Originally posted by: Furen
There's nothing to improve in socket AM2 K8s. K8s are not bandwidth starved, they did quite well on DDR 400, doubling the bandwidth to 800 effective doesnt yield much of an improvement because there is no need for it. What AMD could do is to a) add more cores, or b) have more aggressive prefetching. Both of these options would require a significant redesign of the K8 and both of these will be part of K10/K8L/Barcelona (which, supposedly, will run on socket AM2, even if it isnt a K8).
Interesting you should mention that. AM2 parts don't need the bandwidth offered by DDR2-800 or higher, but they do enjoy considerable benefits from the low latency offered by high DDR2 speeds. It seems the DDR2 memory controller AMD released produces curiously low latencies and low memory bandwidth at any given memory speed, and it thrives when running RAM at high memory clocks even with a 2T command rate (1T doesn't help much, and more memory speed helps more than tighter timings).
An AM2 chip with DDR2-800 3-3-3-8 2T trumps an s939 chip at the same clock speed with the same size l2 running DDR400 2-2-2-5 1T. If you look at the high end, you can get AM2 chips theoretically running at DDR2-1300 or 1400 (depends on how much money you're willing to spend), and getting up to DDR2-1100 or so is doable if you buy the right RAM (this can be done for $200 or less) and motherboard.
Originally posted by: BitByBit
It seems the logical thing to do from AMD's perspective is to focus on increasing clock speeds, until K10. It isn't worth the R&D costs to make any architectural changes to the K8 this late in the game, but it's a shame AMD didn't update the core when it was logical to do so. A 5-10% increase in IPC would have softened the blow dealt by Core.
On the subject of cache, AMD seems to have focused on compensating for smaller caches by making substantial improvements to the memory controller on K10. Its ability to prefetch directly into the L1 cache, its write buffer and RAM prefetch logic all reduce the need for large caches. If K8 sees little to modest improvement from larger caches, expect K10 to see even less improvement, especially with the addition of its 32-way L3.
I agree. There's some question as to how far AMD could really push K8, but we'll probably never know. FX-62s and X2-6000+ chips got as high as around 3.3 ghz on good air and/or water, and phase change got some chips beyond 3.6 ghz on a few occasions (not without complications). However, the memory controller on AM2 chips is already outstanding and can go a long way towards eliminating the need for large, fast blocks of cache if the memory is run at a high enough speed. Had AMD push the X2 to 3.6 ghz and paired it with fast enthusiast-level DDR2, it would have been more than a match for even overclocked Core 2 chips. Core 2s have scaling problems due to their memory controller; this is especially true of the 2 meg l2 Core 2s.
Originally posted by: SickBeast
Memory wise the K8 keeps up with the Core 2 for the most part. I suppose it should be winning considering its integrated memory controller.
Actually, K8 is much better in the memory department, especially on DDR2 platforms. RAM latencies are much lower (measured in cycles at least). Intel compensates with large, fast l2 cache, which works, to a point. At the current stock speeds of AMD and Intel chips, it works very well for Intel. Still, it's amazing when you see a 3.5 ghz E6300 producing nearly identical SuperPi times to an E6700 at 2.66 ghz.
Facebook
Twitter