As mentioned in Dan's July 5 update, the T-bird, seems to have problem scaling above 1.5 GHz. Not that the chips can't do it, per se, but that the performance in some benchmarks between the 1.4, 1.5 and 1.6 GHz T-bird is very close (closer than any previous models). Since P4 seems to be on a perfectly linear scale up to 1.8 GHz, I thought this might make in interesting discussion topic.
I will put forth a couple of my observations, then a couple of my theories:
1) T-bird 1.2 GHz was comparable in the BapCo benches, to 1.5 GHz P4
2) T-bird 1.4 GHz was comparable to 1.7 GHz P4, but not to 1.8 GHz P4
3) P4 is good in bandwidth intecive stuff, and bandwidth seems to be where faster T-birds are failing.
4) T-bird 1.377 (9.5 x 145) does match 1.8 GHz P4!
Conclusion 1) T-bird itself is a very powerful chip, compared to the P4.
Conclusion 2) T-bird is starving for bandwidth once again.
Now some observations on DDRSDRAM:
1) DDR is supposed to offer 2.1 GB/sec of bandwith.
2) RDRAM is supposed to offer 3.2 GB/sec in DC configs.
3) RDRAM really doubles DDR's bandwidth.
4) 760 MP has two memory paths, one for each CPU.
5) Athlon MP gets 160% of the memory bandwidth of regular Athlon.
Conclusion 1) DDR is not the culprit. Single Channel DDR on Tyan Thunder K7 with both paths used is 2/3 the speed of DC Rambus.
Conclusion 2) Athlon is not the culprit. Athlon with higher FSB performs better than Athlon with DDR 266 FSB. That means Athlon can do more.
Proposition: FSB is the culprit. Either the memory controller, the EV6 itself, or the datapath is constricting Athlon's performance.
This is particularly interesting in relation to the nForce chipset, which might help some of these problems work themselves out. The DASP unit is really mystical, since nothing of the kind has ever been employed in X86 cpus before. It seems to have the potential to help alot.
Palomino also has a prefetch unit. Will it work with the DASP, or overlap it? Will Palomino's prefetch help it scale more easily to match the performance of very high clockspeed P4s?
I think Palomino's greatest selling point is it's 25% lower heat dissapation. It should buy AMD about 400 MHz before they move to .13 micron production. It also seems to outperform T-bird by about 10%.
Will a 1.7 GHz Palomino match a 2.2 Ghz Northwood? If Northwood is mostly a cache-enhanced coreshrunk version of the Willy, it very well might. But it has a much better chance on an nForce chipset.
Now, If Athlon is starving for bandwidth, wha
|>> You actually had the answer
The answer isn't the FSB. It is, as you stated, the lack of hardware prefetch in the T-Bird Athlon.
Palomino Athlons should be able to truly take advantage of the bandwidth that DDR offers. Hopefully, it will scale as well otherwise the gap may eventually just become too large. Realistically though, it needs to do the job until Claw Hammer gets here for the desktop. That's where AMD's future most likely lies.
|>> I concur.
Another way to look at the Athlon, TBirds in particular, is that they are a new, rough-n-ready, piece of hardware; everything AMD needs to improve its market share and be taken seriously by corporate IT departments and please the users of AMD-based machines is there...its just that everythingneeds to "mature".
There is no lack of intelligence, skills and labour out in the world willing - and able - to do it, its something that takes time, experience and marketing; I dont know anyone who could accuse AMD of being marketing geniuses.
Sorry, marketing is the catalyst, and, obviously, not an essential skill in designing a processor and chipset