New truths about the Pentium M (AKA Banias) architecture

IF YOU MAKE A Pentium 4 desktop chip you've got to make a mobile version of that too, was the received wisdom inside the Intel Corporation, only for it to realise that perhaps the tail was wagging the dog.

Almost too late the Eureka moment arrived because chip technology is not necessarily a serial kind of thing and Transmeta seemed to have got a few things right.

That realisation led to the birth of Banias, an admission we extracted a bit like a crazed dentist pulling an aching molar -- from Intel over a year ago.

On the 1st of March last year, chief technology officer Pat Gelsinger admitted Intel stood the danger of falling behind Transmeta, when he told the INQUIRER "we were asleep at the wheel".

Now sources from deep inside Intel reveal just exactly what the corporation's thinking was about the new mobile project, then codenamed Banias but now full-blown as the Pentium M.

Said our source: "Most of the speculation until now has been completely off other than some stats like the clock rate, cache sizes, etc. The article I just saw and the link to the Anandtech site were starting to get more info, and I just can't resist any more as to why Intel has kept so quiet about this stuff".

He says the answer is "politics", something which we know from past experience is as common within Intel as spring daffodils blooming against the old city walls of York, Yorkshire.

Continued our source: "Many of us wondered why, if the P-4 wasn't shaping up, we couldn't just super-charge a P-III. The answer was, as much as anything, 'because we're doing the P-4'. Intel's internal projects develop a life of their own. Heck, look at the Itanium".

But Banias, he continued, delivered far better overall performance per watt than the Pentium 4M.

"Intel management didn't take that seriously until it became beyond clear that it needed to compete in the laptop and server blade market. Sigh."

The Banias (Pentium M), he added, is definitely based on the Pentium III. The principles used are the same as the Pentium 4 moving to the Prescott architecture, with Intel making a few microarchitectural changes to improve performance and adding performance tweaks to make the Pentium III a little bit faster.

Those were "flubbed" in the original design of the Pentium IIIM, but in any case gave far better IPC than the Pentium 4, anyroad.

Banias has a better branch predictor, for example, but Intel talking about an expanded branch prediction cache and how it's related to the L2/L1 data caches is mere "marketing babble". Bigger caches, of course, always improve performance because there are fewer misses.

Intel also introduced a bigger ROB, which in this case does not mean a higher price for the chip, but a bigger re-order buffer. The big obstacle to the Pentium III architecture was because of the large number of outstanding instructions "in flight" at once. The chip has to wait until the first instruction in its window is retired, that is to say committed, before it can move along there now, nothing to see. It's the ROB on the Pentium III architecture that decides this, but Banias doubled the number of entries and so helps with the number of outstanding instructions.

Mr Banias, now known as the Pentium M, also has more store buffers, which limit the number of outstanding writes "in flight" and not flushed into memory - these can persist well after an instruction is pensioned off and de-allocated from the ROB.

Further, Intel took a very close look at the power down and clocking logic, compared to the Pentium III, which didn't allow most sub-units to retire gracefully. That was a priority for the re-design in the clocking and circuit network.

The Pentium 4 has a pre-fetcher in the bus logic, which attempts to hide memory latencies by discovering access patterns and nailing them, and it seems Banias has one of these babies too.

This all helped supercharge the Pentium III architecture.

So now you know. The Pentium M is the Pentium III with Krypton added to make it Superpentium III.

But it's hardly building a "chip from the ground up", is it?

Like we said, we're very happy with the PIIIM 1.2GHz chip in the notebook we bought, and look forward to buying a Centrino in about 18 months time when it should be clocking 2.9GHz and more, coupled with a methanol fuel cell and the trillion wi-fi hotspots everywhere.

We must call the airlines later and see if they approve of machines that use methanol on board. µ

See Also
ATI gets licence to produce chipset for Intel Pentium M
Is the Pentium M a next-gen Pentium III