http://www.beyond3d.com/#news35566
Based on the RSX die shots released by websites following the PlayStation 3´s launch in Japan yesterday, we have successfully estimated the chip´s die size by comparing it to the nearby Samsung GDDR3 memory chips. The result is approximately 240mm² and, even with the imperfect precision of our measurement method, we can safely claim it´s between 230 and 245mm². That makes RSX roughly the same size as the CELL revision used in the PS3, which is known to measure 235mm², and it´s a fair bit bigger than NVIDIA´s PC equivalent to the RSX: the 90nm NVIDIA G71, which has a die size of 196mm². While TSMC´s 90nm transistor density might be slightly different from that of Sony´s fabs, it is unlikely to be a major factor here. The chip is also rumoured to run at 500MHz internally.
Based on our analysis and information, it seems that the RSX has 28 pixel shaders internally, with 24 active. G71, on the other hand, has only 24 and NVIDIA sells the defective parts as the 20 pixel shader GeForce 7900 GS. Such an opportunity for multiple SKUs based on the same die does not exist with consoles, and therefore higher redundancy is desirable. It´s also likely that other parts of the RSX are therefore disabled for yeilding, such as the vertex shaders and possibly the raster operations units (ROPs). This is similar to the strategy used by Sony with the CELL, where 1 of the 8 SPEs is disabled.
While the RSX might have some parts removed compared to the G71, such as its PureVideo technology, transistor count is likely balanced back by having more redundancy and larger caches: indeed, we believe that most of the RSX´s internal caches are larger compared to any G7x part, including the post-vertex transform cache and compression caches. In addition to its normal 48KiB of texture cache to its local memory, it also has an extra 96KiB of cache dedicated to communication with the XDR memory pool, in order to improve bandwidth utilization and average latency, and make it possible to use that memory pool for texturing operations - resulting in greater overall system bandwidth utilization.
Xbox 360´s Xenos, on the other hand, is a 180mm² chip (plus the daughter die, with the eDRAM and ROPs) and does not seem to have this level of redundant parts disabled for yield reasons. Instead of coarse-grained yield management (entire functional blocks duplicated), it is rumoured to be duplicating smaller elements in each block (likely 17 shader ALUs per 16 required working - some patents, including this one, hint at this), in order to minimize the chip´s die size while keeping yields in check. Both approaches are perfectly acceptable and industry standard nowadays.
Based on all of this information, RSX does not seem significantly more expensive to produce than Xenos, especially so when counting Xenos´ daughter die, which is estimated at more than 70mm². It should be taken into consideration, however, that the PlayStation 3´s memory costs are most likely higher than the Xbox 360´s, and that the lack of eDRAM on the RSX is a contributing factor to this difference, as this requires the system to have higher overall memory bandwidth.
The memory chips being "nearer" to the graphics core, as visible on the die shots that can be found on the web, will however not give any performance advantage, and while it may increase costs very slightly, it will help for spreading the memory´s heat. It should finally be noted that while the GDDR3 chips people have found next to RSX are rated at 1400MHz effective (700MHz internally), they will apparently only run at 1300MHz effective. This has most likely been done in order to also be able to use chips rated at 1333MHz effective by Samsung and possibly other manufacturers, a step which Sony might have deemed necessary to meet its long-term production goals and lifetime cost reductions.
ok maintenant à toi.
Je te prend effectivement pour un con , mais pas pour ces raisons là