I think the jury is still very much out on that one, but all we can do is speculate, at this point.
Howso? The 0.18 micron part already runs circles around everything. The 0.13 micron part is due out mid next year and is faster with more cache... are you suggesting that it will magically be slower???
I believe you are wrong on that one. Our most recent team order went to Athlons after a comparison done a few months ago between a 1.47ghz 1700+ and a P4 1800. The Athlon was consistently 20% faster, so that team just ordered Athlon 2200+ boxes. They estimated that it would take 2.7ghz P4's to equal the performance of the Athlon 2200+. They do a lot of modeling, running primarily in Java. Could they have gotten a little better performance if they'd spent a lot of time optimizing their code for P4? I'm sure they could have. But they want to spend their time writing new code, not re-writing stable, de-bugged code for the purpose of making Intel's expensive, fussy, processors look good. With the Athlon, their old code runs fine, and they can spend their time working on improving the capabilities of their system.
You seem to have a somewhat naieve team with a very narrow set of computing goals. Differences between a P4 1.8GHz and a P4 2.53GHz include faster bus speed, 2X on-die cache, and other microarchitectural improvements. What exactly does your "team" do?
K8 is quite a bit more than that, and why do you think a direct connection to memory would be more difficult to validate than a segmented one that has to go through a chipset? You almost certainly have that one backwards.
Umm... no it really isn't "quite a bit more than that." Most of the performance boost is coming from integrating the memory controller. It's essentially still the K7 core with a couple of added registers, 64 bit addressibility, a couple of pipeline stages, and an integrated memory controller. Any change in memory standard requires re-taping out the entire processor.
Intel is getting a boost in performance from hyperthreading, especially when the computer is running multiple programs simultaneously. The K8 gets benefit from additional on-die integration. That will work fine for single threads, but won't help the performance of multiple simultaneous threads. Overall, I'm not convinced that the K8 is going to outperform the Prescott or high-end Northwood, especially when running multiple threads.
Errr.... Have you heard of something called "Microsoft Windows?" How about "Linux?"
Yes, I have... and it doesn't appear that 64-bit extensions will be added to Windows until ~2004... so exactly how is the K8 going to be a lot better than the now obsolete K7 again?
That was the theory that SUN, Alpha, etc used to ignore the progress being made by standard computer hardware as it "moved up the chain" to servers and workstations. One size fits all has been fitting very well, and seems to fit a little better each quarter.
Howso? The Suns, IBM's, PA-RISCs etc. still hold the lion's share of server revenue. Sure, that's only 10% MSS, but it's around 60% REVENUE share and holding...
Hyperthreading slows down standard desktop software by about 5%. Some server software shows a net gain.
Not across the board... newer versions of the software are showing some net gains as well. I think Intel sat on this for so long because the OS and software wasn't optimized for hyperthreading. Granted, the single thread benefits of hyperthreading aren't as big as the multi-thread benefits, but they're tangible. |
