Laird: AMD Bulldozer: The mystery of the missing transistors
What are 800 million transistors between friends? According to AMD, not a lot. The bizarre story here involves AMD’s revision of the transistor count for its all-new FX processors, the chips that power the supposedly revolutionary Bulldozer architecture.
At launch, AMD said the FX weighed in at an immense two billion transistors. As I explained on these very pages, that was an impressively large figure, but only served to underline what a train wreck the new architecture had turned out to be. AMD’s old six-core Phenoms consume less than a billion transistors a pop and aren’t far off the FX for performance.
In that context, two billion transistors for a pseudo eight-core design and dubious performance gains didn’t look like a great deal. Well, something odd has occurred. AMD has revised the transistor count for Bulldozer-derived chips. Chips that were once claimed to contain two billion transistors are now said to pack just 1.2.
Yup, that’s 800 million transistors gone walkies. Put another way (purely for dramatic effect), the original transistor count is fully 66 per cent higher than the number AMD now claims. All of which begs several questions. How could AMD have got the numbers so very wrong? Is there even any meaning to transistor count claims? And what does it mean for our understanding of the Bulldozer architecture?
Let’s start with the facts as we know them – if, that is, we can actually believe them. The new count puts Bulldozer and the FX processors at the bottom of the 32nm pile in terms of transistor density. That’s the number of transistors that are squeezed into a given area of silicon. Intel’s dual-core Sandy Bridge processors are in a similar ballpark, but everything else is much more dense.
To put Bulldozer into context, it’s now rated at 3.8 million transistors per mm2. AMD’s own Llano quad-core fusion processor tops the 32nm density list at 6.4 million transistors per mm2. Meanwhile, AMD quad-core Phenom II chips manage 2.8 million, despite being on the older 45nm process.
So the first thing we learn is that the boundaries between chip production processes are not quite as clear and defined as you might have thought them to be. The top of 45nm is awfully close to the bottom of 32nm. But that doesn’t address the question of how AMD could have got it so wrong.
AMD’s line is that the previous figure was simply a very early estimate, but surely the company that designed the thing should know how many transistors reside therein? If it doesn’t know that, what else doesn’t it know? After all, far from being analogue, a computer chip is the epitome of modern, digital engineering.
It turns out that transistor counts are always estimates. For starters, several different types of circuit and transistor reside in CPUs. That’s increasingly true as more and more functions move on-die. Today’s ‘fusion’ processors can contain everything from cores, to system I/O, memory controllers, graphics accelerators and more. And each transistor type has a slightly different footprint and eats up variable amounts of die space. Transistor density isn’t a simple function of area and transistor count.
But even more important, modern chips are typically a mix of hand tuned and machine designed circuits. The intriguing upshot is that nobody quite knows exactly what’s inside them. Chip design is now so complex, humans can no longer grasp the whole of it.
There are shades of Skynet here – of machines outgrowing their wetware masters. I don’t think we need to worry about Terminator overlords just yet. However, the idea that something so conceptually cold, precise and digital as a computer chip is also fundamentally mysterious and fascinating makes you start to wonder just how clueless we’ll be 50 years from now. Maybe then we’ll just have to welcome our new computer overlords.
In the meantime, AMD’s Bulldozer architecture is left looking less catastrophically inefficient. Problem is, it remains far too slow to put Intel under any pressure. And despite expectations of a performance-enhancing revision in the next six months or so, I can’t see that changing. As I’ve indicated before, Bulldozer looks like AMD’s last roll of the dice. It needed sixes. It rolled twos. And it’s unlikely to ever again compete with Intel for performance supremacy. AMD seemed to concede as much recently when it told analysts its main focus was no longer competing with Intel.
Exactly what AMD’s options now are is hard to say. A move to ARM hardly seems plausible given its status as one of the few companies with rights to manufacture x86 chips. At the same time – and despite its claims regarding the Bobcat core – it has nothing remotely suitable for smartphones and tablets. Your guess, then, is as good as mine.
Posted on Friday, January 20th, 2012 at 11:09 am under Articles, Columns. You can subscribe to comments. You can leave a comment, or trackback from your own site.
Tags: Intel, Issue 318, Jeremy Laird
