Clock speed of the CPU (three Broadways put together) at 1.25GHz and a GPU at 550 MHz. Does this give us further hints of what's achievable with the console?
Clock speed of the CPU (three Broadways put together) at 1.25GHz and a GPU at 550 MHz. Does this give us further hints of what's achievable with the console?
If the CPU is truly based on the 750 family, then we have no clue what evolutions the chip has gone through since the last known chip to be using that core, the Wii's CPU broadway (correct me if I'm wrong). IBM has stopped publishing specs to the public on 750 family years ago.
What we do know is that those chips were previously manufactured on 90nm fabrications. Do you think IBM and Nintendo were able to shrink it to 40nm for the WiiU?
Also, at 90nm, these CPUs took less than 3 watts to power. Could the WiiU CPU be closer to 1-2 watts now on a smaller fab? Insanity.
If we assumed that U-CPU is clock-for-clock compatible with the Broadway, using the Broadway/Ontario comparison from a couple of pages ago, we can conclude that for tight scalar FPU code (read: matrix multiplications sans SIMD) a U-CPU core is damn close to an AMD C-60 core:
$ echo "scale=4; 16.3517 / (1250 / 729) / 8.89305" | bc
1.0723
Kenka mentioned me a few pages back, so I might as well give my two cents.
First, it's worth keeping in mind that the general expectation until very recently was a CPU around 2GHz (many estimates around the 1.8GHz mark) and a GPU 500MHz or under (my guess was 480MHz).
The main take-home from the real clock speeds (higher clocked GPU than expected, lower clocked CPU than expected) is that the console is even more GPU-centric than expected. And, from the sheer die size difference between the CPU and GPU, we already knew it was going to be seriously GPU centric.
Basically, Nintendo's philosophy with the Wii U hardware is to have all Gflop-limited code (ie code which consists largely of raw computational grunt work, like physics) offloaded to the GPU, and keep the CPU dedicated to latency-limited code like AI. The reason for this is simply that GPUs offer much better Gflop per watt and Gflop per mm² characteristics, and when you've got a finite budget and thermal envelope, these things are important (even to MS and Sony, although their budgets and thermal envelopes may be much higher). With out-of-order execution, a short pipeline and a large cache the CPU should be well-suited to handling latency-limited code, and I wouldn't be surprised if it could actually handle pathfinding routines significantly better than Xenon or Cell (even with the much lower clock speed). Of course, if you were to try to run physics code on Wii U's CPU it would likely get trounced, but that's not how the console's designed to operate.
The thing is that, by all indications, MS and Sony's next consoles will operate on the same principle. The same factors of GPUs being better than CPUs at many tasks these days applies to them, and it looks like they'll combine Jaguar CPUs (which would be very similar to Wii U's CPU in performance, although clocked higher) with big beefy GPUs (obviously much more powerful than Wii U's).
Well, if you copy-pasted the $-headed line in a unix terminal (sans the $) you'd get the result given below the line. Re the numbers:Is there a "For Dummies" version of this post?
3m Hector Martin ‏@marcan42
No hardware threads. One per core. No new SIMD, just paired singles. But it's a saner core than the P4esque stuff in 360/PS3.
6m Hector Martin ‏@marcan42
The Espresso is an out of order design with a much shorter pipeline. It should win big on IPC on most code, but it has weak SIMD.
8m Hector Martin ‏@marcan42
It's worth noting that Espresso is *not* comparable clock per clock to a Xenon or a Cell. Think P4 vs. P3-derived Core series.
New Tweets from marcan
Well, if you copy-pasted the $-headed line in a unix terminal (sans the $) you'd get the result given below the line. Re the numbers:
16.3517 (count of seconds a certain sw test takes to execute on Broadway)
8.89305 (count of seconds the same test takes on C-60)
I think you can figure out the rest of the numbers ('scale=4' being of no interest ; )
/Layton hat off
Oh, and Thraktor - that was definitely a post worth bringing over - you should've posted it here first! ;p
Well, if you copy-pasted the $-headed line in a unix terminal (sans the $) you'd get the result given below the line. Re the numbers:
16.3517 (count of seconds a certain sw test takes to execute on Broadway)
8.89305 (count of seconds the same test takes on C-60)
I think you can figure out the rest of the numbers ('scale=4' being of no interest ; )
/Layton hat off
Oh, and Thraktor - that was definitely a post worth bringing over - you should've posted it here first! ;p
So GPU Gflops either 352 or 528 or something in between?
