Re: So you accept groo's prediction that K8L will be Q1'08, but dismiss his prediction that revG QC will have superior performance vs. Intel's QC...
Why would it?
Woodcrest is already outperforming Opteron by 30-50% on key benchmarks. Why would this change when both go to quad core, especially when memory bandwidth doesn't matter for a rather large class of server and desktop applications?
Re: 100% pro-Intel selectivity. Hmmm.
Yet, you're assuming that quad core Clovertown won't stand up to quad core K8, even though dual core Woodcrest enormously outperforms dual core K8. Sounds like a case of double standards to me.
Re: A double-die Intel QC part will be very expensive...
It will likely be cheaper than an AMD quad core equivalent. We already know that two small cores of size N yield better than a single large core of size 2N. And Clovertown will reuse the package and socket of Woodcrest, so those costs will be minimal as well. There's some extra assembly, but that is also minimal. So your claim of being expensive, in light of the cost of a quad core K8, is nothing but FUD.
Re: It will be competing with 4x4/2S/4S AMD machines and Intel's own 2S machines, hamstrung by a single FSB's bandwidth.
Look at where bandwidth matters today. SPECfp_rate and Stream seem to be the most bandwidth sensitive benchmarks out there today, and neither of these represent real application workloads. When you look to workstation apps, Woodcrest wins in nearly every single app, even the bandwidth sensitive ones. Quad core Opteron is likely to pull ahead in these relative to Clovertown, but what about the majority of server and desktop workloads? You're telling me that memory bandwidth is going to limit Clovertown and Kentsfield when it is nowhere even close to limiting Woodcrest and Conroe?
Re: 2x the cores will place 2x the demand on memory bandwidth.
Actually, it depends on the workload, and how well the other threads are utilized. In reality, it will place 1-2x more demand on bandwidth (and only 2x in benchmarks that are nothing more than memory accesses, such as the synthetic Stream test).
Re: We know (and you just stated) AM2 has tons of memory bandwidth overhead... Does CMW? Could it be that with QC, AM2 begins to utilize the overhead while CMW is already near saturation at DC and the FSB becomes a bottleneck at QC? (That is the scenario groo is predicting and it is not contradicted your argument.)
I know what Groo/Charlie is claiming, and sometimes he doesn't know what he's talking about. He may be good at gathering inside info, but his conclusions aren't always spot on. In this case, he needs to look at current analyses, which have been published from multiple review sites, and these show that Woodcrest and Conroe continue to perform and scale very well, even when limited by memory bandwidth. In other words, memory latency still plays a significant roll in many apps.
Re: QC is going to be used primarily in servers... And servers live or die on bandwidth.
Wrong. Workstations live or die by bandwidth, and I expect this is where Clovertown will be hampered the most (although I expect a boost in these apps with the 1333MT/s FSB Charlie also mentions). Server benchmarks - such as TPC-C, SPECjbb, SAP, SQL, etc - aren't as memory bandwidth sensitive as you think. Just look at how Woodcrest continues to boast 30-50% performance advantages over Opteron, in spite of being far behind in the memory bandwidth benchmarks. You think quad core K8 will close the gap completely on these? Fat chance.
