Support NeoGAF

[Jburton]

So the Wii U GPU is even weaker than some of the cynics thought, I still don't see how Nintendo get away with charging so much for this subpar hardware.

That controller must really cost a lot.


p.s. Well done guys who put up the money, clearing up once and for all how low powered Nintendo went with this machine ....... also puts a major spanner in the works for those Nintendo fans that told us the gap between Wii U and PS4 and 720 would be very small.

 

[Fourth Storm]

Regarding the large bank of SRAM, I was reading up on the R700 ISA, and I found something interesting related to GPGPU.

Each thread of an R700 SIMD has access to 128 GPRs. However, not all of these are located on the cores themselves. Some are typically located in a set aside "scratch buffer" in main memory. Could this additional SRAM be there to provide more GPRs (or make up for any missing ones in the SIMD cores) so that no GPRs need be off chip?

blu could probably answer this best...

 

[majik13]

...and BLOPs, and Tekken, and Darksiders

yes, whats the point?, you cant play them HD on any console. its how the games were designed and then ported to WiiU. Tekken actually keeps HD res more often on WiiU.
But you are acting that sub HD is the norm and not the rare exception.

 

[Captain Smoker]

so how similar are the innards to the 3ds? i wonder if that's more or less what they were going for, given recent statements from the financial conference dealy.

I thinkt this will have an impact on the next Nintendo generation of systems.

They just opened the building.

 

[ozfunghi]

What's the equivalent number for the 360 gpu?

I see the 240 gflops thrown around but I don't see how it's calculated.

Zomle seems to be throwing a hissy fit out of disappointment.
WiiU is likely 352 Gflops according to the latest speculation on the picture.
X360 indeed is 240 Gflops.

ALU count * clockspeed * 2 /1000 = Glops
320 * 550 * 2 /1000 = 352

 

[Earendil]

They're on the way.

Holy cow! They are seriously going above and beyond.

I know it's been said a hundred times already, but thanks again to Chipworks for being so generous.

 

[Zoramon089]

Apparently you should be comparing it to a 360 rather then Durango and it still in some ways comes out weaker.

No. Unless the 240GFLOPS>350 now...

 

[bequietdrive]

Zomle seems to be throwing a hissy fit out of disappointment.
WiiU is likely 352 Gflops according to the latest speculation on the picture.
X360 indeed is 240 Gflops.

So just under 1.5 times as powerful a GPU, around what most people with realistic expectations were assuming.

 

[Blades64]

The CPU shots are on the way also?? Chipworks is AWESOME!

 

[Thraktor]

Regarding the large bank of SRAM, I was reading up on the R700 ISA, and I found something interesting related to GPGPU.

Each thread of an R700 SIMD has access to 128 GPRs. However, not all of these are located on the cores themselves. Some are typically located in a set aside "scratch buffer" in main memory. Could this additional SRAM be there to provide more GPRs (or make up for any missing ones in the SIMD cores) so that no GPRs need be off chip?

blu could probably answer this best...

I know the R700 series had a "global data share", but I assumed that that would be simply replaced by the eDRAM in Latte. There are also "local data shares", but I haven't heard anything of GPRs being offloaded to main memory.

Here's a block diagram of the R770, as a point of reference:

 

[jerd]

No. Unless the 240GFLOPS>350 now...

I think that the bolded 176 GFLOPS as the second comment of the thread is hurting the posting quality a little bit. I just got here though so I don't really know.

 

[LeleSocho]

Also, regarding the eDRAM, I had the craziest thought: why not just count the number of banks? It seems there are 2048 eDRAM banks in the main block, which may indicate a 2048 bit interface? (Although that contradicts what we currently know about Renesas's 40nm eDRAM).

Edit: For clarity, this would imply a 140.8GB/s interface, assuming it's operating at 500MHz.

You mean the little squares? on the "slow" EDRAM there are 256*64 squares on the "fast" one there are 256*16
don't ask me why i counted them

 

[Branduil]

So the Wii U GPU is even weaker than some of the cynics thought, I still don't see how Nintendo get away with charging so much for this subpar hardware.