I guess we can rule out 704 Gflops... right?
So GPU Gflops either 352 or 528 or something in between?
I guess we can rule out 704 Gflops... right?
Might as well repost what I posted in the clock speed thread:
Plus, as an added bonus, some numbers on the eDRAM bandwidth. It's down to one of these three possibilities:
1024bit interface -- 68.75GB/s
4096bit interface -- 275GB/s
8192bit interface -- 550GB/s
Barring anything crazy (like a 28nm GPU or Renesas having developed some super-secret eDRAM just for Nintendo), it has to be one of those three numbers.
I'm guessing 528. Though I'll take 704 in a heartbeat.
Okay, (semi)good news: We now know the Wii U's CPU and GPU clock speeds.
bad news: we still don't know anything ELSE about the components(other than the fact that the CPU is based off of Broadway).
There's not really much we can say about Gflops from this. The R700 line normally divides the SIMD units (SPUs) into arrays of 80, but that's not a hard restriction of the architecture. The Wii U's could be divided into arrays of 60 or 100 or any number, really.
What speed is the 360 eDRAM ?.
Wouldn't that just make it unnecessarily exotic and difficult to program for? I would think they would keep the SIMD units at 40 or 80 unless there was some real benefit. More than, "Hmm. I want more than 320 shaders but 360 would be too much...how about 345!?"
Not really. Keep in mind that the R700 layout isn't designed so that developers can create code which is specifically optimised to it, it's designed to run general DirectX/OpenGL code that's written in a hardware-agnostic manner. So, if you were deriving a console chip from it (ie a chip where coders would be performing low-level optimisation specifically for the hardware), then the optimal configuration may be different.
So GPU Gflops either 352 or 528 or something in between?
I guess we can rule out 704 Gflops... right?
So again they made a machine not friendly to ports. Both this gen (gpu centric) and next (super slow). I guess we need to know ROPs etc now.Might as well repost what I posted in the clock speed thread:
Plus, as an added bonus, some numbers on the eDRAM bandwidth. It's down to one of these three possibilities:
1024bit interface -- 68.75GB/s
4096bit interface -- 275GB/s
8192bit interface -- 550GB/s
Barring anything crazy (like a 28nm GPU or Renesas having developed some super-secret eDRAM just for Nintendo), it has to be one of those three numbers.
So again they made a machine not friendly to ports. Both this gen (gpu centric) and next (super slow). I guess we need to know ROPs etc now.
I really didn't expect down near 1GHz. Nintendo just seem so paranoid about size/heat.
Regarding EDRAM bandwidth. Pick the smallest as that seems like Nintendos MO.
Erm, after today's revelations, I think we can safely cast off anything BG said as made up.If the CPU is really only using 1-5 Watts, that's a lot of Wattage left for the GPU, i think it's going to surprise people with how powerful it is.
Maybe BG's estimate of 600 GFLOPs wasn't far off afterall .
Problem with that approach is that latency would be higher as would be the power usage as data would have to fetched from further a way and from bigger memory pool. (future GPUs will most likely use 1KB L0 caches for this reason.)
Using eDRAM saves a lot in terms of power usage when compared for DDR3.
I definitely would rule this out, likely impossible judging by the die size.
I hope for at least 440 GFLOPS (400 shader units).
So again they made a machine not friendly to ports. Both this gen (gpu centric) and next (super slow). I guess we need to know ROPs etc now.
I really didn't expect down near 1GHz. Nintendo just seem so paranoid about size/heat.
Regarding EDRAM bandwidth. Pick the smallest as that seems like Nintendos MO.
Erm, after today's revelations, I think we can safely cast off anything BG said as made up.
If the CPU is really only using 1-5 Watts, that's a lot of Wattage left for the GPU, i think it's going to surprise people with how powerful it is.
Maybe BG's estimate of 600 GFLOPs wasn't far off afterall .
Erm, after today's revelations, I think we can safely cast off anything BG said as made up.
