Prediction update time. I think everyone is getting to the point where we're agreeing on the specs. This is just a writeup of what I think might be an explanation of some of the specs.
CPU: Based on lherre's comments, the devkit CPU is both of near final design as well as some weird assymetric unit. I'm speculating here that it will be built around one single full-fledged POWER7 core. This core will still be highly customized when compared to the server chip, but will share a similar amount of execution units to support a similar high performance as the original POWER7. This single core will be accompanied by simpler cores, I think either highly simplified POWER7 units. These cores will support the instruction set, but will be much simpler to reduce die size and power usage. All cores will support the VMX128 instruction set like the Xbox 360 CPU. All parts of the CPU will be barely be recognizable as sharing a similar base as the POWER7 server chip. One customization will be that L3 cache is scrapped for a larger L1 cache and a larger L2 cache using EDRAM. Customizations will allow a small die size and a high clock speed - rumours indicate the devkit is running at 3.2+ GHz. I think server POWER7 is a very bad indication of what power usage it'll have at that speed however.
Prediction: 1 core w/ 4-way SMT, ~2MB L2 EDRAM cache, 2 cores w/ 2-way SMT, ~1 MB L2 EDRAM cache. Cores share a common base with server POWER7, and will be clocked at ~3 GHz. ~30W TDP on a 45nm process, though 32nm shouldn't be out of the question (despite IBM's June press release).
GPU:The RV770LE sets a considerable number of constraints. The final GPU will feature 640 SPUs - if the final GPU would have 800 SPUs Nintendo should have used an RV770Pro - at 600 MHz clock speed. The final unit will probably be very similar in featureset as an RV740. That's probably only the case because the raw numbers make sense for a small efficient GPUs, and the GPU will feature a completely different design with several magic Nintendo tweaks we won't know the details of until some time in the Wii U's life. 32 MB of eDRAM like the one in the Xbox 360 is pretty much a given.
Prediction: 640 SPUs (probably VLIW5, VLIW4 is not ruled out yet I think), 32 TMUs, 16 ROPs, ~600 MHz clock speed, 32 MB eDRAM intelligent framebuffer. Competely new design although similar in featureset as an RV740/RV770LE. ~35W TDP on a 40nm process.
Memory: Assuming the devkit uses GDDR3 with the GPU having a 256-bit bus to that memory, we can only say that that's probably not final. The 256-bit bus does tell us that the memory in the devkit uses a number of memory chips that is a power of 2. This suggests that there won't be 1.5GB of memory - which also makes less sense for other reasons. If they end up using a 128-bit bus for the GPU (which is very likely) GDDR5 is an obvious choice, but I won't rule out XDR2 just yet because it may allow the usage of less memory chips and increased performance for lower power usage.
Prediction: 1GB unified. Could be anything.
Other stuff: One mystery is that lherre has indicated that the devkit indicates the final unit will have more than 1GB, however slightly that might be. 1.5GB is unified is a weird number that is problematic for other parts of the system design, and 2GB is probably too much. I think Nintendo will use a big chunk (at least 32MB, maybe even up to 256MB) of IBM EDRAM, 1T-SRAM or a similar type of low latency memory in the Wii U to achieve that. The reason I think they will do so is because the main memory it will have is high latency, even if it's GDDR3. A second reason is that this allows the Wii U design to be approached as some sort of superpowered Wii - a powerful single core OOE Power-based CPU (with optional additional extra cores), a small but quickly accessible chunk of primary memory, a large chunk of video memory and a GPU with super fast memory on board. I don't know whether this will be very advantageous, but I think it could definitely help speed up design for simpler games and avoid the headaches developers had with 360 and PS3.
Another thing we don't know nothing about is tesselation. I'm thinking it will definitely be there, but I'm wondering how much. I'm guessing that it should in principle be possible for the Wii U GPU will be at least as capable as a Barts or Southern Islands chip of similar specs.
bgassassin's got me convinced about the BC stuff. There'll probably be some ARM9/ARM11 chip doing various system stuff. I'm guessing total system TDP with CPU, GPU, motherboard, I/O chip and disk drive and other functionality will be around 80W.
CPU: Based on lherre's comments, the devkit CPU is both of near final design as well as some weird assymetric unit. I'm speculating here that it will be built around one single full-fledged POWER7 core. This core will still be highly customized when compared to the server chip, but will share a similar amount of execution units to support a similar high performance as the original POWER7. This single core will be accompanied by simpler cores, I think either highly simplified POWER7 units. These cores will support the instruction set, but will be much simpler to reduce die size and power usage. All cores will support the VMX128 instruction set like the Xbox 360 CPU. All parts of the CPU will be barely be recognizable as sharing a similar base as the POWER7 server chip. One customization will be that L3 cache is scrapped for a larger L1 cache and a larger L2 cache using EDRAM. Customizations will allow a small die size and a high clock speed - rumours indicate the devkit is running at 3.2+ GHz. I think server POWER7 is a very bad indication of what power usage it'll have at that speed however.
Prediction: 1 core w/ 4-way SMT, ~2MB L2 EDRAM cache, 2 cores w/ 2-way SMT, ~1 MB L2 EDRAM cache. Cores share a common base with server POWER7, and will be clocked at ~3 GHz. ~30W TDP on a 45nm process, though 32nm shouldn't be out of the question (despite IBM's June press release).
GPU:The RV770LE sets a considerable number of constraints. The final GPU will feature 640 SPUs - if the final GPU would have 800 SPUs Nintendo should have used an RV770Pro - at 600 MHz clock speed. The final unit will probably be very similar in featureset as an RV740. That's probably only the case because the raw numbers make sense for a small efficient GPUs, and the GPU will feature a completely different design with several magic Nintendo tweaks we won't know the details of until some time in the Wii U's life. 32 MB of eDRAM like the one in the Xbox 360 is pretty much a given.
Prediction: 640 SPUs (probably VLIW5, VLIW4 is not ruled out yet I think), 32 TMUs, 16 ROPs, ~600 MHz clock speed, 32 MB eDRAM intelligent framebuffer. Competely new design although similar in featureset as an RV740/RV770LE. ~35W TDP on a 40nm process.
Memory: Assuming the devkit uses GDDR3 with the GPU having a 256-bit bus to that memory, we can only say that that's probably not final. The 256-bit bus does tell us that the memory in the devkit uses a number of memory chips that is a power of 2. This suggests that there won't be 1.5GB of memory - which also makes less sense for other reasons. If they end up using a 128-bit bus for the GPU (which is very likely) GDDR5 is an obvious choice, but I won't rule out XDR2 just yet because it may allow the usage of less memory chips and increased performance for lower power usage.
Prediction: 1GB unified. Could be anything.
Other stuff: One mystery is that lherre has indicated that the devkit indicates the final unit will have more than 1GB, however slightly that might be. 1.5GB is unified is a weird number that is problematic for other parts of the system design, and 2GB is probably too much. I think Nintendo will use a big chunk (at least 32MB, maybe even up to 256MB) of IBM EDRAM, 1T-SRAM or a similar type of low latency memory in the Wii U to achieve that. The reason I think they will do so is because the main memory it will have is high latency, even if it's GDDR3. A second reason is that this allows the Wii U design to be approached as some sort of superpowered Wii - a powerful single core OOE Power-based CPU (with optional additional extra cores), a small but quickly accessible chunk of primary memory, a large chunk of video memory and a GPU with super fast memory on board. I don't know whether this will be very advantageous, but I think it could definitely help speed up design for simpler games and avoid the headaches developers had with 360 and PS3.
Another thing we don't know nothing about is tesselation. I'm thinking it will definitely be there, but I'm wondering how much. I'm guessing that it should in principle be possible for the Wii U GPU will be at least as capable as a Barts or Southern Islands chip of similar specs.
bgassassin's got me convinced about the BC stuff. There'll probably be some ARM9/ARM11 chip doing various system stuff. I'm guessing total system TDP with CPU, GPU, motherboard, I/O chip and disk drive and other functionality will be around 80W.
