~1 TFLOPS GPU? 14nm? HBM?
I'd like to see that but have we learned nothing from the past?
I'd like to see that but have we learned nothing from the past?
- Status
We're thinking a 1TF GPU is unlikely, while nVidia is putting that in Shield TV already in a super slim form factor using a mobile part?
Yes because nVidia != Nintendo
nVidia are specialized in high tech GPUs and devices while Nintendo historically has almost* only used the cheapest hardware for its devices.
*Gamecube
D
Deleted member 13524
Guest
HBM to the rescue?
(we can dream)
With a single 8-stack HBM rev.2 the NX could have 8GB, have the luxury to clock the memory below its maximum spec and even then beat the PS4's 176GB/s total bandwidth.
Would it be totally cool? Yes.
Will it happen? 99% no.
Even then, Maxwell 2 seems to be much more efficient at using their VRAM bandwidth than anything from AMD so far, yet it gets around 20 GFLOPS-per-GB/s, which don't have to be shared with a CPU.
With 50GB/s and without eDRAM the NX would be severely constrained, since up to ~20GB/s (as much as the other consoles) would eventually be used by the CPU so the GPU would have to deal with 30 GB/s at times.
With even a 750 GFLOP/s GPU and 30GB/s, the GPU would have to deal with ~25GFLOPS-per-GB/s. That's 25% more than nVidia's best.
Esram?
There's a GPU vendor out there with a 16 ROP / 32MB template just lying around ...
Looking at the XBone and PS4... my impression is that with the density and speed of GDDR5 chips that are currently available (not to mention HBM), spending transistors and die-area on ESRAM + cheaper DDRx seems like a huge mistake.
At least Sony had the uncertainty of having to gamble with the density of GDDR5 chips available for 2013. Nintendo would never have that concern. In fact, they could even get 7000MT/s GDDR5 for relatively cheap nowadays and have 112GB/s with just a 128bit bus.
Well they did release a console with presumably 175GFLOP/s in late 2012. That's about half of the lowest-end mobile GPU that AMD had to offer since the beginning of that year.
So half of the lowest-end mobile GPU AMD has to offer right now would be... ~330 GFLOPS
Probably the question of having a GPU with less or more than 1TFLOP/s has to do with it being a mobile console or not.
Nintendo can lowball the marketplace with a cheap ass 1TF piece, I'm sure (if not portable). That should be the minimum we should be expecting IMO. Else they're console is properly doomed plus, unless it's like $100 including new Mario or something.
It's a rudimentary comparison. As architectures improve, what can be done with the available processing power improves. So 1 TF on a later GPU can do more than 1TF on an earlier GPU. However, we can't quantify that and for the sake of discussion need recourse to fall back to the raw metric, until someone invents a holistic measure and we can talk about a 3.4 Dougie GPU in NX beating the 2.9 Dougie GPU in PS4 regardless how many TFLops it manages. It's benchmarks after the fact that give us more realstic performance comparisons, and these generally aren't available on consoles leaving fanboys to Make Shit Up about how powerful their console of preference is.
D
Deleted member 13524
Guest
Wii U's Latte was ~156mm^2 at 550MHz and 40nm. 156mm^2 using 28nm and including north+southbridge would be enough only for a 10CU Cape Verde. If they clock it towards 800MHz like the other consoles then it'll do 1024 GFLOP/s.
But this is only if they don't decide to put a big chunk of EDRAM/ESRAM in there that occupies almost 1/3rd of the chip (like they have for the last 3 generations of home consoles).
That would mean Nintendo acknowledges the hypothesis that their consoles' success depends on their processing power relative to the competition.
So far, they haven't. At least not publicly.
For a higher end system (relative to consoles) esram does seem like a bad move at the moment, especially if you were to factor in the BW saving features that will be on Polaris GPUs.
However, for a lower performance, more power constrained system embedded memory still offers very high bandwidth for very low power and footprint. On 14 nm 32 MB should be less than 40 mm^2 and take very little power, while offering BW potentially in excess of a 256-bit GDDR5 bus drawing many times the power. Plus, as the process matures cost will continue to fall. XBone is strange in that despite having embedded memory they still require a 256-bit bus with no less than a whopping great 16 memory chips soldered to the fekkin huge mobo.
Would this ever make sense for Nintendo? Don't know. But in addition to fitting better into a small and power constrained handheld, you could scale up clocks to north of 1 gHz and make good in a (reasonably) competitive console too.
In this scenario, you might be looking at:
366 mhz esram , 64-bit LPDDR4 2333, half size GPU and esram
1000 mHz esram, 128-bit DDR4 3200, full size GPU and esram
... depending on configuration. You get a small and simple mobo, low power, scalable architecture, and DDR4.
Not very likely, but under the right circumstances it might work out.
(Just a minor point, but Latte ended up being 45 nm NEC process, CPU being 45 nm IBM.)
