Hello all, Don has provided some interesting information about chips and multi-cores that provide background justification for our 2 year doubling of compute power. Thanks!
Jeff ------ Forwarded Message From: Don Dossa <[EMAIL PROTECTED]> Date: Thu, 28 Sep 2006 17:40:59 -0700 To: Jeffrey Kantor <[EMAIL PROTECTED]>, Tim Axelrod <[EMAIL PROTECTED]>, Ray Plante <[EMAIL PROTECTED]> Subject: projecting moves to multicore CPU chips Jeff, Tim, Ray, Here are some important points in the 2005 International Technology Semiconductor Roadmap. More complex uni-core processors will continue to be designed for a few more years, but competition from multicore designs will win out because 1. Multiple cores permit reuse of processor design and verification; typical CPU designs consume thousands of man-years and still results in chips with hundreds of bugs. 2. Provides for easy and flexible power management; if a core is idle, power it down. 3. Multiple cores counter the global interconnect problem of cross-chip propagation delay. The interconnects are local to a core and have no significant impact. CPU maximum clocks has doubled every semiconductor technology generation (this refers to a major transition in mfg technology). There are 2 components to this speedup. A factor of 1.4x comes from device size scaling in the new technology. Another 1.4x comes from reducing the number of logic stages in a pipeline. A standard measurement of logic stages is the fanout of 4 (FO4) inverter delays. In the 180 nm technology, 32 FO4 were possible. In 130 nm, this was reduces to 24 FO4. This cant continue because properly shaped clock pulses cannot be generated below 6-8 FO4 delays. Look at AMD for examples of different manufacturing technologies, all of which are referred to as 90 nm SOI. The Athlon 64 runs at 2.4 GHz. The AMD Athlon 64 dual core runs at 2.6 GHz and has the same size L1 and L2 caches per core. The AMD Operton runs at 3.0 GHz while the equivalent dual core is a 2.4 GHz part. Investigation shows that the Athlon dual core runs faster than the single core because the 90 nm technology has been significantly improved over the Athlon introduction (it's called an E4 to F2 step). Compare to the single core Opteron, the dual core Operton is built using a only slightly improved manufacturing technology (E4 to E6 step). This slight manufacturing improvement does not overcome a less mature dual core design. Why do you care about the process steps? At product introduction, CPU die size is far too large to be profitable. With experience, the manufacturing process is improved, the die size shrinks, and so do the prices. The reuse of cores is a big win for us because the chip design and validation process is faster and cheaper and the vendors can take advantage of process step improvements much more quickly. An aside is to realize that there are 3 major versions of most CPU chips of the same name. They are the cost-performance versions for desktops, the high performance versions that we care about, and the power-aware versions for laptop/mobile devices. There are 7 versions of the AMD 2.4 GHz dual core Operton. Don't let people give you grief about chip prices if they are quoting desktop or laptop chips. Even though the name and speed may be the same, they are different parts reflecting different designs and manufacturing processes. Due to a combination of many factors, AMD will introduce a quad core chip in 2007 and an 8-way core in late 2008/early 2009. All of these previous comments serve to explain hows and whys of the chip design and manufacturing technologies. When you look at the AMD availability dates, you might conclude that the doubling time is under 2 years. This would be true if we were buying unpackaged die. By the time you get to the system vendors, they are able to keep to the approximately 2 year doubling time since they amortize the NRE for system software and integration over multiple product generations. Jeff - I don't recommend putting this into the DM chapter. Just hang on to it incase we get some questions. --Don ------ End of Forwarded Message _______________________________________________ LSST-data mailing list [email protected] http://www.lsstmail.org/mailman/listinfo/lsst-data
