From Tim Sweeny @
http://www.tech-report.com/etc/2001q1/sweeney-p4/
=============================================================================
The P4 is a good architecture, it's just not a good architecture for any of the current program executables on the market. It will take a while (a year or two perhaps) for Visual C++ to catch up, and for game and applications developers to recompile and ship new versions of their software that really take advantage of the P4.
I look at the P4 pragmatically. In terms of current applications, the P4 is basically equivalent to a P3 running at a few hundred megahertz lower clock rate. The P4 architecture, with its really long pipeline, doesn't seem to be designed for optimal instruction-per-second throughput -- the P3 probably beats it there -- it's designed to scale up to higher clock rates better than the P3. So, Intel's focus on the P4 is a good and necessary step for the industry, but don't expect much now. The big benefits to come in a couple years, when they get up into the many-gigahertz range.
The Unreal engine is pretty sensitive to both memory latency and memory bandwidth. The P4's Rambus focus is OK for this, but it's not optimal. I expect DDR will perform better for most of the high-performance 3D games on the market. The Xbox architecture (unified memory architecture, DDR memory) is really the model of what's ideal for us. However, with the P4's bus, Rambus is no longer the huge, glaring flaw as it was on Pentium 3 PC's equipped with RDRAM -- where the added memory latency made (more expensive) Rambus memory significantly slower than SDR memory for 3D applications.
SSE2 is an improvement over SSE, 3DNow, and Intel's past floating point architecture, since it contains a new, rapidly-accessible register set. However, these improvements will only show up in applications that are recompiled for SSE2, and it could be a year or more before good tools are available for this. With past architectural improvements like MMX, we went through and rewrote parts of our game in assembly language, taking advantage of the new instructions that way. Now, virtually no developers (even performance 3D game developers) use assembly language anymore, so our ability to take advantage of SSE2 is limited by compiler support. And virtually all developers are using Visual C++, and are limited by waiting on Microsoft for compiler improvements. (No, compiling with Intel's "benchmark" compiler isn't an option).
I think "Netburst" is a buzzword, not an architecture. Intel is very marketing savvy, and recognizes that users' biggest desire now isn't megahertz, but Internet performance. So, they have crafted this silly marketing campaign around promoting the P4 as improving your Internet experience. That might be good marketing, but it's lousy technology. The notion of a P4 improving your modem or DSL line's bandwidth is complete bullshit. As developers and gamers, we always laugh about that, and wonder (seriously) if Intel really thinks users are dumb enough to buy a P4 thinking it will give them better Internet performance.
In current apps, the 1.5 GHz P4 ends up being about equivalant to a 1.1 GHz Athlon. I think that's a fair comparison. Of course, both Intel and AMD are always releasing bogus benchmarks to show how their processor is much faster, but those tend to be very contrived situations. As a software developer, I also view it as unrealistic to run benchmarks using a new compiler designed just for those benchmarks -- this puts the emphasis on compiler tricks rather than real application performance.
The Intel/AMD benchmarking seems a lot to me like the Bush/Gore contest and all its cynicism: You know, "This country should count ALL the votes. Oh, just not those sitting over there in that pile. And you should count THESE votes a few more times just to be sure." The same thing is true of CPU benchmarking. Users don't run benchmarks; users run real applications, so you need to compare application performance. Not some special new applications recompiled with someone's trick compiler, but real applications on store shelves. And in that view, both Intel's and AMD's top-of-the-line offerings come out very close.
-Tim
==============================================================
I'll wait the year or two it takes for applications to actually take advantage of SSE2, then I might actually purchase a P4. But, a lot can happen in a couple years. Just in case you missed it, "In current apps, the 1.5 GHz P4 ends up being about equivalant to a 1.1 GHz Athlon." I don't see any reason to fork out another 400 to 500 bucks for the processor let alone the Rambus. I don't care what processor I'm using as long as I win. No one comes over to my house and looks in my case. I can assure you my wife is less than impressed with my processor selection to boot. If you want to believe that the p4 is the best thing going, good for you. Enjoy your processor.
Facebook
Twitter