backtop


Print 65 comment(s) - last by saratoga.. on Apr 28 at 12:51 AM

June 6, 2006 didn't fit AMD's schedule

We just got official confirmation from motherboard and chipset manufacturers in Taiwan -- AMD has moved the official launch date of Athlon 64 DDR2 up two weeks to May 23, 2006. AMD roadmaps have previously put the AM2 launch at June 6, 2006 (during Computex 2006), but since motherboards and CPUs are already completed, the launch will be pushed up. AMD insiders tell us Conroe's launch date was also a factor in pushing the AM2 launch date up, though even we do not know the exact date Intel's Conroe will launch.

AMD's latest advisories claimed the following:
  • May 16, 2006: Global announcement of Energy Efficient Processor roadmap and pricing
  • May 23, 2006: Global announcement of Socket AM2 and new desktop product availability and pricing
  • May 31, 2006: Global announcement of AMD LIVE! desktop system availability


Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: reverse hyperthreading?
By niknot on 4/19/2006 6:39:09 PM , Rating: 5
SMT refers to presenting the functional units of a core as separate logical processors. This allocation of execution resources can be done concurrently to compensate for a lack of ILP in the code being scheduled or sequentially to hide stalls caused by dependencies. There is no contradiction in providing SMT to a single-core processor. That however wouldn't be "reverse hyperthreading," it would be SMT.

Making a parallel execution engine appear to the software being run as a sequential execution engine isn't in of itself novel. That is what superscalar cores do already. A lot of hardware is added to decode, re-order, and speculatively execute instructions in parallel and present state that appears to be modified sequentially. As everyone knows--and this is the motivation of SMT in the first place--in certain types of code there is little ILP that can be obtained and functional units are left underutilized. Working with a simulator and sample code an engineer can model the effects of issue-width on the number of cycles necessary for executing code. When performing this modeling the engineer will find that while the size of the circuitry grows quadratically with issue width, performance does not. For typical code, as the issue width tends toward infinity, the rate of improvement tens toward zero. Naively increasing the issue width thus won't solve any problems, because the code being executed doesn't contain sufficient ILP.

Taking two n-issue cores and naively adding the necessary scheduling and re-ordering hardware to make them appear as a single core will never provide the same performance of a 2n-issue core. You're reducing the space complexity of an actual 2n-issue device, at the cost of genuinely sharing the same resources. At the lowest-end you'll waste a lot of resources performing speculative execution that will be thrown away 97+% of the time. Somewhere in the middle you'll schedule work on both processors realizing that functional units on both are going to be underutilized because of limitations in scheduling and re-order possibilities caused by both cores sharing only some state. At the highest end you won't pretend to be a single processor but will instead present logical processors to the operating system where some of the functional units of one of the physical processors will be scheduled and the results re-ordered as if they belonged to the other physical processor. This way the functional units in the second processor are not wasted prematurely when there operating system has work for them to do. This is necessarily less optimal than SMT on a 2n-issue device, but the cost of producing the device is lower in design and manufacturing.

What AMD has in the future is a problem: their cores are probably going to trail Intel's by 15%, and they have only so many resources to devote to remaining competitive and designing future products. While enjoy a commanding performance and power dissipation lead for years, AMD has only enjoyed small increases in market share that could evaporate more rapidly than it was obtained if Intel suddenly starts just being better. They cannot financially afford that sort of problem.

What AMD has, is a multicore process that is working quite well for them. What they also have, is that a lot of the software people use isn't aggressively multithreaded. Even if they can't obtain 2x improvements in performance in these programs through scheduling and synchronizing work on two cores, perhaps they can obtain enough to remain competitive with Conroe on computationally-intensive tasks with a lot of ILP. If AMD's is doing anything but tossing feces on the wall in hopes of preventing post-IDF rumors from sabotaging their future, it's probably that.


"DailyTech is the best kept secret on the Internet." -- Larry Barber

Related Articles
AMD's Next-gen Socket AM2 Revealed
February 6, 2006, 6:42 PM













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki