Solution : dual-cores for everyone
The dual core should be 2005’s leitmotiv at Intel! During the IDF, Intel made demonstrations and intrigued us with the Montecito, Smithfield and Yonah. These names correspond to:- Montecito :
Itanium dual-cores- Smithfield :
Pentium 4 dual-cores- Yonah :
Pentium M dual-cores
The Montecito is a very ambitious CPU. The Montecito will raise the Itanium to the level of a technological marvel. Does it really happen in practice? It’s hard to tell. But on paper the Montecito is really impressive! The number of transistors reaches 1.72 billion. It’s useless to mention the yield, multi Gigahertz, die surface or fabrication costs.
Why so many transistors? Is it because the CPU has two cores? This is not exactly accurate, but partly true. Each core has its own cache L1, L2 and L3. These cores are rather impressive; 1256 Kb for the cache L2 and 12 MB for the cache L3 multiplied by 2. The Montecito integrates a total of 26.5 Mb of cache. With the redundant areas it represents more than 1.6 billion transistors. Compared to the Itanium 2, the new core benefits from several optimizations that Intel didn’t explain. The cache 2 has been split into Data Cache (256 Kb) and Instruction Cache (1024 Kb) like the cache 1. The Hyper Threading has been integrated into a slightly different version from the Pentium 4 in order to take into account the Itanium’s architecture. The Montecito has an advanced thread management system for optimal or at least not random use. Sometimes two threads will be processed by two cores and sometimes they will be processed faster on a single core via HyperThreading.
We witnessed an impressive demonstration using a system made of 4 Montecitos, but no information was revealed on the CPU frequencies. Intel announced, however, the CPU electrical consumption at around 100 watts, this despite the chip’s complexity. A power management system reduces voltage to below 100 Watts and therefore CPU frequency if the system overheats.
Intel put on display an avant-garde style desktop model with a dual-core CPU. Without clearly mentioning it, Intel insinuated that it was equipped with a slightly modified Pentium 4 dual-core Smithfield CPU.
In the demonstration opposing a Pentium 4, Pentium 4 HT and Pentium 4 dual-core in several video flows our interest was limited. Intel didn’t specify the computers’ details, so it was useless to speculate on performance.
This processor should be released in 2005 and reach respectable frequencies, equivalent to the current Pentium 4’s. Initially, it won’t be HyperThreading compatible. There are two reasons for this. Programs optimised for 4 process are still far from being available and the OS are completely shifted with the future CPU. More than two threads managed corresponds to a server for Microsoft.
For laptops this CPU belongs to the Napa platform (“Centrino 3”). It integrates 2 Dothan cores (corresponding to the last Pentium M) with 2 Mb of cache L2 and will be compatible with LaGrande Technology and Virtual Technology but not with the other T (no HyperThreading and no 64 bit support).
Electrical consumption will be higher than the current Pentium M but the power management system will be able to reduce it by deactivating one of the cores. This processor should be the first with a 65 nm fabrication process. We have attended a conference with Marc Bohr, in charge of the development of the manufacturing process. According to him the process is well advanced. The equipment used for the 90 nm has been designed to migrate easily to 65 nm. This new process should allow putting cache areas not used at a specific time in stand-by, a little like the current Pentium M.
During our meeting with AMD, we saw a demonstration of two servers equipped with 4 dual-core Opterons. It was impossible, however, to see them or know their frequencies. The frequency is still most likely below the Gigahertz level. With AMD yield problems, will it allow them to build high frequency dual-core CPUs? If the yield isn’t very good and the low number of CPUs able to work at high frequencies, the probability of having two high frequency dies is strongly reduced. This might be a problem for AMD if the 90 nm yield isn’t improved. AMD is planning on 1.6 GHz for this CPU release. This confirms that our ideas on the yield issue are probably true.
One of the two servers had a strange configuration. The screen displayed 8 CPUs, or 4 dual-core CPUs, but only 512 MB of memory were present and placed in the memory bank of a single processor. This configuration is suspicious and we doubt the real functioning of the computer with the indicated configuration… The other server didn’t have this anomaly.