power consumption, as long as the average heat output and power consumption don't pass the TDP barrier.


AMD's TDP is the absolute maximum electrical power the chip can take, which is usually a lot higher than what the chip is consuming (power measurements of Llano desktop 100W APUs are actually hovering the ~65W while stressed).

Nonetheless, AMD also gives away the ACP (Average CPU Power), which is in fact the equivalent of Intel's TDP and is a lot more relevant for building a system.

And how dare you say TDP is irrelevant it serves as a guideline for OEM? You think that they are to fine tune the cooling to the Watt in precision? Neither you think it's a good idea to have just a bit of overhead just in case? Come on, AMD for its APU have 3 TDP I think 100, 65 and a last one I don't remember, on a really practical case all the products in a given TDP will use the same grade of cooling solutions.

AMD's TDP is actually a little bit irrelevant, except maybe for designing critical power circuits. I doubt OEMs use it as guideline, since the ACP value is a lot more important for designing the cooling system.

 

[Brimstone]

I also think Microsoft would loath to use blue ray if they some how could get around it, the only scenario i think they would is if they added that rumoured DRM rubbish to stop second hand resales..other than that i think it would be both cheaper and more beneficial to use a propreity flash storage, then they could save face, reduce the cost AND reduce the over all size of the console(no royalty payments to Sony..).

Developers would also get to only buy the amount of flash they need tailored to the exact game size..no redundant disk space,Cut down on noise and if its good quality flash streaming shouldn't be problem like it is on blueray currently, the prices of flash are decreasing year on year.

A better question in my opinion is what DISC technology will Microsoft use? The physical optical drive will be essentially the same for either a Blu-Ray disc or a GE Holographic disc. GE made a breakthough with the materials they use to make the physical media disc.

With a holographic disc you wouldn't need to have a hard drive standard, which leads to lower prices as the console has CPU/GPU die shrinks later in the console cycle. It's essentially a hologram based disc versus magnetic based disc. There is nothing special about a hard drive really. A consumer could buy either a hard drive add-on or use flash memory sticks.

Talk about anti-piracy...build a game engine with Partially Resident Textures (John Carmacks Mega-Texture technology) and have your game weight in around 500 Gigabytes or more. How many would bother to pirate with such massive downloads?

 

[pjbliverpool]

Also bear in mind custom built power pc cores are far more efficient than x86...

Why would it be? Are there benchmarks proving this?

 

[3dilettante]

"All else being equal..." would be the start of any answer.
The rest of the answer would involve laying down how much of "all else" is equal.

 

[Rangers]

Just thought of Moore's law earlier for some reason. Back to basics.

2X transistors every 2 years.

November each year

Xenos 330m 2005
2007 660
2009 1.3
2011 2.6
2013 5.2
2014?

2.6 actually sounds quite reasonable for a console launching in late 11. Would have basically been a Caymen which are right at 2.6. Need not be the 6970, would be downclocked et al.

5.2 in Nov 2013 sounds a little high, but I wonder if it will when we get there. SI is 4.3b, Almost two years so we'll get another AMD gen on 22nm around that time? It might be kind of close. SI wasn't 2X Caymen, so lets peg Sea Islands at 6-7b. 5.2 in Nov 2013 might be something from the mid-upper Sea I range.

That's a lot of power. I'll stake my claim it'll be closer to that than the, 5w lily livered box that seems to be the de jour.

Also I noticed 360 followed Moore's law pretty perfectly, if not beating it. NV2a=60m, 4 years later should have been 240m, 6 yrs=480m, 5 years later Xenos weighed in at 337m, split the difference.

2014 (my expectation for next box) works too. By moore's 2015 would see 10.4b. So 2014 should be between 5.2 and 10.4. Sea Islands should be 6-7b, right in the middle.

Put another way if I'm expecting 5.2 for Xb3, it shows just how far the rumored 6670 would deviate from moore's. It's a .7b chip, so something like 1/7 what moore's would expect, aka major sea change.

More idle numbers:

I am adamantly opposed to the 6670 rumor, but lets discuss it as true:

716m transistors. 480 SP's. 800 mhz.

the 60% clock increase over Xenos (lets assume such a puny chip could run at full clock) give it theoretical 3.2X Xenos shader power. At 1080P (2.25X720P) it would have about 40% more shader power per pixel than Xenos at 720P. (Yes, weak).

Transistor wise it cannot even get to the supposed 6X 360 power. 1.6X716=functional ~1.15b compared to Xenos 337m. Only 3.4X.

Of course were this scenario true maybe MS think the last 2X is in 6670's superior features, or something. Overall though it's just not there to be 6X.

