Wii U hardware discussion and investigation *rename

Status
Not open for further replies.
The CPU cache is too small,even in the recent GPUs.
Every WII game expecting a fully random access of 1 megs of texture memory with 4ns latency.And the same for the frame buffer,with 2 megs of memory within 4 ns.
The caches are not too small. If you don't have a hit in the L1, the data get fetched from the L2 or main memory (or eDRAM if present). Because of the latency hiding mechanisms in recent GPUs, this rarely induces significant stalls. You can live with way higher latencies without losing performance. How often do we have to tell you that?
Performance isn't the problem. Nintendo has to find a viable solution for intercepting the GPU commands and translate it into something the new GPU understands and leads to similar and correctly rendered visuals.
 
Xenos eDRAM is 256 GB/s. I doubt Wii U will be struggling for BW.

:???: 2ns latency is faster than 4ns latency. It's also immaterial. The reason you need emulating hardware to be faster than what it's emulating is to turn instructions for different hardware into equivalent instructions on the emulating hardware. When it comes to data access, you only need the emulating hardware to have access to the data to work on at a suitable speed. That data moving system can be completely different architecturally.

That often quoted number of 256 GB/s is deceptive. Its really only 32GB/s, you only get 256 GB/s with the ROPs on the eDRAM die.
 
That often quoted number of 256 GB/s is deceptive. Its really only 32GB/s, you only get 256 GB/s with the ROPs on the eDRAM die.
How much bandwidth between shaders and ROPs do you need to completely saturate the theoretical throughput of 16 ROPs running at 500 MHz? ;)
 
That often quoted number of 256 GB/s is deceptive. Its really only 32GB/s, you only get 256 GB/s with the ROPs on the eDRAM die.
You're right, that's not really the R/W rate to eDRAM. My bad. I wasn't thinking through bomlat's argument properly (don't really understand it!) and took the internal BW figure. But then Wii doesn't have 125 GB/s bandwidth either. The BW from the caches is plenty enough to texture from.
 
That often quoted number of 256 GB/s is deceptive. Its really only 32GB/s, you only get 256 GB/s with the ROPs on the eDRAM die.
You "only" get it there because that's where it's most needed. Like PS2 only has what, 1.2GB/s graphics bandwidth? ...Because that's the capacity of the connection between the graphics processor and the main processor, right?

...Nevermind that on-chip there's a 1kbit wide read port, and a 1kbit wide write port, and another read port half that width for texture reads, all running at 150MHz. ...So yeah, only 1.2GB/s. No more than that! ;)
 
PCPer did a teardown. Great stuff... they actually read off the RAM chip (SMRT!). There's only 4 chips.

Samsung K4W4G1646B -> DDR3 4Gbit, 1.5V, 800/933/1066 speed bins -> I'm pretty sure that means "DDR3-1600/1866/2133" data rate.

256Mx16 would imply 16-bit I/O per DRAM... So... 17GB/s at most for main memory bandwidth. They didn't mention the rest of the numbers on the DRAM, but they ought to have high res photos later.
 
Is there a weblink to that, or just something they haven't published yet? Google shows nothing, as does hitting the site directly. :(
 
PCPer did a teardown. Great stuff... they actually read off the RAM chip (SMRT!). There's only 4 chips.

Samsung K4W4G1646B -> DDR3 4Gbit, 1.5V, 800/933/1066 speed bins -> I'm pretty sure that means "DDR3-1600/1866/2133" data rate.

256Mx16 would imply 16-bit I/O per DRAM... So... 17GB/s at most for main memory bandwidth. They didn't mention the rest of the numbers on the DRAM, but they ought to have high res photos later.

I always suspected the main memory would be on a pretty narrow bus. 64bits would not surprise me.
 
PCPer did a teardown. Great stuff... they actually read off the RAM chip (SMRT!). There's only 4 chips.

Samsung K4W4G1646B -> DDR3 4Gbit, 1.5V, 800/933/1066 speed bins -> I'm pretty sure that means "DDR3-1600/1866/2133" data rate.

256Mx16 would imply 16-bit I/O per DRAM... So... 17GB/s at most for main memory bandwidth. They didn't mention the rest of the numbers on the DRAM, but they ought to have high res photos later.

Is this good, bad, average?
 
17GB/s is bordering on smartphone/tablet SoC bandwidth levels.
Of course, those don't have dedicated edram for the GPU, so it may not be an issue.
 
It's slower than the 360's main memory. It's slower than either of the PS3's memory buses.

17GB/s is bordering on smartphone/tablet SoC bandwidth levels.
Of course, those don't have dedicated edram for the GPU, so it may not be an issue.

Thanks. Doesn't sound too hot, but I guess the proof will be in the pudding once we see some optimised games that utilise the EDRAM and GPGPU functionality.
 
PCPer did a teardown. Great stuff... they actually read off the RAM chip (SMRT!). There's only 4 chips.

Samsung K4W4G1646B -> DDR3 4Gbit, 1.5V, 800/933/1066 speed bins -> I'm pretty sure that means "DDR3-1600/1866/2133" data rate.

256Mx16 would imply 16-bit I/O per DRAM... So... 17GB/s at most for main memory bandwidth. They didn't mention the rest of the numbers on the DRAM, but they ought to have high res photos later.

i don't think nintendo use the max chip frequency (because more failure and heat).
probably use intermediate 1866mhz data rate = 15Go/s.
X360 have 50% more bandwith (22.4Go/s). it's a bottleneck, WiiU game probably need max use of eDRAM for map and not just frame buffer. it will be difficult to use MSAA or 1080p...
 
Is this good, bad, average?
That's a relative measure. Compared to other devices, as mentioned already, it's lower performance. PS3 had 22 GB/s from its GDDR3. A lot depends on the eDRAM implementation. Total aggregate BW should be ~PS3 I think (taking 32 GB/s as per Xenos, but we really don't know). If they've included 'smart eDRAM' like Xenos (ROPS on board), the benefits could be much better.

So it's pretty uncertain so far. ;)
 
i don't think nintendo use the max chip frequency (because more failure and heat).
probably use intermediate 1866mhz data rate = 15Go/s.
X360 have 50% more bandwith (22.4Go/s). it's a bottleneck, WiiU game probably need max use of eDRAM for map and not just frame buffer. it will be difficult to use MSAA or 1080p...
Do you know what a peak BW for the eDRAM might be? As I say above, if the eDRAM is smart like XB360, there should be plenty of BW, but if the ROPs are working from the same BW, the platform will be about as BW limited as PS3.
 
i don't think nintendo use the max chip frequency (because more failure and heat).

Right, that's just an upper bound given that we don't know what the rest of the print actually says (to identify which speed bin in particular). It's just a ballpark figure for comparison or understanding of what the I/O & clocks mean.
 
A lot depends on the eDRAM implementation. Total aggregate BW should be ~PS3 I think (taking 32 GB/s as per Xenos, but we really don't know). If they've included 'smart eDRAM' like Xenos (ROPS on board), the benefits could be much better.

So it's pretty uncertain so far. ;)
I guess bandwidth from eDRAM will be higher. It's not a separate eDRAM chip somewhere (as with the XBOX360). It's on the same die as the whole GPU. If the eDRAM can be used in a general way (and not only for the framebuffer), it could prove pretty useful.
 
Status
Not open for further replies.
Back
Top