Elmer, I just got back from Microprocessor Forum. Great show, but not as good as last year.
Re: Hypertransport scaling.
Not gonna happen in SPECint, which is a scalar benchmark. Only one thread will be active, so I don't expect any difference between the score of a Pentium 4 3.06GHz with Hyperthreading, and one without. I also think that Spokeshave's estimates are probably closer than yours.
Re: Opteron performance
Now having said what I said above, I think you bring up some interesting points. Let me just add some additional info from the Microprocessor forum. First, Tom Halfhill from MDR was the one that picked up on the Hammer cache size, since Fred Webber didn't want to disclose that. Apparently, AMD wanted to get away with throwing out an unlabeled SPEC score. An Intel employee asked about the cache size at the Q&A session, and Webber clammed up. It was then that Halfhill noted an earlier slip from Webber where he unintentionally disclosed the cache size. So thank Halfhill for that information.
As for any of your other assumptions, Spokeshave is right that it is not good to assume peak or base SPEC score, since that was not listed in the Presentation. Typically, most companies disclose base scores, and even though you might think AMD would disclose peak scores, it is a bad assumption to make the way you did with extrapolated benchmark results.
Back to Microprocessor Forum, I was able to see and touch a dual processor Hammer system that was being shown off by a Newisys employee. Apparently, they have 4P and 8P designs ready as well, but I don't know how quickly AMD will be able to supply the silicon for that. The DP design was in a 1U extended form factor, which is a great sell in the server market (kudos for Newisys). The thermals of the design will support up to 5GHz operation, according to what AMD has disclosed to Newisys as 5GHz thermal operation (note that there will be no 5GHz Hammer, but AMD is apparently planning ahead for future drop-in replacement cores).
I was able to get some interesting information about the demonstration. For one, these processors were running at 1.4GHz, and they were the newest samples from AMD. Functionally, the Newisys guy confirmed that they have been functionally very stable ever since first tapeout. They have run extensive validation without running into many issues at all. The system on display was actually compiling a Linux kernel, and system management windows alerted the user of fan outages from the chassis. Very impressive, even though Hammer was running so slow. One has to wonder if AMD will make 2GHz by their newly "revised" Q2 launch, or if they do, how much they will scale after that.
It's also interesting to note that the board was a 12-layer design, which is far more complex than AMD's previous claims of being able to realize 4-layer designs. The Newisys guy said that they might be able to get it down to 8-layers. It is worth noting that Opteron chips based on the Sledgehammer core have memory capacity capabilities of 4 DIMMs per channel, and Sledgehammer has 2 memory channels. Additionally, each DIMM can use up to 2GB technology without using double-stacked designs. Thus, the Newisys system can supported up to 32GB of memory. It is also worth considering a single processor system with the Clawhammer core, which uses only one memory channel. In that case, the system would only be able to support up to 8GB of memory. So much for all the hype of needing 64-bit addressing, since AMD will only make use of 33-bits on the desktop (note that Clawhammer and Sledgehammer only support 40-bit addressing, not the 64-bit number that sometimes gets tossed around).
Lastly, Re: 3400+ model number.
I don't know how AMD will classify their latest processor at in terms of model numbers. Right now, though, SPECint does not seem to be a consideration. The Athlon scores well below the Pentium 4 here, and yet AMD still uses 2800+ model numbers to describe their latest (unavailable) incarnation. So I believe that relative SPEC performance to be irrelevant to determining market- er - model numbers. I think it's safe to say, though, that AMD will have the upper hand in SPECint, provided they have 1MB of cache, and they can get to 2GHz. With 256KB or 512KB cache, one should expect a 5-10% lower SPEC score, probably bringing it below the score capable of a 3.06GHz Pentium 4. A lower frequency part may additionally do the same.
wbmw