You obviously never read much of what he had to say...
2GB's of Ram, a slower than PS360 CPU and a GPU around the 600 GFLOP mark was what he predicted months ago.
A this gewgaws all about learning to use multicore chips efficiently. Will next gen be about how to use a limited amount of edram and a large amount of slow main ram efficiently?
My guts tell me that Nintendo went for the 1Mo interface.
It would still be an absolute improvement over the 360 but not by an order of magnitude. Maybe it wouldn't make sense to have a higher bandwidth given the other components? Sorry, I don't know what I am talking about.
Did we ever get semi-confirmed info on how many GFLOPS the Wii runs at?
Some of you guys really need to get the stick out of your asses when it comes to BG.Erm, after today's revelations, I think we can safely cast off anything BG said as made up.
Some of you guys really need to get the stick out of your asses when it comes to BG.
I've already said if you guys want to bitch at someone for not lowering the bar it should be me.
I had access to the same info as BG, and had a slightly more pessimistic outlook, but didn't post with the same degree of certainty.
This thread has been civil so far. I'd appreciate if it remained so in the near future. Thanks.Erm, after today's revelations, I think we can safely cast off anything BG said as made up.
You mean the GPU block diagrams? Well, I'm not a GPU designer, so I'm hardly qualified to tell you how wrong you are, particularly given that my personal expectations of that GPU and its access to its eDRAM are not far from yours ; )Edit: Actually, Blu, do you have any idea as to the feasibility of what I was posting above (regarding access to DDR3 and eDRAM from the various internal components of the GPU)? Or is it hard to put into words just how wrong I am?
Only thing listed on the wiki is that the Wii CPU is 2.9 GFLOPs.
Wwise 2012.2 CPU Load (lower is better)
| Wii U | A6-3500 | i7-920
------------------------------------------------
Peak Limiter | 0.20% | 0.08% | 0.05%
Delay | 0.07% | 0.04% | 0.01%
You mean the GPU block diagrams? Well, I'm not a GPU designer, so I'm hardly qualified to tell you how wrong you are, particularly given that my personal expectations of that GPU and its access to its eDRAM are not far from yours ; )
Wondering, is this test still fairly valid? http://www.neogaf.com/forum/showpost.php?p=41793263&postcount=1
Code:Wwise 2012.2 CPU Load (lower is better) | Wii U | A6-3500 | i7-920 ------------------------------------------------ Peak Limiter | 0.20% | 0.08% | 0.05% Delay | 0.07% | 0.04% | 0.01%
If so, given the clock speed of Wii U, doesn't that make it more efficient per clock than A6-3500, a triple core AMD CPU. (not to say anything about wattage)
I'm not as technically inclined as a lot of folks around here but from the sound of it, this is an expansion of the CPU architecture that originated with the GameCube.
That Power PC architecture was never able to attain the higher clock rates of X86 based architecture. It's one of the reasons that Apple ditched Power PC in favor of Intel and that Xbox 360 has had such issues with heat.
From a performance perspective even at much lower clock speeds (sometimes less than half IIRC) the Power PC benchmarked better than it's X86 counterpart. In other words a Power PC CPU clocked at 1.25GHz performed the same as an X86 CPU clocked at 3GHz.
Mind you this was several years ago when I remember seeing these benchmarks, but going strictly off the clock speeds of the CPU isn't a way to give a clear indication of what we can expect going forward out of the Wii U. It's already able to handle modern day ports and despite what anyone says about performance, you have to take dev time and the fact that it's a new architecture into the equation when looking at Wii U's launch line up.
So I'm not trying to defend Nintendo's lackluster stance on the graphics race - quite the contrary. What I am saying is that it's a bit premature to judge the graphical capabilities of the system based on specs alone. I honestly think the Wii U will remain a generation behind, but just like the Wii compared to Xbox and PS2, I think the Wii U will be able to deliver some really great feats graphically. In other words, don't worry so much about specs. Instead, try and speculate what can be done with them.
As a reminder
Anything made by Nintendo developers on Nintendo hardware will look LIGHTYEARS better than what 3rd-party devs can make. It's been proven generation after generation. Those .gif's are a moot-point.