High End CPU Battle in 2012: Back to the Future
Intel, AMD and, at the very top, IBM, had the total performance and feature lead in high end general purpose microprocessors for a decade now. However, few 'strategic mistakes' that the US microprocessor and system vendors made a decade ago now haunt them back, as China and Japan embraced the superior technologies that the US discarded for short-term interests of the time.
By now, you are almost certainly used to hear the words Intel and AMD, and sometimes IBM, if talking about big machines, when it comes to high end microprocessors. These 10 past years, the general purpose X86 CPU leadership was with Intel for 7 years, and with AMD for 3 years. On the other hand, IBM firmly had the RISC – and overall – microprocessor performance lead per socket with its POWER 4, 5, 6 and Power7 CPU families, of course at the matching prices. Do not confuse these with much lower performance PowerPC CPUs used in PowerMac systems of the past.
Looking at the current roadmaps, that domination is expected to continue, with all the Intel Sandy Bridge, Ivy Bridge, Haswell and so on microarchitectures being followed by AMD Bulldozer, Piledriver and so on, and IBM keeping the performance lead with Power8 and beyond. Aside the now very common delays – as you can see in our stories nearly daily – things sound pretty simple. Or, do they?
Let's rewind the clock back some 13 years ago, to 1998, when yours truly was busy creating systems based on the 64-bit Alpha microprocessor, the Formula 1 of CPUs for the whole of late '90s. Being much faster than not just Intel, but also other processors at the time, yet offering extremely compact and elegant RISC architecture – simpler than Power or Sparc or ARM – with comparatively small but fast cores and high system bandwidth, coupled with excellent compilers, Alpha was popular in the high end computing circles – and also in the nascent 64-bit PC community, since it had native Windows NT without all the X86 viruses – around the world. Even Samsung and Mitsubishi were making the Alpha CPUs, including their own derivatives.
Do note that one of the most interested parties in the Alpha was, of course, China. It even officially licensed the complete Digital-Compaq Tru64 UNIX OS, one of the best at the time, and the associated software stack. The Chinese saw the benefit of a superb architecture and its programming model, and, I was told at the time in the internal meetings, would have loved it to continue the development at a maximum pace.
That was not to be then, unfortunately. The US corporate shenanigans surrounding the buyout of Compaq by HP and its nasty boss at the time, the political pretendent Carly Fiorina, and the seeming need to diminish the company's value before it's bought, saw Alpha literally murdered at the peak of its prime, for HP to justify its – later proven foolish – Itanium move, which also killed its own HP-PA RISC architecture. That reminded everyone of an equally foolish move a few years earlier by Silicon Graphics, today known as SGI, killing the MIPS RISC architecture it helped create to subscribe to the Itanium story – which we all know how is ending now, and I advised everyone on that outcome exactly ten years prior to now.
For the sake of the later read, do note that MIPS architecture was the close second best to Alpha in performance, simplicity, intra-core and, see today, inter-core scaling, and it was truly 64-bit since 1989!! Alpha was, of course, 64-bit only, without any 32-bit modes, since 1991, and, guess what, ARM will only be 64-bit in shipping products in 2013… tells you something, doesn't it? Add that MIPS cores are faster and smaller yet than ARM, and have vector FP for ages already…
Fast forward to 2011: China, which saw the true goodness of what were the best CPU architectures of the time, and without the 'software support pressures' of companies like Microsoft, or the corporate plays that saw the best things go, continued to develop both MIPS – read: Loongson – and Alpha – read: Shenwei – all this while. With the ground-up approach, and (fast shrinking) semiconductor process gap, it took them a decade, over multiple generations, to perfect things up. I have seen what Shenwei and Loongson can do, and the commitment of their teams with the huge resources of central government behind them. Even at the 65 nm process, the FP power per socket of each of these CPUs is matching or exceeding that of current Intel and AMD high end processors. What do you think they will do at 32 nm, and then 22 nm?
The 8-core and 16-core Loongson and Shenwei CPUs may not have catchy marketing names as their US counterparts, but they have the performance, low power consumption – the Jinan Petaflop Shenwei system consumes less than 1/3 power per teraflop compared to its US counterpart Jaguar – and clean (both technically and legally, IP-wise) vector-enhanced RISC architecture that can ultimately scale well from smartphone to supercomputer. Add to this the indigenous Chinese ARM architecture design companies, and you'll see that, despite its low starting position, China is managing to catch up with the US on the 'trophy prize' high end microprocessor front.
|Japan is not giving up, either – Fujitsu, which for years designed faster SPARC RISC CPUs than Sun Micro (now Oracle) the original SPARC creator, used a vector version of its SPARC64 to make a 10 PFLOPs K-system in Kobe, the world's fastest this year. The incredibly fast memory and interconnect systems come along as well. Interestingly, the design approach is similar to the vector-enhanced units that Chinese develop as well. Some common thinking there…|
Now, as far as I understand from the people in charge, the 2012 may be the year when China starts more aggressively marketing the fruits of its design. Isn't it an irony that the best that US design long ago, and then foolishly discarded, now, many years later and IP rights expired, appears new, enhanced and sped-up, from China, to compete at the global level? Well, maybe it is about time for some more variety in this business, or "back to the future".