First, there are a number of issues which are flat out contradictory to what happens in the real world. I think you were referring to the P4 when saying ?by giving it a shorter pipeline??well, they didn?t. They doubled the number of stages over the P3. This is completely contradictory by making a ?backwards step in processor technology?, as pipelining has been using extensively (in the x86 world) since the 486, and the pipeline has gotten longer every time to increase performance through higher frequencies.
However, you are right that fewer pipeline stages means less aggregate wattage to be dissipated. But, contrary to what you said, a shorter pipeline, it makes the chip less amenable to increases in clock speed. Clock for clock (both same bus, same cache, same frequency, etc), everything else the same, a chip with a shorter pipeline will outperform a chip with a longer pipeline. This is due to the branch mispredict penalties being smaller on the chip with the shorter pipeline.
I tend to agree that a bit of it is marketing. Higher clocked chips, in the past, tended to sell better (but historically, they also had better performance, so the impression almost makes sense).
Sure, Intel could have gone back to a 5 stage pipeline, reduced the branch mispredict quite a bit, required a simpler branch prediction algorithm with fewer entries, but then we?d be stuck at ~450mhz or so (this is just a guestimate ? the P55 made it >= 300mhz on the .25 micron). Don?t tell me that a 450Mhz Pentium with MMX is faster than a 1.7ghz P4!
You can go the reverse way, and instead of trying to go for high frequency, go for more IPC. Of course, this has its own problems. They burn up a lot of silicon in search of ILP that doesn?t always exist. You think that all the functional units on those sweet Athlon?s are always in use, even most of the time? They?re not. They?re not on a P3, and they?re not on a K6, K5, or even a Pentium MMX. Worst yet, the added complexity wickedly decreases the maximum frequency headroom. So going for the ?brainiac? (lots of functional units) hasn?t proven to be the best choice. Of course, a ?speedster? has it?s drawbacks too. It spends a lot of silicon (and power) on attaining high frequencies, often with large mispredict penalties.
The balance is somewhere in the middle. AMD claims that they are at the ?sweet spot,? but if you look at the power consumption, it?d show you they?re wrong too.
The Pentium 4 architecture will be around for years to come ? it has MANY forward looking technologies. Just because it doesn?t perform well today doesn?t mean that it won't perform far better in the future.
AMD spent a lot of silicon attempting to be compiler agnostic (well, as agnostic as a processor can be, which isn?t very), because they knew that few developers would compile just for them! They will (at least, that?s the hope) for Intel. They have in the past. The 486 with a short, single-issue pipeline performed better than a Pentium (in some cases, early on). The Ppro had terrible 16-bit performance: the Pentium trounced it! But for what the Ppro was intended for, who cares?
Okay, I?m done for now. If you still want to badmouth the P4, go ahead.
As for the rest of this thread, I?m going to say this:
Get a T-bird. It?s the best deal right now.
- what a great chipset that was...).
Facebook
Twitter