16.
@TA152H
>The article seems to miss some of the more salient characteristics of the Alpha. Mainly it was a clock speed whore, and at the beginning of its life had a horrible IPC. This conflicted with the high IPC processors from IBM and HP.
Was it bad to feature high clock speeds? I guess not. Though IBM POWER and POWER2 were "brainiacs" clearly, but they were far more expensive because consisted of 3 (POWER) to 8 (POWER2) chips within the same ceramical package. HP CPUs started to feature high IPC when 64-bit PA-8000 came out (4-issue core), but 32-bit PA-7xxx were no better (1- or 2-issue cores). EV6 could execute up to 6 instructions per clock (4 instructions dispatched by the decoder), EV5 -- 4, EV4 -- 2. Initial Alpha CPUs were limited by cache/memory bandwidth mostly (bus multiplexing).
>Everyone now like to romanticize about how incredible the Alpha was, except it never was. It was a horrid chip in a lot of ways; it ran very hot and was a huge processor that was extremely expensive. It never showed a big performance delta over contemporary chips for any period of time, they would leapfrog each other and each had an advantage depending upon what point in time you were referencing.
I'm afraid you exaggerate. Alpha CPUs were hot when compared to most other RISC CPUs, but not the leaders at all. Check the power consumption figures for PA-8000 or PA-8200, also POWER series. Different datasheets of modern Pentium 4 and Athlon CPUs are also informative very much. http://www.spec.org is a very good source to start talking about performance.
>DEC did not kill the Alpha, the P6 did. The Pentium Pro and its successors proved so powerful, inexpensive, and overall superior (particularly with regards to compatibility) it spelled the end to the dominance of RISC processors once it matured.
First of all, P6 series CPUs were "RISC inside". Second, their stack-based FPU was rather poor. The SSE engine of P3 allowed single-precision calculations only, and that was of no to very little use for anything beyond 3D rendering.
>Again, compare the modern processors with the Tualatin and you can see Intel and AMD have produced processors that offer much, much less performance per transistor than 1995 based P6 core. Sooner or later, it had to happen.
Check SPEC.org, and see that Tualatin lags in dust when compared to P4 or Athlon CPUs. No wonder, memory bandwidth is the key. Banias/Dothan are much better examples, though not the definite winners indeed. Modern technical processes allow to utilise large on-chip cache areas, that's why performance per transistor figures are much lower. Cache size increase is the last thing to do when improving performance.
>CISC chips were mainstream chips that were much smaller and less expensive than the RISC chips out at the time.
Wrong completely. RISC CPUs have much lower transistor counts. EV4 was of 1.68 mln., EV45 was of 2.85 mln. If to discount cache areas, that would make ~0.5 mln. for the core itself. Now Intel Pentium (P5) -- 3.1 mln., or ~2.0 mln. for the core itself. By the way, 64-bit vs. 32-bit.
>The failure of the Alpha was that it did not gain critical mass early, mainly because it was such a disagreeable machine and extremely expensive. They also did not attempt to make it popular outside of DEC, with only relatively minor companies offering Alpha based machines.
That's true. Though not extremely expensive, many vendors offered RISC workstations priced even more aggressively. For example, Silicon Graphics Power Indigo^2 vs. DEC 3000 Model 900 and Model 700.
>Of course, the processor was all wrong for this, particularly wasting all those transistors on 64-bits when it made little or no difference. How often does one need to do arithmetic on numbers greater than two billion anyway?
Not really. Some algorithms benefit very much from 64-bit calculations.
>The code density sucked too. And they generally had this super-expensive cache arrangement to compensate for this.
Something like that. Although on-chip cache was never large.
>More recently, Intel failed with Rambus memory (which I suddenly find myself in the need of, and this memory is expensive!) and of course very recently with the Itanium line.
Intel has failed with Itanium so far, agreed. More, Itanium2 and the primary supporting chipset (zx1) have been developed at HP mostly.
>Incidentally, I remember testing one of these machines, and ran so hot and was so expensive, no one liked it.
Would you specify the machine model?
>Another problem was it was made by DEC, and even if I, and others, had pushed the technology, management was not too crazy about buying from DEC at that time. They were perceived as having poor support and being beleaguered.
Unfortunately, it was. Bad for DEC.
>By the way, in the article it was mentioned that David Cutler went on to Microsoft to work on Windows NT in 1998. This is obviously false, as there was no Windows NT and no one was working on it at that time. Windows NT was originally called OS/2 NT, or OS/2 3.0. It was intended to be a platform unspecific version of OS/2, with a microkernel for the specific machine but otherwise very portable. When IBM got annoyed with Microsoft for being so successful with Windows 3.1 and WFW 3.11, they demanded they stop making new apps for this OS and devote themselves exclusively to OS/2. Microsoft demurred, and they split. IBM at the time had been working on OS/2 2.0 (x86 only) while Microsoft was working on OS/2 NT. So, Microsoft kept working on it, and it became Windows NT. An irony of all this is that IBM largely financed the development of it.
In 1988. There was no Windows NT brand those days, but the department was. Yes, Microsoft developed OS/2 3.0, and when relationships between IBM and Microsoft got worse, the latter released it as Windows NT in 1992.
[
Posted by: PVB
|
Date: 08/31/05 02:10:06 AM]
Make Home Page
Add to Favorites
XbitLabs is using Twitter
Windows Live Alerts
