Design a console with better specs than PS4 and Xbone, constraints apply

P

Proelite

Veteran
Supporter
Constraints:

TDP that you would consider reasonable.
Should be capable of being manufactured in 2013.
Should fit into a backpack size wise.
<=$500 at retail
Aiming for 50 million consoles sold and profitability within 5 years.

Mine:

AMD APU 8 core jaguar + 512 SPUs + > 64MB E-ram, ~ 1 teraflops

AMD Discrete GPU, 1.0ghz, 1024 SPUs with 4GB of GDDR5 ram, > 2 teraflops

8GB of DDR3 main rain, ~6 available for games

Split controllers with gyroscope, accelerometers, and IR.

Controller sensor has 720p video camera with night-vision and microphone.

No Kinect-like full body motion gaming.
 
Bagel seed said:
Easy, PS4 overclocked to 1.8ghz CPU 1.0ghz GPU plus a bigger fan, heatsink and box. Based model priced at $499.
Click to expand...

Maybe increase the CUs to 24, add some embedded ram to provide enough bandwidth for the 32 ROPs and extra CUs?

> 3 teraflops

225 TDP total for the system.
 
8 core amd jaguar with a 12 meg l3 shared cache .

12 cu built in apu with 16 rops and 64 megs of sram

12 cu discrete gpu with 16 rops and 64 megs of sram

12 gigs of ram for the apu and 8 gigs of ram for the gpu. DDR 3 of course.


That way the OS has enough ram . Both gpu's have acess to 8 gigs of ram and you have more than enough cpu power.

I'd drop the optical drive since its a waste of space. I'd also give the sound chip 2 gigs of ddr 3 ram
 
@Proelite: Wouldn't that be a little bit too big with 6 additional CUs AND embedded RAM?

@blakjedi: 32 CUs and 68GB/s bandwidth?

@eastmen: Sounds complicated. Two different GPUs? Three different RAM pools?
 
intel i7 4 core 8 thread - 1-2 ghz - 17watt
nvidia gk104

both in a soc with 256bit memory controller - and 8 gb of gddr5 ram at 6ghz
 
My guess is everybody is going to assume a too low BOM for their project here. Since we dont know actual costs we're all just throwing crap at the wall.

One thing's for sure, Microsoft should have had more CU's.
 
Trevor said:
@Proelite: Wouldn't that be a little bit too big with 6 additional CUs AND embedded RAM?

@blakjedi: 32 CUs and 68GB/s bandwidth?

@eastmen: Sounds complicated. Two different GPUs? Three different RAM pools?
Click to expand...


Not very complicated in all honesty .


Just like cross fire today the two pools of memory would be hidden from developers. They will just clone the same data. The extra 4 gigs attached to the apu would be where the 3 oses stay.

The ram on the sound chip would be handled by the sound api .
 
Well I would think that within what seems to be their silicon and power budget it seems tough to do better than the ps4 and Xbox1 combo before Sony went for that expensive memory bump.

It looks tough to do significantly better, I would think that Intel could do a better durango using a set-up ala high end Haswell (with the 128 MB of EDRAM) using their up coming Atom (that's if their current GPU architecture is scalable enough which it might not be).
But then there is the PS4 and imo it is tough to get any better, it is a plain UMA system oozing with bandwidth, you can't really beat that without spending more silicon (and power) than what Sony did.
The last Killzone seems to leverage this convenience with its 800MB of render targets ( read the prez yesterday night :8O: ).

Then you enter the realm of custom hardware but the costs seems to high now and actually beating existing design is not a given even if one were to pour insane amount of money in the project (Durango design (SoC only) could have cost 3billions on its own).

On the other hand design a cheaper but competitive design (what could have been the WiiU) is definitely doable though.
 
I would wait for 20nm early next year and use the roughly 2x transistor budget advantage to make a real difference. That would mean 24-32CUs/32-48Rops(if that's what it needs to make effective use of the ESRAM), 64-96MB ESRAM(less power/better yield than logic) and 3-4x DDR4 channels.

and adjust the numbers to fit into a 175-200W TDP and reasonable cost/yield. WIth 20nm and DDR ramping up rapidly next year MS might only have to buffer the premium costs for 1-2 years.

Kinect optional for the people who want it like now.

When the 20nm GPUs come out next year XBone will look ever more pathetic than it does now.
 
A Kaveri with 256bit gddr5 and enough CUs (i.e. like the PS4, but better CPU and confirmed HSA)

Or a triple core Haswell / 6 threads (why? because I can) running at 4.0GHz along with a GPU similar to GK104, but with actual GK110 stuff inside it (minus the ECC and fast double precision). Maybe two less SMX : a full GK104 is too hot for a console.