That controller must really cost a lot.


p.s. Well done guys who put up the money, clearing up once and for all how low powered Nintendo went with this machine ....... also puts a major spanner in the works for those Nintendo fans that told us the gap between Wii U and PS4 and 720 would be very small.

How weak is it? Please go into detail on the remaining hardware since you are apparently an expert.

Were many people still claiming there would be a "very small" gap between Wii U and PS4/720?

 

[test_account]

Zomle seems to be throwing a hissy fit out of disappointment.
WiiU is likely 352 Gflops according to the latest speculation on the picture.
X360 indeed is 240 Gflops.

ALU count * clockspeed * 2 /1000 = Glops
320 * 550 * 2 /1000 = 352

Before we got this picture, what was the general assumption of how powerful (in Gflops) the WiiU GPU was?

 

[PaulLFC]

So the company basically gave away a $2500 photo? That is awesome of them.

 

[Zoramon089]

I think that the bolded 176 GFLOPS as the second comment of the thread is hurting the posting quality a little bit. I just got here though so I don't really know.

Yeah. The result of kneejerk reactions, especially when there was barely any info posted at that point in the thread

 

[ozfunghi]

I think that the bolded 176 GFLOPS as the second comment of the thread is hurting the posting quality a little bit. I just got here though so I don't really know.

Indeed. I have no idea why he made such a childish and extremely premature post. Likely disappointment, as he was a big Wii U "defender".

 

[Branduil]

Zomle seems to be throwing a hissy fit out of disappointment.
WiiU is likely 352 Gflops according to the latest speculation on the picture.
X360 indeed is 240 Gflops.

ALU count * clockspeed * 2 /1000 = Glops
320 * 550 * 2 /1000 = 352

Has anyone come up with a reasonable conjecture for what all the extra custom stuff on the chip is doing?

 

[btrboyev]

Before we got this picture, what was the general assumption of how powerful (in Gflops) the WiiU GPU was?

Around 600 or 650

 

[pieatorium]

Before we got this picture, what was the general assumption of how powerful (in Gflops) the WiiU GPU was?

400-550 i think was the most readily assumed?

 

[ozfunghi]

Before we got this picture, what was the general assumption of how powerful (in Gflops) the WiiU GPU was?

Anywhere from 352 to 528.

Meaning 320 to 480 ALU's.

 

[Thraktor]

You mean the little squares? on the "slow" EDRAM there are 256*64 squares on the "fast" one there are 256*16
don't ask me why i counted them

I'm talking about a line of 8 orange squares with one white-ish square at the end (interfacing with the bus). There are 2048 of these in the main eDRAM pool, with 16,384 little orange squares in total (each of which would be 2KB of eDRAM).

 

[guek]

so wait, is 352 now the newest consensus?

 

[Fourth Storm]

I know the R700 series had a "global data share", but I assumed that that would be simply replaced by the eDRAM in Latte. There are also "local data shares", but I haven't heard anything of GPRs being offloaded to main memory.

Yeah, it struck me as odd. Here's where I read, it though:

http://developer.amd.com/wordpress/media/2012/10/R700-Family_Instruction_Set_Architecture.pdf

 

[Jburton]

How weak is it? Please go into detail on the remaining hardware since you are apparently an expert.

Wow, way to get on like a dick.


Do you need to be an "expert" to post in this thread now?

Who is it up to? ... you?

 

[Glorified G]

So just under 1.5 times as powerful a GPU, around what most people with realistic expectations were assuming.

I was expecting under 400 GFLOPS. Not surprised.

 

[Raist]

Actually the saving from my old tv to this one is probably twice that amount every month

Are you seriously saying that your old tv cost you at least $20 per month in power consumption? Seriously? What's your monthly bill? $300?

 

[shinra-bansho]

Before we got this picture, what was the general assumption of how powerful (in Gflops) the WiiU GPU was?

Before we got this picture, people were still throwing around a 600 GFLOPS number and people seemed to have settled on 450+ GFLOPS. And I recall high EDRAM bandwidth expectations (multiple hundreds GB/s) comparative to the outcome.

With some proclaiming Durango being in the "middle" or closer to Wii U.

 

[Captain Smoker]

so wait, is 352 now the newest consensus?

Probably (also going by some comments @ B3D).