Ahhh, but I'm forgetting EDRAM, add a quarter billion trans of EDRAM or something things change. Now it's a 1b chip, 1.6b equivalent after clock, equivalent to 4.7X Xenos by this figuring (simple clock X transistor count). Closer but still light. And 3.2 the shader power still tells the tale. That's not enough jump to even get anybody buying, another reason I dont believe this rumor.

For the low spec rumors, the lowest I would even be remotely happy with or see as reasonable would be a Barts. 1120 SP's, 1.7b transistors. The HD 6870 is a nice card. Now you are also talking at least 6.5X Xenos shader power assuming a 700 mhz clock as well.

 

[function]

The ENTIRE SYSTEM pulled 90 watts from the wall.

<edit> checking some more reviews, the 90 was on the low side. Anand dragged 126 from the whole system at load.

Consoles are probably going to get much closer to their TDP than any PC part would during gaming based purely on the fact that the hardware is better utilised and spends less time idling or stalling.

After the disaster that was the 90nm Xbox, MS appear to have taken testing on the 360S very seriously. They made sure it could handle a power virus pushing both the CPU and GPU simultaneously, and their testing showed that existing games (and this was a couple of years back) were drawing 80%+ of the power virus load.

The Anandtech A8-3850 review* I think you're referring to lists a Metro 2033 "load" test as drawing 126W and the CPU x264 encoding "load" test as drawing 123W. 126 CPU & GPU vs 123 CPU only. So, basically the same. I doubt their Metro 2033 "load" test is pushing the A8-3850 anywhere remotely as close to its maximum power draw as a console game using the processor would push it (and the kind of CPU+GPU power virus that the 360S was required to handle without issue would be another level still).

*http://www.anandtech.com/show/4476/amd-a83850-review/9

 

[liolio]

The ENTIRE SYSTEM pulled 90 watts from the wall.

<edit> checking some more reviews, the 90 was on the low side. Anand dragged 126 from the whole system at load.

Well hardware.fr which usually has really precise measurement give it 136 max (furmark). Anand figure has to be correct.
I can find north of 100Watts through undervolting. Still AMD doesn't feel comfortable shipping the product at this tension.
NB I don't know how the 3870K fare in this regard.

Tottentraz if you want to discuss a hand full of Watts and the discrepancy in Intel and AMD naming go for it.
126 Watts while playing Metro 2036
44 watts idle
That's 82 watts, take in account the power supply (in)efficiency 90% and you get 76 Watts.

There is a difference but I can't see how it is relevant while either comparing Llano to Trinity or to the the monster that French Toast was describing. Trinity will have a 100W TDP like Llano.
If you prefer Trinity will consume around 75Watts "for real" like Llano.
For French Toast SoC you can easily consider a 200 Watts of TDP and 150 Watts of "real power consumption, and that really low balling it (a simple x2 over llano which is not).

So now what is your point? You think it made his point more likely?

 

[liolio]

The Anandtech A8-3850 review I think you're referring to lists a Metro 2033 "load" test as drawing 126W and the CPU x264 encoding "load" test as drawing 123W.

FYI Hardware.fr got 136 running furmark and Prime 95

 

[almighty]

Your right about the heat, thats why i suggested a custom water cooling solution...

Think you'll find that was me like 20 pages back..

 

[function]

FYI Hardware.fr got 136 running furmark and Prime 95

Good find!

I've been tinkering around with my Sandybridge 2500K recently, and at 4 gHz, Core Temp 1.0 lists 4 x Prime95 as drawing ~ 70W and pegs CPU temperatures at about 53 C. Running Intel Burn Test instead of Prime95 raises power to ~ 105W and temperatures go up to about 61 C. I don't know how accurate the power figures are but the temperatures are right. Prime95 is definitely not the furtherest you can push a CPU, and by quite a long way - at least not on my CPU.

I wouldn't be surprised if the Hardware.fr power figure could be pushed somewhat higher still!

 

[AlphaWolf]

Still that's the entire system.

 

[french toast]

Why would it be? Are there benchmarks proving this?

HA, I knew i was opening my self up to that, funny enough i did find such a graph comparing all architecturers performance per watt, i think it was at a forum discussing Chinas ICUBE UPU or something, anyway for the life of me i cant find it now

Anyway you dont need graphs to see the truth, ALL consoles manufacturers uses IBM PPC last gen, and ALL consoles will use them again this gen, outside of desktop where other things come to play such as software, POWERPC makes total sense, else why would they all use it?

From a performance per watt perspective, and performance per mm die, POWERPC dominates, i think only mips beat it and this new UPU uarch from ICUBE.

Transistor wise it cannot even get to the supposed 6X 360 power. 1.6X716=functional ~1.15b compared to Xenos 337m. Only 3.4X.