Have all outputs : DVI dual link (compatible with VGA), latest hdmi, mini displayport, high quality stereo sound, optical sound and a selectable refresh rate for the display (including a true 120Hz output)
 
I think that the current Microsoft and Sony designs are good. It's hard to improve them and keep the thermals / cost / ease of programming (for all kinds of programmers) intact.

---

Personally I would have liked to see more CPU cores in the next gen consoles. It's true that some developers still have these old style "dedicated thread per task" engines that do not scale well to high core counts. But those kind of engines are bad in many other ways as well (waste lots of cycles because of load balancing issues). Fortunately many developers have already transitioned (or are transitioning now, see KZ Shadow Fall slides) to fine grained task based threading systems. These kinds of systems scale very well (and scale automatically) to high CPU core/thread counts.

For these kind of optimized game engines, it would be better to have more low clocked cores than less high clocked ones (because clock scaling usually results in nonlinear x^2 power increase).

For my own console, I would have thus included 4 Jaguar CUs (16 cores) and ran them at slightly lower clock to compensate the thermals. It seems that Jaguar follows x^2 power envelope when clocking it higher (1 GHz Jaguar core is 1W according to AMD Temash slides, while 2 GHz core is around 4W based on GPU less Opteron "Kyoto"). Thus approx 30% lower clocks should halve the core power consumption. That would allow us to double the number of cores. And thus we could have 70% * 2 = 140% CPU total multicore throughput compared to the current designs (but of course the trade off would have been 30% less single core perf). Jaguar CUs are so small that this change doesn't affect the chip size/cost in a meaningful way.
 
Constraints:
Power consumption: < 120W
Cost: ~$350
Price: $399

APU:
8 x 1.6GHz 2-way superscalar OOOe CPUs. One reserved for yield boost, one reserved for OS
GPU: As many ALUs as the power budget and costs constraint allow (let's say 768 in 28nm).
32MB on die RAM

Memory:
8GB DDR3

Storage:
500GB single platter 2.5" HDD

Human interface:
1 x wireless controller
1 x combined aural and visual input system

Market goals:
Profitable from day one.
200 million sold over 10 years.

Cheers
 
Last edited by a moderator:
Gubbi said:
Constraints:
Power consumption: < 120W
Cost: ~$350
Price: $399

APU:
8 x 1.6GHz 2-way superscalar OOOe CPUs. One reserved for yield boost, one reserved for OS
GPU: As many ALUs as the power budget and costs constraint allow (let's say 768 in 28nm).
32MB on die RAM

Memory:
8GB DDR3

Storage:
500GB single platter 2.5" HDD

Human interface:
1 x wireless controller
1 x combined aural and visual input system

Market goals:
Profitable from day one.
200 million sold over 10 years.

Cheers
Click to expand...

This is hilarious. No one can say you don't have a sense of humor. lol
 
Single die with 2x AMD 'Richland' APU
8 CPU Cores (limited to 2,8 GHz)
768 GPU Cores (800MHz)
Replace the 128 DDR3 interface with a 192 bit GDDR5, 6 GB of RAM
70W TDP (A-10 5757 has 35W TDP)

Minimal OS and non-gaming apps that take 512MB or less Max

Bluray drive, 64GB solid state storage for OS and user space.

External mass storage supported, via USB 3.0, can be used for installing games too.

Ethernet, WiFi, HDMI, VGA and Composite (SD).

349 at launch
 
Im not sure why almost everyone would put eSRAM in their console in such quantity (and die space). It makes less and less sense to have that kind of memory, while machines gets more and more powerfull, easying up developpement is (should) be the only concern now (aka having one and unique memory pool and bandwidth to deal with). Unless it's interfaced for a lot of bandwidth that really makes a difference (x10 times we can talk).

Microsoft didn't put eSRAM because they thought it was great or fun, they would have been way too far behind in bandwidth without it, that's all and it was a relatively cheap way to work around this issue while keeping a huge amount of cheap memory (same might apply with Nintendo design)

Also, there s no way you can launch a console early next year on 20nm process.. it's not a limited edition graphic card to launch.. it's million of consoles and on a proven process.

I would just put for more bandwidth, cause you never have enough... At least 200 GB/s and im good with 18/20 CU.
 
With the size of the XB1 allowing a chunkier cooling system and more airflow what is a reasonable max TDP of a theoretical console in a similar size box? Previous to the reveal most speculations on potential console specs were shot down if they involved a large case, i guess we can now be more generous in this regard when revisiting what could be done.
 
I'd take Jaguar and add AVX2 compliance to them (doubling FP throughput), up CU count to 24 and increase the GDDR5 back to at least 192 GB/s. All hail my huge APU.
 
You must log in or register to reply here.

Similar threads

R
Building a PC with XB1X specs *spawn
Replies
10
Views
2K
Blazkowicz
B
Back
Top