Replies to post #26086 on Intel Corp (INTC)

[wbmw]

Re: It's a bit strange to see your post lacking any comment on AMD's quad core that's coming relatively soon on the heels of Woodcrest and won't have the bandwidth bottleneck issues of Kentsfield.

I'd hardly call 17GB/s vs. 21GB/s a bandwidth bottleneck. See my other post.

[alan81]

Been round and round on quad core...
However, if we look at the dual socket market we see Intel with a bandwidth of 2.1 GB/s per core, and AMD at 2.65 GB/s per core (assuming DDR2-667 registered is available, and not DDR2-800 registered). IF registered DDR2-800 does show up that moves AMD up to 3.2GB/s/core. This puts the bandwidth advantage for AMD between 25% and 50%. Note that we have been talking about a 3 to 5% performance improvement for 100% increase in bandwidth. I would assume the "bandwidth advantage" that AMD gets is less than 3% in performance. As a sanity check to this number, Intel reported than a "modern" cpu consumes about 1GB/s of bandwidth to run specint, putting an Intel quad core chip at about 2X the needed bandwidth for good specint performance. I suspect applications that are more like specfp will do better than the 3 to 5% number though.

There are several factors that will play a bigger role than this small bandwidth difference...
How effective are the offload engines Intel is putting into the chipsets?
Can Intel get the snoop filtering in the chipset to work well?
What are AMD plans for additional offload engines, large caches in quad core, or other cache coherency improvements?

That said, it brings us back to the whole debate on timing, process generation, power, and speed relative to quad core. We can redo that debate, but I doubt there is much more progress to be made on the subject.
I do think INTC can do the following:
2.2Ghz 65W quad core, late Q4/ early Q1 2007
3.0Ghz 130W quad core, same timeframe
2.6Ghz? 90W quad core, same timeframe
I am not sure what power/performance bins INTC will offer, and just suggest these as possiblities.
--Alan