The processor wars Part I, “The death of Alpha”: Q&A with Nebojsa Novakovic
digitimes.com
Chris Hall, DigiTimes.com, Taipei [Monday 23 January 2006]
To the outside observer, improvements in PC architecture are evolutionary but logical. Processors advance inevitably in speed and performance, in happy accordance with Moore's Law. For Nebojsa Novakovic, a consultant in high-end computing systems, that's hardly the case. The demise of the DEC Alpha processor is a case in point. A performance leader was killed off by corporate whim.
Novakovic argues that despite the turn by AMD to a 64-bit platform, vindicated by the success of the Athlon 64, AMD had better watch out. Intel will be fighting back in 2006 with Merom, Conroe and Woodcrest, as processor design adopts a multicore approach. Even so, life won't be a bed of roses for Intel either. Rumors abound in the industry of the "retirement" of the Intel Itanium, an expensive attempt to break away from the x86 architecture at the high end, and Intel sorely lacks an interconnect that could compete with AMD's HyperTransport.
Novakovic comments on the major technology issues, in this in-depth interview, as Intel and AMD square off for the next rounds of the processor wars.
This is Part I of a five-part interview. Part II will follow on 24 January.
Q: What was originally the DEC Alpha processor has always been very highly regarded in the industry. Nevertheless, for one reason or another, the platform was essentially killed off, creating a new term in the IT lexicon, Alphacide. What comments and insights do you have on the death of Alpha? What remaining influence does Alpha have for CPU designers?
A: Since 1992, when the first Alpha chip, the 21064, officially saw the light of day, the Digital Equipment (DEC) design team that was in charge of it focused on sheer speed, in a way that is not dissimilar to the Intel Pentium 4. It had a very high clock rate, quite high power consumption also, but contrary to the x86 processors, a very elegant micro-architecture. It used a simple, fast instruction set that was easily programmable, with lots of high-end features. You would have had to wait until this year or next to see an equivalent feature-set available for the desktop.
Most importantly, this was an architecture that was in need of sales, marketing and business execution that would match its performance potential, but that type of business support was never put in place for the Alpha platform. At this time, in the early 1990s, there were many factors involved in the market positioning of Alpha, just as there were for HP's PA RISC and MIPS, which can be regarded as the other two leading RISC architectures of the time, in terms of performance and architectural elegance. Pressure from Intel was one factor. The sale of DEC to Compaq was another factor. If DEC had not been sold to Compaq, with Compaq subsequently being acquired by HP, maybe things would have turned out very differently.
However, the reality is that what I and some others in the industry call the murder of Alpha, in 2001, had nothing to do with deficiencies in the Alpha core or Alpha's lack of performance potential, as was claimed by certain executives at the time. In fact, even now, some of the Alpha machines that came out of HP before 2004, out-perform some of the best current Itanium or POWER chips in specific tasks and applications. This demonstrates that the architecture, despite its corporate demise, still retains its performance and functional leadership in many areas.
The other thing you need to bear in mind is that what transpired with Alpha was not like the fate of HP's PA RISC and MIPS; it was not canned overnight. I feel there was a strategic decision to kill off Alpha, but not to kill it immediately. The decision would have been to do this over several years, and the first step in the destruction of the Alpha architecture would then be to make it unavailable at the PC desktop and workstation levels. That strategy greatly reduced the number of available programmers who would otherwise have created software for the platform. That led to the elimination of the desktop user base, and that user base is a necessity if an architecture is going to survive. That loss of the user base is what happened to Alpha after 1999, and that is also what has been happening to the Intel Itanium, this past year.
The other thing you would do to destroy a platform is to get rid of the operating systems that would normally run on the system, and that's exactly what happened to Windows NT for Alpha in 1999. That version of Windows NT was killed off by Microsoft and Compaq in 1999, even though it was the most stable Windows ever, with built-in support for the AlphaBIOS, which effectively eliminated security vulnerabilities. For example, if you had tried to run spyware on an Alpha machine, under Windows NT, you most likely would have been unsuccessful.
So a plan was most probably created to kill off Alpha, but I cannot say whether this plan originated with DEC or with Compaq. Whatever the root cause, obviously we cannot place all the blame on Intel. The vendors who owned Alpha (remember the almighty Samsung and Mitsubishi were in the game too) had enough fire-power to figure out how to make the Alpha a viable computing platform in the market, if that's what they had really wanted to do. The EV8 (21464) core, that was killed just about as the design was completed in 2001, had 8 instructions per cycle sustained per single thread, and 4-way multithreading, something no other general-purpose CPU architecture even attempted – such a core could cover everything, from high-end PCs to supercomputers, easily providing the Far Eastern giants of the industry a good entry into those markets.
Alpha still has the potential to be used as a competitive computing platform. China, for example, would be able to adopt the platform, particularly since Eckhardt Pfeiffer gave the source code of Digital Unix to the Beijing government, plus related software, and this was one of the best Unix distributions of the time. Alpha has the same potential for adoption in Russia and Eastern Europe. In fact, it could be adopted anywhere where you are not dependent on software support from Microsoft.
And even if you admit that there is probably little or no chance of that kind of belated adoption, the Alpha EV9 core with its vector capability, which was co-designed by the Alpha team at the University of Barcelona and scheduled for release in 2005, could probably still be the performance leader, by a very wide margin. It would outperform the Intel Itanium by a wide margin, for instance.
So my basic point is that what happened with Alpha and the other two leading CPU architectures of the 1990s, MIPS and HP's PA RISC, is that basically three of the best RISC architectures available at that time are now gone. Today, the only general-purpose server processors with a similar target market are POWER and SPARC (the latter only for heavily multithreaded usage), while the AMD Opteron, basically, has become the mainstream processor for Sun Microsystems – as I predicted over three years ago.
Q: Does Alpha still have residual influence in terms of CPU design?
A: Yes, the Alpha architecture remains influential, and occasionally some of its features have been successfully integrated into other platforms. In the case of Intel, they took various bits and pieces from the Alpha design and might have tried to integrate them with the Itanium – its interconnect and so on – but they never succeeded. Alpha in fact has far more implications for AMD's HyperTransport, which essentially has roots in an Alpha EV7 interconnect, but one that has been made widely available in the industry. In the meantime, it's become the subject of jokes that AMD's Opteron processor is basically an x86 version of the Alpha architecture. We know, of course it was designed by almost the same team.
HyperTransport is very similar to the Alpha EV7 interconnect, but in its way, the EV7 interconnect was always far more advanced in terms of scalability. The EV7 interconnect integrated many more channels (4 interprocessor plus 1 dedicated I/O per CPU, compared to 3 mixed-use per CPU on current high-end Opterons), and it has other advantages. Nevertheless HyperTransport has now come of age, and it's no longer simply an AMD bus. You see it implemented in MIPS CPUs and Broadcom processors, and a few other places as well, including Intel machines. Intel uses it in their chipsets. So in that sense, HyperTransport has proven to be a very good board-level interconnect standard, for anything from a CPU interconnect to HyperTransport switching using Newisys' Horus chipsets, to very high performance I/O.
The other aspect is Intel's Pentium 4 processor. Essentially, while limited by all the quirks of the x86 platform, the Pentium 4 has a very similar approach to that of the earlier Alpha processor, in that it emphasizes sheer clock speed and delivers maximum megahertz. The essential difference is that with the Intel design you have problems with power consumption, and of course the Pentium 4 is limited by the x86 architecture and could never perform as well per clock. For a while, the Pentium 4 did manage to retain its performance advantage, but once it inevitably reached its limits, possibly the limits of the process design, but possibly also the limits of the x86 platform itself, the Pentium 4 lost ground to AMD's Athlon 64 and Opteron processor. This is what we've seen happening over the past six months.
In fact, there has been no significant performance boost to the Pentium 4, the kind of boost that would push it beyond 4GHz, over the past year. If you bought the 2003 3.2GHz P4, you won't have seen any noticeable performance increase – whether from a Pentium or an Athlon – over the past year and a half. That is the unfortunate reality. AMD's Athlon processors have managed greater improvement in their performance, but not by a really significant amount – by 20-30%.
And if you look at the overclocking capabilities, the current short-term manufacturing process advantage that Intel enjoys with its new 65-nanometer lines gives the newest batch of “Presler” and “Cedar Mill” Pentiums some pretty nasty unlocked performance – at fully stable operation if you are using good heatsink-and-fan sets. I ran my dual-core 3.46GHz Presler benchmarking unit at 4.26GHz dual-core, with low heat and full stability, for an hour, while the competing 2.6GHz AMD FX60 could only reach 2.8GHz, cooled in the same way, in stable operation. Over two weeks later, I’m still trying to tune the FX60 system to enable 3.02GHz operation.
This is Part I of a five-part interview. Part II will follow on 24 January.
Nebojsa Novakovic is a Singapore-based consultant for high-end computing as well as maglev transportation systems. He has been active in various projects throughout the Asia Pacific for over 10 years, the most recent being high-end technical computing clusters using top-end Intel and AMD platforms in combination with a Quadrics high-speed interconnect. His IT commentaries, covering high-end computing issues in particular, have appeared in numerous publications, including Singapore's The Straits Times, and he is a frequent contributor to the well known www.theinquirer.net website.
Photo: Nebojsa Novakovic |