Of course were this scenario true maybe MS think the last 2X is in 6670's superior features, or something. Overall though it's just not there to be 6X.

You don't really believe that 6670 nonsense do you? besides a modern gpu would not need to have 6x the transisters and 6x every hardware metric to be 6x the performance of 360.

The ROPS on GCN for instance are much more powerfull than the ROPS from Xenos, so will the rasterizer and Likely TMU's, the shaders would be much more efficient per ALU, also you could increase the ALUs more cheaply to get better performance than just replicating the whole architecture 6x, not counting likely bandwidth increases just by replacing GDDR3 for GDDR5.

So a more modern design of only 4 x the transisters could equal 6x or more quite easilly IMHO.

LILIO; I did forget to add the TDP of 250w..which is not that unreasonable compared to original 360, I think my design is feasable, but only in a custom layout, not off the shelf parts, there are many parts of the gpus such as video decode/encoders/eyefiniy etc etc that would not be needed, that and the chop down to say GCN architecture improvements, with say 32 ROPS 64TMU's VLIW 5 only 256bit bus, and the use of powerpc cores which are small and effecient, especially on something like 28nm high k..i think it is possible.

Certainly im expecting shed loads of ALU's.

EDIT; What i am sugesting is a modifed and updated Barts XT aka 6870.

For comparisons sake..

900mhz clock
2016 TFlops, 1120 VLIW 5 ALU's 32 ROPS 56 TMU's 256bit bus
That had 1.7B transisters, 151w load.
255mm die on 40nm.

Evergreen..
850mhz.
2.7 TFlops, 1600 VLIW 5 ALU's, 32 ROPS 80 TMU's, 256bit bus.
That had 2.17B transistors, 188w load, 334mm die on TSMC 40nm,

If you take the Barts XT core, update the various components, taking inspiration from GCN but keeping Barts modified VLIW 5 arch, cutting out any un needed decoder/eyefinity stuff, increasing the TMU's and ALU's to 64 and 2000+ respectively and you would have a very powerfull graphics card.

If you then take that and produce it on a mature 28nm high k process, clock it at 700mhz and you would get something very interesting both in die area and power/heat consumption.

This along with say 4-6gb GDDR5 and octo core OoO PPC 4XSMT @ 3.2 + 256bit simd would easilly be 10x the x360, and be well within power/thermal/die constraints IMHO,

Perhaps someone with more knowledge could expand on those points...

 

[Deleted member 13524]

Well hardware.fr which usually has really precise measurement give it 136 max (furmark).

That's measured at the wall, so that's as precise as.. well..
There's no info on the PSU they're using.. something like "Silver 80+" wouldn't do much more than 83%, so that's 113W for APU + RAM + Southbridge + Voltage regulation + Cooling + Hard Drive + whatever.
So even with the furmark "power virus", the APU isn't hitting those 100W, which is exactly what AMD's erroneous definition of "TDP" means.

Tottentraz if you want to discuss a hand full of Watts and the discrepancy in Intel and AMD naming go for it.
126 Watts while playing Metro 2036
44 watts idle
That's 82 watts, take in account the power supply (in)efficiency 90% and you get 76 Watts.

There is a difference but I can't see how it is relevant while either comparing Llano to Trinity or to the the monster that French Toast was describing. Trinity will have a 100W TDP like Llano.
If you prefer Trinity will consume around 75Watts "for real" like Llano.
For French Toast SoC you can easily consider a 200 Watts of TDP and 150 Watts of "real power consumption, and that really low balling it (a simple x2 over llano which is not).

So now what is your point? You think it made his point more likely?

Given that the original X360's power brick was rated at 203W then yes, I think it did.

Some things wouldn't fit very well on a console like 6GB GDDR5 on a 256bit bus, and 8 cores w/ 4 threads each would be a bit of an overkill. Sacrificing CPU cores in order to switch to GCN could make more sense IMO, since I think developers would rather keep all the visual processing in the GPU.
Four 4-threaded OoO PowerPC cores at 4.5GHz would probably be better, though we're yet to see what frequencies the 28nm HP will allow at a reasonable power consumption.

This along with say 4-6gb GDDR5 and octo core OoO PPC 4XSMT @ 3.2 + 256bit simd would easilly be 10x the x360, and be well within power/thermal/die constraints IMHO

I don't think it'd be that much interesting to increase the X360's CPU performance by ten-fold through a lot of parallelization.
As I said, I think developers would be more attracted to something like 3x the single-threaded performance, while adding a core or two.

I say this as I look at Intel's dual-core i3 sandybridge performance in games, compared to the 6-core Thuban from AMD.