 

[cyberheater]

so wait, is 352 now the newest consensus?

It's the newest made up number.

We don't know about a lot of the stuff in the GPU and how it would aid performance.

 

[EloquentM]

Wow, way to get on like a dick.


Do you need to be an "expert" to post in this thread now?

Who is it up to? ... you?

Well, your post was pretty ignorant for someone who posted like they were so well informed.

 

[test_account]

Around 600 or 650

400-550 i think was the most readily assumed?

Anywhere from 352 to 528.

Meaning 320 to 480 ALU's.

Before we got this picture, people were still throwing around a 600 GFLOPS number and people seemed to have settled on 450+ GFLOPS. And I recall high EDRAM bandwidth expectations (multiple hundreds GB/s) comparative to the outcome.

With some proclaiming Durango being in the "middle" or closer to Wii U.

Thanks

 

[Durante]

Regarding the large bank of SRAM, I was reading up on the R700 ISA, and I found something interesting related to GPGPU.

Each thread of an R700 SIMD has access to 128 GPRs. However, not all of these are located on the cores themselves. Some are typically located in a set aside "scratch buffer" in main memory. Could this additional SRAM be there to provide more GPRs (or make up for any missing ones in the SIMD cores) so that no GPRs need be off chip?

This is an interesting theory, considering that the this behaviour (the "spilling" of registers) was often bandied about as a reason for the R700 architecture's disappointing GPGPU performance.

Before we got this picture, what was the general assumption of how powerful (in Gflops) the WiiU GPU was?

350-600 was the usual range.

 

[Branduil]

Wow, way to get on like a dick.


Do you need to be an "expert" to post in this thread now?

Who is it up to? ... you?

I'm genuinely curious since a lot of the chip is still not analyzed but you seem to have insight into exactly how powerful it is. If you have insight into something on the chip, please share it, don't keep us in the dark.

 

[Reallink]

yes, whats the point?, you cant play them HD on any console. its how the games were designed and then ported to WiiU. Tekken actually keeps HD res more often on WiiU.
But you are acting that sub HD is the norm and not the rare exception.

What's your point? You said, and I quote "the only thing sub HD on the system is VC content....." I provided examples of retail games. Seems to me if it can't even manage PS360 fare, the odds of pared down PS420 ports (assuming they're even possible at all) running at 720p are slim to none. Same will likely apply to a number of ground up Wii U games that make some effort to push the hardware. Doesn't really matter whether it's a norm or a rare exception, the fact that it happens at all is a shame.

 

[Jburton]

Indeed. I have no idea why he made such a childish and extremely premature post. Likely disappointment, as he was a big Wii U "defender".

Indeed he was and going by his reaction was enough to estimate how much weaker it is compared to the projected power the likes of he and others have been saying around here since it was unveiled.

 

[EloquentM]

Probably (also going by some comments @ B3D).

Link to thread and comments?

 

[Pie and Beans]

With some proclaiming Durango being in the "middle" or closer to Wii U.

I saw all that happening across a few threads the other day. I guess the crazies all felt the impending hammer of facts coming down, and gave it their all for one last nutty ride.

 

[Fourth Storm]

I'm talking about a line of 8 orange squares with one white-ish square at the end (interfacing with the bus). There are 2048 of these in the main eDRAM pool, with 16,384 little orange squares in total (each of which would be 2KB of eDRAM).

Hmmm, that's a damn good point, Thraktor. Meanwhile, if I counted correctly, there are 512 of them on the upper die - coincidentally the same bus width as Flipper's texture cache.

 

[AniHawk]

I saw all that happening across a few threads the other day. I guess the crazies all felt the impending hammer of facts coming down, and gave it their all for one last nutty ride.

could have also been some stealth durango bashing considering the comparison of specs in the orbis and durango threads.

 

[ozfunghi]

It's the newest made up number.

We don't know about a lot of the stuff in the GPU and how it would aid performance.

Right, because AlStrong has been known to pull numbers out of his ass.

 

[wsippel]

Also, regarding the eDRAM, I had the craziest thought: why not just count the number of banks? It seems there are 2048 eDRAM banks in the main block, which may indicate a 2048 bit interface? (Although that contradicts what we currently know about Renesas's 40nm eDRAM).

Edit: For clarity, this would imply a 140.8GB/s interface, assuming it's operating at 500MHz.

That's actually not crazy at all, looking at the macros. But I have my doubts. There are three off-the-shelf macros:

8MB, 256bit
1MB, 256bit
1MB, 128bit

And this is exactly where things get weird. There are apparently eight macros of 4MB each. That doesn't seem to make sense. 4 x 8MB would give you 32MB on a 1024bit bus. Why not use that? Why go with 8 x 4MB instead? It looks like each macro is connected via a 64bit bus in this particular case, so 8 * 64bit * 550MHz = 32.8GB/s. Not bad, but not really enough to "emulate" eFB and eTC. Maybe that's where the second eDRAM pool and the SRAM pool come into play?

 

[Jburton]

Well, your post was pretty ignorant for someone who posted like they were so well informed.

Who said I was well informed?

My opinions are based on the subpar performance of the Wii U graphically so far, a lack of technically impressive looking games on the horizon and now a chip that according to many on here falls short of the projected power levels they had been estimating for so long.

 

[test_account]

350-600 was the usual range.

Thanks

 

[cyberheater]

Right, because AlStrong has been known to pull numbers out of his ass.

He's looking at what he knows. There is custom silicon on that chip will isn't there doing nothing.

 

[Branduil]

could have also been some stealth durango bashing considering the comparison of specs in the orbis and durango threads.

Doesn't sound very stealthy to me.

So no one has any theories at all yet as to what the large custom sections of the chip are for?

 

[mantidor]

Around 600 or 650

400-550 i think was the most readily assumed?

Anywhere from 352 to 528.

Meaning 320 to 480 ALU's.

Gotta love these answers, right next to each other.

 

[blu]

Regarding the large bank of SRAM, I was reading up on the R700 ISA, and I found something interesting related to GPGPU.

Each thread of an R700 SIMD has access to 128 GPRs. However, not all of these are located on the cores themselves. Some are typically located in a set aside "scratch buffer" in main memory. Could this additional SRAM be there to provide more GPRs (or make up for any missing ones in the SIMD cores) so that no GPRs need be off chip?

blu could probably answer this best...

I don't know where you may have read that, but if I understand you right (which I may not, as it's past bedtime here), that's not true. Each shader stage (or was it a clause - don't remember ATM; anyhow, think of it as a subroutine) gets assigned a certain number of GP registers (up to a limit, as you note, of 128, IIRC) from a common pool, but it also gets constant regs and registers which carry over results from previous clauses/to following clauses. The registers of the active shader code could/should not be swapped out, or else the entire clause (across the entire SIMD wavefront!) will suffer badly.

 

[majik13]

What's your point? You said, and I quote "the only thing sub HD on the system is VC content....." I provided examples of retail games. Seems to me if it can't even manage PS360 fare, the odds of pared down PS420 ports (assuming they're even possible at all) running at 720p are slim to none. Same will likely apply to a number of ground up Wii U games that make some effort to push the hardware. Doesn't really matter whether it's a norm or a rare exception, the fact that it happens at all is a shame.

My original comment was worded poorly yes, it was meant to only specify the diffrence in HD content between the released consoles. it is common knowledge that some games are sub HD on all consoles, so I didnt think it needed to be explained. Yeah it is lame that they cant be HD on wiiU, but considering they were ports on brand new archeticture, most devs were focusing on just gettign it straight ported and figuring out game pad usage, and thats it. I dont think that is a shame or surprising that the games were not upscaled. Anyways, it is not the point.

Also how exactly is the WiiU not managing PS350 fare?


and as for your original comment

Nice, with Wii U you can feed your big new fancy HDTV beautiful sub-HD content and laugh all the way to the bank with your $10 in annual electricity savings.

what exactly are you feeding your HDTV that has so much more HD than WiiU?

my point is that your comment is troll worthy

 

[Branduil]

So, does this explain to a certain degree why 360/PS3 ports don't look great? Is this a similar situation to the GC/Wii and its TEV where you need to write custom code to get the best quality and ports using standard coding will be nothing special?