Author: Tom Kerrigan
Date: 14:25:44 06/15/00
Go up one level in this thread
On June 15, 2000 at 16:15:51, Adrien Regimbald wrote: >Hello, > >>There are two interesting things going on with the Willy: >> >>1) Trace cache. This replaces the instruction cache. Instead of caching main >>memory, it stores decoded instructions in the order that they're run. The >>advantage is that instruction decode is eliminated in many performance-critical >>situations. The disadvantage is that a trace cache has to be several times >>bigger than a regular I cache to store a similar amount of code. Also, because >>instruction decode is often eliminated, the Willy will only have one decoder. So >>in cases where branch prediction isn't doing well, the decoder will be a >>bottleneck. > > >This is very interesting .. I have two possible perspectives on this - >1. Their branch prediction is phenomenal, and they are so confident about it >that they aren't concerned with mis-predictions so much. (I have seen some >incredible results for prediction using some very simple cache setups using a >very simple algorithm .. I can only imagine what Intel managed to achieve with >this new and powerful machinery, and the algorithms they made for it, devised by >some of the brightest minds in the industry!) They are using some new BP scheme that's supposed to be really great. I forgot the name of it though. >2. That this is simply a blunder by Intel - an extra decoder wouldn't cost too >much (I think? I admit I'm not up to speed on the prices of CPU components :P), >so really - why NOT just add a second one, just in case certain applications >can't be predicted well at all? Decoding x86 instructions is relatively expensive. The other thing is that if the Willy is forced to decode instructions, they're coming from L2 cache, which is probably more of a bottleneck, now that I think about it... >>2) Double-pumped ALU. The ALU is designed to run on the rising and falling edge >>of the clock, which means that it effectively runs at 2.8GHz while the rest of >>the chip runs at 1.4GHz. This is probably so the ALU frequency isn't bounded by >>the speed of addition (or some more complicated ALU op). > > >Could one interpret this as meaning that Intel fully believes that this chip >could reach an overall 2.8 GHz speed??? Not necessarily. It just means they designed the ALU to run twice as fast as everything else. >>One other thing to note is that the pipeline is now 20 stages, so any branch >>mispredictions will basically result in disaster. > > >Geeeeez! :P Most of the models that I've worked with (simplified in my hardware >classes just so that we didn't all go insane ;) were between 5 and 10 stages .. >I couldn't even imagine dealing with ** 20 ** !! Ahh well, it's a good thing my >interests lie more to the software aspect of things :) > > >>It'll be interesting to see if the Willy is significantly faster than the >>PIII/Athlon... > > >Yes, very interesting indeed! I suspect that it will be - I doubt Intel has any >interest whatsoever in advancing the field - Intel is only interested in making >money and LOADS of it, so I doubt very much that they would invest what is >obviously a substantial sum of money to make such a big advancement if they >didn't expect some MAJOR returns on it. Then again, nobody has ever really >declared that major businesses had much business sense, so only time will tell >:P Back in the day, Intel was at the leading edge of processor design and fab technology. Now they're more commercial. Another thing to ponder about Willy is that it was an underfunded, understaffed project that was more or less forgotten until recently. Intel expected everybody to be using IA64 by now, so there wouldn't be a need for another 32-bit processor. But now that IA64 is several years late, and the Athlon is eating up market share, it looks like Willy is going to have to save the entire company's ass. -Tom
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.