270 replies to this topic

[3Dude]

routerbad, on 03 Apr 2013 - 07:49, said:Aha, it would be 
Ahh, it would be 17.6 PER MODULE though, each module is on a separate 88 bit channel width.
Funny thing, 17.4 * 4 modules is 69.6, which isn't far off from what we've been saying.


----------------------------------
across 11 lanes per chip.... hmmmm.... I want to see the underside of the wii u motherboard.

routerbad, on 03 Apr 2013 - 07:49, said:Aha, it would be 
Ahh, it would be 17.6 PER MODULE though, each module is on a separate 88 bit channel width.
Funny thing, 17.6 * 4 modules is 70.4, which isn't far off from what we've been saying.
Basically the 800MHz number we are using for the memory clock rate rather than 200MHz is effectively giving us the same number based on four chips.  So I think goodtwins was correct with using 200MHz as the memory clock rate (though strangely the Micron website lists it as an 800MHz memory clock, not IO clock).  Going by that, we have:
200*2*4*88/8=17,600Mb/s * 4 Modules = 70,400Mb/s  exact same number 


Yeah, the memory clock was doing me in the head too. I wonder if one of the other ram suppliers is correctly listed.

----------------------------------------------
ah the underside of the wii u motherboard.

http://guide-images....VPDtCY6ppv.huge

[Goodtwin]

Should'nt that Marcan guy be able to run benchmark test on the Wii U memory since he has hacked the Wii U? 

[routerbad]

routerbad, on 03 Apr 2013 - 07:49, said:Aha, it would be 
Ahh, it would be 17.6 PER MODULE though, each module is on a separate 88 bit channel width.
Funny thing, 17.4 * 4 modules is 69.6, which isn't far off from what we've been saying.


----------------------------------
across 11 lanes per chip.... hmmmm.... I want to see the underside of the wii u motherboard.

routerbad, on 03 Apr 2013 - 07:49, said:Aha, it would be 
Ahh, it would be 17.6 PER MODULE though, each module is on a separate 88 bit channel width.
Funny thing, 17.6 * 4 modules is 70.4, which isn't far off from what we've been saying.
Basically the 800MHz number we are using for the memory clock rate rather than 200MHz is effectively giving us the same number based on four chips.  So I think goodtwins was correct with using 200MHz as the memory clock rate (though strangely the Micron website lists it as an 800MHz memory clock, not IO clock).  Going by that, we have:
200*2*4*88/8=17,600Mb/s * 4 Modules = 70,400Mb/s  exact same number 


Yeah, the memory clock was doing me in the head too. I wonder if one of the other ram suppliers is correctly listed.

----------------------------------------------
ah the underside of the wii u motherboard.

http://guide-images....VPDtCY6ppv.huge

I think I may have gone overboard there, it wouldn't be an 88bit channel per chip, it would be 22 (44 per chip in dual channel mode) which would make the total bus 176.

 

If we go by that we have 800*2(data rate)*2(interfaces, or dual channel)*88 (bus width)/8 = 35,200Mb/s  I think that is the most reasonable number, and one you provided earlier.

 

Another way to look at it is 200*2*4*88/8 for 17.6 then double that for dual channel for the same number.

Edited by routerbad, 03 April 2013 - 02:16 PM.

[3Dude]

routerbad, on 03 Apr 2013 - 08:28, said:I think I may have gone overboard there, it wouldn't be an 88bit channel per chip, it would be 22 (44 per chip in dual channel mode) which would make the total bus 176.
If we go by that we have 800*2(data rate)*2(interfaces, or dual channel)*88 (bus width)/8 = 35,200Mb/s  I think that is the most reasonable number, and one you provided earlier.
Another way to look at it is 200*2*4*88/8 for 17.6 then double that for dual channel for the same number.


That makes my gut happy.

But these lanes... Somethins not right. Is their a memory controller inside those ram housings?

Edited by 3Dude, 03 April 2013 - 02:27 PM.

[routerbad]

routerbad, on 03 Apr 2013 - 08:28, said:I think I may have gone overboard there, it wouldn't be an 88bit channel per chip, it would be 22 (44 per chip in dual channel mode) which would make the total bus 176.
If we go by that we have 800*2(data rate)*2(interfaces, or dual channel)*88 (bus width)/8 = 35,200Mb/s  I think that is the most reasonable number, and one you provided earlier.
Another way to look at it is 200*2*4*88/8 for 17.6 then double that for dual channel for the same number.


That makes my gut happy.

But these lanes... Somethins not right. Is their a memory controller inside those ram housings?

Yeah 12.8 wasn't sitting well with me either.

Edited by routerbad, 03 April 2013 - 02:28 PM.

[Robotic Sunshine Commander]

This isn't bad. The power consumption is much less, but the console is much more powerful.

[3Dude]

Those pins on the gpu gddr3 io.... they should be 1 for 1 with the bits of the bus....



typical 64 bit gddr3 1GB ram card. 4 ram chips (removed you can see the grid ball array)



16 lanes per chip, adding up to 64 bit bus. Normal.

They go through the card, and are condensed into 16 thicker lanes, these lanes make their way back to the memory sockets,and are split back up to 64 pins, in 3,4 pin groups to be plugged into the mother board.

are our 88 lanes being split up to a 160 something pin layout on the gpu?
 

usually consoles lanes are wired like the individual ram chips, and not grouped like the traces leading to the pins....

I NEED to see how the ram connects to the gpu.

Edited by 3Dude, 03 April 2013 - 02:45 PM.

[routerbad]

Those pins on the gpu gddr3 io.... they should be 1 for 1 with the bits of the bus....



typical 64 bit gddr3 1GB ram card. 4 ram chips (removed you can see the grid ball array)



16 lanes per chip, adding up to 64 bit bus. Normal.

They go through the card, and are condensed into 16 thicker lanes, these lanes make their way back to the memory sockets,and are split back up to 64 pins, in 3,4 pin groups to be plugged into the mother board.

are our 88 lanes being split up to a 160 something pin layout on the gpu?
 

usually consoles lanes are wired like the individual ram chips, and not grouped like the traces leading to the pins....

I NEED to see how the ram connects to the gpu.

Just did another count, 158 pins.  79 per side.  Unless I'm counting some of them wrong, a few of them are hard to make out.

[Goodtwin]

If anything wouldnt the GPU having so many pins/lanes debunk the idea of 16bit right away?  The GPU is custom, so there would be no point in having more memory lanes at the GPU than required. 

[routerbad]

Apparently the hynix model used is DDR3 "graphics memory"  Same specs list though.

[3Dude]

Goodtwin, on 03 Apr 2013 - 09:11, said:If anything wouldnt the GPU having so many pins/lanes debunk the idea of 16bit right away?  The GPU is custom, so there would be no point in having more memory lanes at the GPU than required. 



------------------------
Exactly.


routerbad, on 03 Apr 2013 - 09:07, said:Just did another count, 158 pins.  79 per side.  Unless I'm counting some of them wrong, a few of them are hard to make out.


That gddr3 64bit card has 63 pins.

I think weve had a breakthrough. I thought those lanes were abnormally large.

158 pins... 158 bit bus????

(fixed, lmao)

800x2x158=252,800/8=31,600

31.6GB/s?

Edited by 3Dude, 03 April 2013 - 03:18 PM.

[routerbad]

If anything wouldnt the GPU having so many pins/lanes debunk the idea of 16bit right away?  The GPU is custom, so there would be no point in having more memory lanes at the GPU than required. 

The fact that it's DDR3 should have debunked it right away.  They applied a 16 bit width per chip which isn't the case in a dual die package.  They should have assumed 32 bit per chip or count the lanes, which is normal for graphics applications.

Goodtwin, on 03 Apr 2013 - 09:11, said:If anything wouldnt the GPU having so many pins/lanes debunk the idea of 16bit right away?  The GPU is custom, so there would be no point in having more memory lanes at the GPU than required. 



------------------------
Exactly.


routerbad, on 03 Apr 2013 - 09:07, said:Just did another count, 158 pins.  79 per side.  Unless I'm counting some of them wrong, a few of them are hard to make out.


That gddr3 64bit card has 63 pins.

I think weve had a breakthrough. I thought those lanes were abnormally large.

158 pins... 158 bit bus????

(fixed, lmao)

800x2x158=252,800/8=31,600

31.6GB/s?

I'm good with that.  What's going into the GPU is going to tell the whole story with regard to the bus width.

Interesting, maybe we're missing 1 pin per channel here.

Edited by routerbad, 03 April 2013 - 03:33 PM.

[3Dude]

routerbad, on 03 Apr 2013 - 09:46, said:The fact that it's DDR3 should have debunked it right away.  They applied a 16 bit width per chip which isn't the case in a dual die package.  They should have assumed 32 bit per chip or count the lanes, which is normal for graphics applications.

I'm good with that.  What's going into the GPU is going to tell the whole story with regard to the bus width.



I think i remember starting to go down that path. Then i got distracted by something and forgot what why i was doing it XD.

Whelp,i am pretty confident in this bandwidth. It was a bit of a weird layout, but thats nintendo for you.

Mystery solved gentlemen?

Edited by 3Dude, 03 April 2013 - 03:36 PM.

[routerbad]

routerbad, on 03 Apr 2013 - 09:46, said:The fact that it's DDR3 should have debunked it right away.  They applied a 16 bit width per chip which isn't the case in a dual die package.  They should have assumed 32 bit per chip or count the lanes, which is normal for graphics applications.

I'm good with that.  What's going into the GPU is going to tell the whole story with regard to the bus width.



I think i remember starting to go down that path. Then i got distracted by something and forgot what why i was doing it XD.

Whelp,i am pretty confident in this bandwidth. It was a bit of a weird layout, but thats nintendo for you.

Mystery solved gentlemen?

Solved.  Anandtech analysis debunked.

[Goodtwin]

Anandtech debunked for sure.  Too bad no reputable website will every go through the trouble that we all did here to try and make sense of things.  Seriously, if Anandtech is a hot shot tech website, wouldnt they have a contact within the ram industry to back up their findings?  I mean, when you are looking at something that doesnt seem quite right, and your supposed to be a professional journalist, you would think that you would at least try to contact the memory manufacture to get clarification on the issue.  They make no mention of contacting anyone.  They found the same model number decoding sheets that we have found, and took the X16 as being 16 bit.  Good work guys.   

[meitantei_conan]

So what does this all mean in the grand scheme of things?

[3Dude]

meitantei_conan, on 03 Apr 2013 - 11:08, said:So what does this all mean in the grand scheme of things?


Wii u's bandwidth is not half of ps360's, like the gaming media is circle jerking about, its much higher than ps360's.

It means more data (textures models, ai etc from wii u's gigabyte of ram can be moved out of ram to the gpu into the edram faster.

More textures, higher res textures, more effects, bigger wider open areas like X.

[meitantei_conan]

Wii u's bandwidth is not half of ps360's, like the gaming media is circle jerking about, its much higher than ps360's.

It means more data (textures models, ai etc from wii u's gigabyte of ram can be moved out of ram to the gpu into the edram faster.

More textures, higher res textures, more effects, bigger wider open areas like X.

I was suspecting as much , just needed to clarify.Finally some good news in this ever darkening sea of Wii U negativity!

[Alex Atkin UK]

It just makes sense all round. 

 

It always seemed rather unlikely that games like Sonic & All-Stars Racing Transformed could basically be the same game as running on PS360 but with better textures, if it didn't have increased RAM bandwidth to actually access those higher resolution textures.  The more detailed the textures, the more memory bandwidth you need.  Upon switching from the Wii U version to the Xbox 360 one, I'm sure I noticed a reduction in graphical effects on a few levels too.  Although I found the frame rate more constant on Xbox 360, likely down to the fact the game really was launched on too many platforms and as the Wii U was the easiest to hit the target frame rate, it seems they slacked off on a few sections where it takes a real nose dive (eg the NiGHTS level).

 

Then there is LEGO City Undercover which a lot of people seem to be complaining doesn't look impressive graphically.  I admit, my first impression was really bad, the frame rate was so low it initially gave me a headache and I had to enable the frame-doubling on my TV to compensate (but it adds horrible latency).  Over time I adjusted to it and started to appreciate some of its charm, especially as a first-gen title.

 

The texture work in particular is superb, not a single blocky texture in the whole game that I saw.  I'm fairly convinced its using tessellation as the hills and banks are really smooth compared to PS360.  Of course it has some bugs too and the draw distance/depth of field effect is something I hate.  But it certainly shows off things I just do not see could have been possible without greater memory bandwidth.

Edited by Alex Atkin UK, 03 April 2013 - 05:32 PM.

[routerbad]

It just makes sense all round. 

 

It always seemed rather unlikely that games like Sonic & All-Stars Racing Transformed could basically be the same game as running on PS360 but with better textures, if it didn't have increased RAM bandwidth to actually access those higher resolution textures.  The more detailed the textures, the more memory bandwidth you need.  Upon switching from the Wii U version to the Xbox 360 one, I'm sure I noticed a reduction in graphical effects on a few levels too.  Although I found the frame rate more constant on Xbox 360, likely down to the fact the game really was launched on too many platforms and as the Wii U was the easiest to hit the target frame rate, it seems they slacked off on a few sections where it takes a real nose dive (eg the NiGHTS level).

 

Then there is LEGO City Undercover which a lot of people seem to be complaining doesn't look impressive graphically.  I admit, my first impression was really bad, the frame rate was so low it initially gave me a headache and I had to enable the frame-doubling on my TV to compensate (but it adds horrible latency).  Over time I adjusted to it and started to appreciate some of its charm, especially as a first-gen title.

 

The texture work in particular is superb, not a single blocky texture in the whole game that I saw.  I'm fairly convinced its using tessellation as the hills and banks are really smooth compared to PS360.  Of course it has some bugs too and the draw distance/depth of field effect is something I hate.  But it certainly shows off things I just do not see could have been possible without greater memory bandwidth.

Well said, and I'm just about to jump back into Lego City.