Patrick -
Merced is at least as real as the R12k
If that's true, I'll expect Merced to ship in systems within the next four months.
The most recent stuff I read said that it was expected to dissipate around the same amount of heat as a microwave, and Intel themselves has said that they don't anticipate being able to run it in a desktop machine. The Merced stuff that I've read also looks similar to HP's 'PlayDoh' thing that never really worked out. How many transistors is it supposed to have? It's gone up by a factor of about ten from the original PR. The spec numbers have gone up accordingly, as the schedule slips.
I'll respond to the rest of your Merced stuff on Monday, when I get back to work, where all my notes are.
I just checked out SGI's recent tpc-d announcement and found an 8-way 250 MHz R10K Origin system running Oracle 8 did significantly worse than a 4-way Pentium Pro system running NT and Oracle 8. So how are you faster again?
1. We're running 8.03. The only other 8.03 mark that I know of is Data General, whom we beat quite well.
2. Look at the Q1 & Q9 results. Some in the industry consider these
queries to best reflect DSS workloads and are the most BW demanding.
Compare your PC results against those! As I recall we are ~2X faster
than the best NT system on these two queries combined - again using
Oracle 8.03. Oracle 8.04 should make the situation even better.
3. It's our first TPC-D. We'll get the top position, but it will probably take some tuning, as TPC submissions usually do. It's also worth noting that the modular design of the Origin makes our TPC configuration much more useable in the real world than the TPC configs of most of our competitors whose machines require forklift upgrades to do anything more.
but I have a hard time remembering when I've seen such blatant crap from an SGI employee.
I'm the first to admit I don't know all the issues, and am not a CPU designer. You should note that I prefixed my comments with a disclaimer that my comments are just that - mine. I'm just calling them as I see them. I see merced as unworkable, as I currently understand it. I'd be interested to know if you've heard that it's got less than 50M transistors, how they're going to debug code written for a processor that reorders and parallelizes code even without optimization (executing more than 1 line of code at a time), what kind of form factor they're going to use in a desktop to dissipate the kind of heat that their microwave CPU is going to generate, how they expect to get proportional speed improvements when they're essentially wasting a large number of bits in each instruction, and why the projected speed has gradually increased as it's slipped.
-justinb |
