I HAVE STOPPED guessing when the Phenom 9900 speed grade, the one we reviewed here recently, actually ships to real buyers - especially since it now has to be that new B3 stepping. The one that will be its last 65 nm revision.
What's more interesting is AMD's recent commitment to rush out the 45 nm Leo and Shanghai out really quickly - maybe as quickly as the end of the third quarter this year. Let's not talk about the server parts, still stuck in those HT2 sockets - the inputs from trusted Far Eastern friends about the 45 nm desktop parts say that, if AMD does its job right, we're gonna have a quad-core 6MB cache part running at 3.2 GHz with DDR2-1066 memory a few months before yearend. Hopefully, I'll add.
The added cache plus whatever improvements AMD squeezes into the CPU cores themselves should also give the 45 nm round some 10% extra per-clock performance.
So, let's be optimistic and assume we'll see those 45 nm 3.2 GHz Phenoms, say, this September. What will Intel have then, in the best and in the worst case?
Best case, of course, is Bloomfield Nehalem out - probably at not much more than 3.2 GHz, but a fast 4-core, 8-thread monster with native integrated memory controllers, QuickPath and 3-channel DDR3-1600 memory subsystem. Expect it to be at least as fast as a 4 GHz Yorkfield, and twice that in memory benchmarks.
Talk about Yorkfield, that is the "worst case". Assume that Intel doesn't move at all - they could still use the existing steppings to ship 3.6 GHz / FSB1600 (multiplier 9x) Yorkfields without any problems.
Let's give AMD a headstart then. How would a 3.2 GHz DDR2-1066 45 nm, 6 MB cache Phenom, AMD's best case this September, compare with Intel's 3.6 GHz FSB-1600 Yorkfield, their worst case that same month?
Based on the existing benchmarks, a 2.6 GHz DDR2-800 Phenom 9900 is, on average, 15%-20% behind 2.66 GHz FSB1333 Yorkfield Q9450. Assume that, on aver age, the above mentioned 3.2 GHz 45 nm AMD part would perform some 35% better than the Phenom 9900 - a generous allowance, shurely. Also, give the 3.6 GHz Intel, say, 30% extra over the Q9450.
The result, after a bit of rounding: Intel would still be some 10% - 15% faster, except for the memory benches. And that ignores completely at least two major steppings they'd do in such period of time on the Penryn dies - if the 65 nm history is any guide, we should expect at least a 10% direct frequency jump at the same voltage or even TDP in that process, with corresponding high-MHz yield increases.
If that is taken into account, we're then talking about 4 GHz high end " extreme" Yorkfield parts, and 3.2 - 3.6 GHz mainstream range, easily by September. And, all this completely ignores the Bloomfield, mind you.
In summary, the quick 45 nm delivery by AMD will definitely help them be more competitive in the mainstream - keep in mind the lower power consumption as a big plus. However, unless they can quickly ramp up the speeds towards some 3.5 GHz or so by yearend, I wouldn't call it a full comeback - yes they will "come" , but, performance wise, they will still look Intel in the "back". µ
Sign up for INQbot – a weekly roundup of the best from the INQ