On 14 nm (I know, I know, daring to dream) for 156 mm^2 Nintendo would be able to more or less get the Xbox One SoC. Roughly .5 scaling, half bus width for DDR4 so much less IO, probably a third or more off the power consumption.
If they went with fewer CPU cores maybe they could go even smaller. If Nintendo decided to target 14 nm and say 200 mm^2 and 80 Watts system draw they could easily be banging with the Xbox One in a case half the size.
Well they did release a console with presumably 175GFLOP/s in late 2012. That's about half of the lowest-end mobile GPU that AMD had to offer since the beginning of that year.
So half of the lowest-end mobile GPU AMD has to offer right now would be... ~330 GFLOPS
Probably the question of having a GPU with less or more than 1TFLOP/s has to do with it being a mobile console or not.
The NX console will be much more competitive hardware wise than WiiU was, the main reason it was so weak computationally was because the controller cost Nintendo ~$150 to manufacture at launch, they still sell Gamepad replacements in Japan for over $100, three years after launch. This left them with very little of the build costs for powerful hardware components, Iwata himself said if they were to make the console more powerful along with the Gamepad it would have been priced much higher. It's also the major reason the console has never had a significant price drop over three years.
1TFLOP is the absolute weakest the NX consoles GPU will be imo with closer to 2 or even 3TFLOPs (if Nintendo have decided to re-enter the arms race). 3TFLOP will be nothing in late 2016, there are already several PC GPU's that hit the 8TFLOP mark.
D
Deleted member 13524
Guest
Oh if they go FinFet with the same area they have for the Wii U and 80W at its disposal then it would probably beat the Durango SoC-
Remember Latte is only one of 3 chips in the MCM. There's also the 33mm^2 CPU and 3mm^2 thingie-that-no-one-knows-what-it's-for.
With an APU over 190mm^2 using FinFet they could probably fit more of everything.
As for ESRAM/EDRAM, with the GDDR5 currently available I simply don't think it'll happen unless they absolutely needed it for backwards compatibility. No matter how you look at it, Microsoft is still paying dearly for it.
I don't believe this for a second.
And you don't think it's because they're making a ridiculous profit from it, like all hardware makers do by selling replacement parts?
You don't have to believe it, Nintendo had to sell a first party game at launch to break even meaning they had to bring in $410 to cover their hardware costs and going on the power of the chips inside the console, the Gamepad controller was a large chunk of that $410.
Now that you mention it, Nintendo is in the position to skip compatibility altogether
Power (Watts) would be a valid reason for considering embedded memory in addition to backwards compatibility, but only for a certain lower band of performance where it allows you to use smaller external buses. For higher levels of performance I absolutely agree though. With the BW saving features that Polaris would have over PS4Bone, the 7 gHz 128-bit GDDR5 configuration would seem ideal for challenging XB1 ~ PS4 level performance at reduced cost.
With regards to the XB1, I think the embedded ram work out so badly because it's paired with the a terrible configuration of external memory. Fat DDR3 bus, lots of memory chips (now twice the number of the PS4!) that all need identical trace lengths, and low bandwidth. In a hypothetical XB1 where the APU had been paired with 128-bit GDDR5, you'd have seen reduced cost and complexity and kept the benefits of the esram while removing the biggest issue with the slow DDR3 (would probably have had less die area dedicated to memory IO too, so had a smaller chip). Obviously that would have meant only 4GB of memory, but most people aren't interested in the shit that takes up most of the other 4GB anyway.
They could probably emulate the WiiU on any remotely fast modern system, assume they were prepared to go down that road. Nintendo seem to like maintaining hardware BC for one generation and then ditching it.
I'm afraid that's not evidence of anything. A DualShock 3 controller is $50. A DualShock 4 controller with more tech is also $50! The tablet teardown doesn't look massively expensive relative to what mobile devices manage either - full fledged Android tablets can be got for $100.
Read my last post... $460 they had to make before they broke even on WiiU at launch. Android tablets also weren't $100 in 2012 and they most certainly didn't have chips in them that could receive a 60fps video signal from a console with extremely low latency.
Fourth Storm
Newcomer
While I'm sure the Gamepad factored into the cost more so than an average controller, I'd also think the processor(s) probably ended up costing them a bunch. Despite the relatively lean amount of silicon, you've got 3 different companies (Renesas, AMD, IBM) and the licensing fees involved with that. IBM gave them a custom chip using an SOI process, which adds to cost. You had the Renesas eDRAM and all the costs and custom process node to go along with that. Also, the Xbox 360 is sitting at $199 even now. The last generation of consoles simply weren't able to be reduced in price as much as previous generations.
I guess the whole question to me is if they plan to re-gussy Wii/WiiU tech for the 14th time.
If they dont, then yeah no theoretical reason they cant do something nice and tidy with the hardware.
Of course, unless it is some kind of hybrid mobile and that also hampers it...
- Status
Similar threads
