Slashdot Mirror
AMD Announces 1GHz Athlon Imminent
Foxpaw writes, "Not to be outdone by Intel, according to ZDNet UK, AMD is planning on simultaneously releasing the 900MHz, 950MHz and 1GHz Athlons, maybe as early as later this month. " I do have to say that the corporate peeing match between these two is pretty amusing - but if Intel is still having production problems, then AMD can continue to leverage their high-end chips.
IBM is pretty sure 4Ghz is doable in the next two years by doing multiple independent clocks with async buffering between them. Slashdot ran something about that within the last month or so.
I don't know how fast CPUs can really get, but clock speed isn't the only factor. Work done per cycle is also important. If you look at today's (desktop/workstation) CPUs they all can execute more then one instruction per cycle. Two is a minimum. Three is the sustained max for Intel/AMD. Some RISC CPUs can do 5 instructions (UltraSPARC) if they are just the right set. This rarely happens, so most of hte time your two/three/five banger is executing zero (waiting for main memory), or one (too many read-after-write hazards) instruction. Transmeta is trying to fix this by making the (emulated) CPU really smart. The Alpha is trying to fix this by executing instrctions from more then one process at the same time (i.e. an ADD and MUL from your MP3 player, a LOAD and FMUL for your Quake server, and a STORE and a CMP for the re-compile of your Linux kernel). Compaq's simulations say this is really the shit. In two years I guess we will be able to see if it is. (the idea isn't new, the Terra supercomputer does something similar, and barrel processors have been doing something similar for decades).
As for 1Ghz being unthinkable four years ago, I assure you the design cycle for CPUs is long enough that the team that built the Athalon was almost certonally thinking of it. About eight years ago the Alpha archature manual was talking about it. Seven years I would have taken a bet that the Alpha would be the first "desktop" CPU to hit it. Apparently later this month I can be thankful nobody took me up on it. :-)
I have a 200Mhz machine at work, and a 266Mhz at work. Both do nicely for a lot of tasks. I bought a new machine to ray trace on (you could give me a million 1Ghz machines and I could still use more CPU to raytrace on...well maybe not me, but a real artest I guess). Neither do well for modern games, which seem to be a big driver for fast desktop CPUs. My C++ compiler could use a faster CPU too, which is why I don't do compiles on my 200Mhz box at work.
A fast desktop machine doesn't have the I/O of an oldish supercomputer. It does have a faster CPU. It's CPU is even competitave with modern supercomputers (sometimes it shares the CPU!). It's memory system and I/O is staggeringly far behind. Then again so is it's microscopic price tag.
So a fast desktop is good for anything that can get away with less memory, slower memory, and slower I/O, and just pain can't afford a supercomputer.
Since supercomputers are staggeringly expensave, I would guess lots of stuff "makes do" with desktop CPUs. Off the top of my head I would say: