From paste bin
Durango Final Specifications Development (Reasons and Considerations):
The Original Target Specs which were issued in Feburary 2012 to developers were inline with the one u probably heard from Eurogamer etc.
Of course developers werent happy with them.
The 3 areas of complaint were:
1. Low Peak perfomance of GPU
2. Memory Bandwidth
3. System reservation
MS listened (mainly to Epic and EA) and increased the Compute Units (Remember mid-last year when some User on NeoGAF posted that the Team at AMD for Durango was bigger and that they considered it better than Orbis? Well there was truth to that, because they were upping the chip at that moment)
With the increase of the Compute Units the bandwidth became an even bigger problem for MS. At that time GDDR5 wasnt feasible for MS , so they decided (or in other words AMD recommended) to increase the bus width of the memory controller AND increase the amount of memory at the same time. The resulting 384-bit memory controller was a custom created one, simply because DDR3 MCs in normal consumer PCs only allow 64, 128 or 256-bit controllers.
After the increase in bandwidth and size of the main memory, they had to do something about the embedded SRAM on the die.(there wouldnt have been an apparent advantage in using eSRAM with the same bandwidth as the main memory pool)
This resulted in the "doubling of the eSRAM".
All those changes took place April/Mai 2012. In September 2012 the first sampling of the chip took place.(Remember the design of the system hasnt changed)
At the same time, the original APU (Specifications from Feb 2012) was being fabbed for usage in 3rd Party Development Kits. (Beta Kits V1 December 2012/January2013; Beta Kits V2 in March 2013 contained the final chip, though many developer experienced heating issues, so downclocking for final retail kits can be expected)
The yield rates of the chip have been worse than expected (personal note: This is such a Microsoft thing; last minute spec upgrade and then expecting it to have good yield rates...); as of April 2013 Microsoft still expects enough units for a global release at the end of the year, but expect shortages or even a push to early 2014.
Upgrades (in short):
1. GPU: 12 to 20 CUs; 16 to 32 ROPs; 48 to 80 TMUs
Peak Performance: 1,2 TFlops to 2,0 TFlops
Low Level Changes:
Register Read (GB/s): From 7372 to 12287
LDS Read: 1228 to 2047
L1 Read: 615 to 1025
L2 Read: 307 to 512
Die size (part of the APU for GPU): 145 mm2 to 215 mm2
2. Main Memory: Size from 8GB to 12GB; bus width from 256 to 384; bandwidht from 68GB/s to 102,4GB/s
3. eSRAM on die: size from 32MB to 64MB, bus width from 1024 to 2048, bandwidth from 102,4GB/s to 204,8 GB/s
(personal note: eSRAM is NOT 6T-SRAM; its 1 transistor per bit
Die size: from 18 mm2 to 40mm2 (estimates)
4. whole APU: die size from 300 mm2 to 400mm2/450mm2 ( personal note: this one is weird, but there were conflicting reports in the doumentation); TDP: 100 W to 160 W (personal note: those are power usage numbers for the APU alone; for comparision the Orbis APU wont draw more than 145; the efficiency of those parts is by the way much higher than those found in the 7000 Radeon series)