Dear Combjelly:
At the time Mustang was cancelled, AMD specifically stated that OEM's were not interested in it (some probably were but, not enough to make it pay). They were not believed by many including on this thread. Now after the OEMs tried them out on the likely platform, 760MP, they decided that they were interested enough to ask for it where it would pay to do it.
With the current situation:
1) Tbirds now can go to 1.33G and show promise to go to at least 1.5G with current infrastructure.
2) P3 and its derivatives will not go beyond current Tbird speeds.
3) P4 performance not matching Tbirds at 1.2G, sell poorly, limited by RDRAM production, expensive to boot, and likely not to exceed Tbirds in either performance or value for at least 6 months.
4) No chipsets planned for next 3 months with large L3 embedded in sizes that would eliminate the need for larger L2s (read Mamba and similar chipsets).
5) No dual channel DDR chipsets within 3 months.
Large L2s are thus necessary to allow greater performance with existing chipsets (760MP) with server type loads. Although, if points 4 or 5 come to fruition, either could make the larger L2s unnecessary (an 8MB L3 of wide multiport embedded DRAM would more than dual DDR of PC2100, PC2600, or PC3200 in typical server (read TPC and like) and workstation (read SPEC, ScienceMark, and like)). However for the Hammer family, both L3 and dual DDR (QDR too) would be probable (the first because it reduces wasted silicon and the second because it allows greater memory sizes) in chipsets.
Opinions change with new information. Emperical results are more trusted than theorectical ones. It is easy to make chips that look good on paper. It is typical that unforseen and thus never looked at problems arise when making the paper chips real. The real world has this nasty habit of showing flaws in areas unconsidered in designs. A good example is computers that can talk to people with meaning and intelligence like HAL in 2001. It is always thought to be easy and seems to be always 20 years off. Even a subset of the problem, speech recognition is quite far from the ideal of converting speech by anyone who talks to printed word like any good stenographer. It is getting better but, not the trival task it was assumed long ago.
It is remarkable if, something is as easy as it appears at first glance. It is astonishing when it is easier.
Pete |
