Conroe close to be really busted

Remember I told you that Conroe is only good for those single threaded applications with working sets comparable to 4MB, so they can run from cache? Now, there is more proof. A Chinese site got hold of one of this Con E6300 chip (1.86 GHZ) and benchmarked it with BAPCo SysMark 2004.

BAPCo SysMark was created by Intel. In July 2002, AMD was allowed to join in its development. Today, BAPCo's members include INTEL, AMD, Microsoft, HP, DELL, CNet, ZDNet, Nvidia, ATI and others. SysMark is about doing a series of real activities, mimicking real computer usage. These tests have large memory requirement, as they chew on large documents, etc. Also, multiple applications may be running at the same time during a SysMark test. Intel had avoided using this benchmark in all the Intel controlled benchmark runs.

In SysMark 2004 test, the Con E6300 got 210 points. In comparison , an Athlon 64 X2 3800+ got 232 points (2/1.86 = 1.07, 232/210=1.1). In Business Winstone 2004, the Con E6300 got a score of 25.2, an Athlon 64 X2 3800+ got 26.2. In Content Creation Winstone 2004, the E6300 got 32.2, the Athlon X2 3800+ got 33.2 (note that WinStone is not a multi-tasking load, it just runs through more memory). Recall that the X2 3800+ has only 512KB L2 cache per core. Suddenly, the Conroe doesn't look so rosy at all. Intel priced this chip lower than X2 3800+ for a good reason.

No wonder Intel so far doesn't allow reviewers do their own benchmarks. The most recent tomshardware.com comparison of Con E6700 and FX62 was again done with Intel configured hardware and benchmark suites. When a lot of amateurs and script kiddies cheered Intel's IDF guerilla benchmarketing, I smelled something fishy from the very beginning. Later I pin pointed the origin of Conroe's scores.

Why is this so important? Intel's Israeli team has bragged about the 4 issue core and various minor hacks. The amateurs cheered about a 33% IPC improvement. They say 4/3 = 133%. But anyone with a good understanding of CPUs know this is impossible. The Conroe has 14 stages, 4 issue, this means 56 instructions in flight at any moment, the amount of dependency is inherent in normal code. AMD has long concluded that 3 issue is efficient and cost effective.

So, again, just as we have seen from Woodcrest benchmarks, from SysMark results, we see CORE2's IPC is not any higher than K8. CORE2 marchitecture is no miracle. Clockspeed and bandwith will determine Conroe and K8's relative performance.

I expect that once Conroe is released in the wild, it will be soundly fragged by AMD's offerings in most multi-tasking loads. Remember, the whole purpose of going multi-core is for multitasking/multithreading.

Of course, the AMD 4x4 technology will permanently pin Intel at half of AMD's performance.

PS: I am glad that the Chinese have started running benchmarks. What da heck. It's just installing and running programs. Any kid can do it.

PPS: Let's enjoy some music.

PPS: FX62 benchmark results can be found here.