Help
Quote
AMD informed us over the weekend that it had erred when it described Bulldozer as an eight “core” / four module CPU with a total of two billion transistors. Bulldozer is apparently a 1.2 billion transistor CPU (the die size remains unchanged, at 315mm sq.)
This change, while notable, raises as many questions as it does answers. One of the puzzles of Bulldozer was what, exactly, the chip’s two billion transistors were doing. The chip’s density put it on par with Llano, but performance lagged far behind any Intel solution and barely matched AMD’s old CPUs. The consensus was that at least some of those transistors were byproducts of a fully automated design process with an unusually high degree of overhead.
If two billion transistors seemed awfully fat, however, 1.2 billion is pretty darn thin. Information AMD gave long before Bulldozer’s launch pointed towards 211M transistors per Bulldozer module (this includes the L2 cache). 2MB of L2 works out to 120 million transistors, which would leave 91M for each module. This means that CPU + cache alone would be 1.32B transistors, before we add I/O, integrated memory controller, or HyperTransport.
The only way to make AMD’s figures even provisionally fit is to assume the company shrank Bulldozer’s core from 91 million to 60 million transistors. Even then, we aren’t counting the other on-die components. When we contacted AMD for clarification of this point, the company insisted that it was simply correcting a mistake and clammed up thereafter with a “Move along, nothing to see here, folks” attitude.
AMD’s decision to replace a purportedly incorrect number with a figure that doesn’t much sense has given rise to rumors that the company is trying to gloss over Bulldozer’s problems by quoting the number of functional transistors as opposed to the actual number being laid down on the chip. There’s no proof that this is true, but the company could’ve avoided the problem by providing a bit more data and explaining how the new count matches up with previously disclosed information.
This change, while notable, raises as many questions as it does answers. One of the puzzles of Bulldozer was what, exactly, the chip’s two billion transistors were doing. The chip’s density put it on par with Llano, but performance lagged far behind any Intel solution and barely matched AMD’s old CPUs. The consensus was that at least some of those transistors were byproducts of a fully automated design process with an unusually high degree of overhead.
If two billion transistors seemed awfully fat, however, 1.2 billion is pretty darn thin. Information AMD gave long before Bulldozer’s launch pointed towards 211M transistors per Bulldozer module (this includes the L2 cache). 2MB of L2 works out to 120 million transistors, which would leave 91M for each module. This means that CPU + cache alone would be 1.32B transistors, before we add I/O, integrated memory controller, or HyperTransport.
The only way to make AMD’s figures even provisionally fit is to assume the company shrank Bulldozer’s core from 91 million to 60 million transistors. Even then, we aren’t counting the other on-die components. When we contacted AMD for clarification of this point, the company insisted that it was simply correcting a mistake and clammed up thereafter with a “Move along, nothing to see here, folks” attitude.
AMD’s decision to replace a purportedly incorrect number with a figure that doesn’t much sense has given rise to rumors that the company is trying to gloss over Bulldozer’s problems by quoting the number of functional transistors as opposed to the actual number being laid down on the chip. There’s no proof that this is true, but the company could’ve avoided the problem by providing a bit more data and explaining how the new count matches up with previously disclosed information.
http://www.extremete...onfusion-ensues
Haha, when in doubt, fix the specs on paper.
