Hi all,
This is an interesting topic, and although I do not have any technical knowledge on processor architectures (I come from a background of large scale scientific computing for basic research), I have hands-on experience for the last 25 years and I will share some of my thoughts and observations based on the experience.
There was some discussion on reliability of the processors. I tend to think that in recent times all processors are very very reliable. Even the cheapest of processors never give up these days (For example, at home we still have a 6 year old AMD Sempron based PC. It is way outdated and but has been run about 15 hours a day all these years. It's still running flawlessly. There are many other examples I could give here).
There was a time (about 10 - 15 years or perhaps 20 years ago), when for reliable applications like servers, people started choosing RISC processor based systems (like Sun Sparc, DEC alpha etc) primarily because the CISC was deemed less reliable due to very high clock speeds. These days I do not hear people complaining too much about unreliability of CISC processors due to high clock speeds.
The RISC processors were also quite powerful. But they have always remained very very expensive compared to CISC. The last experience I have had with some RISC is IBM Power 4, Power 4+ and Power 6. But in scientific computing we look for performance to price ratio and in my mind the RISC does not qualify for real high end computing just because of their poor performance to price ratio (an exception was the DEC alpha processor. Digital Equipment Corporation was run by academic minded people and they offered great performing machines at great prices in the nineties, but could not survive the corporate world and ultimately was gobbled up by Compaq and in turn by HP, the latter two together destroyed any future road map for the alpha processor and basically killed the most likable RISC processor ever in my mind). Another thing that goes against RISC these days is their high requirement of electrical power for a given performance.
The most unique Supercomputer company CRAY Inc (unique because they produce only supercomputers, and one of the 2 current major makers of supercomputers along with IBM with their Blue Gene) incorporates AMD Opteron multi-core processors for their latest petascale (and beyond) supercomputers XT4, XT5 and XE6 with lean linux kernels. Many brands prefer AMD Opteron based parallel cluster machines (for medium range computing). The reason behind this was the unique hyper-transport feature of these processors which enable each processor to communicate with the others directly, bypassing the system bus (the small bandwidth of the system bus is always a major bottleneck in large scale parallel computing). In addition, memory and cache access is also superb.
In contrast I remember the 1024 node Intel Paragon supercomputer housed in Oakridge National Lab (thought to be the biggest computer of the time) between 1992-94 was a very unreliable machine. It was breaking down all the time and even gave wrong results at times. But I believe the poor result was due to an unstable OS more than the processors themselves.
In desktop computing, in the last 3-4 years Intel has caught up with AMD in overall performance (including memory and cache) and perhaps has gone a bit ahead. But it's also true that if one applies the performance to price test (less relevant for desktop use, unless the user is a serious gamer and ultimate performance is needed regardless of the cost), Intel may just lose out to AMD for processor-motherboard combination. For example, I recently configured a PC for our home and we got a Phenom II X4 955 processor with a Asus 880 EVO board and I think it's value for money. In comparison an i5 based system would have given me 10-15% more performance (somebody please correct me if I am too much wrong) but at a significantly higher price. In addition the Asus AMD-board comes with a ATI Radeon 4250 chip based graphics which is adequate for us (we are not serious gamers, FIFA11 is our limit).
So basically it boils down to your desktop usage and the money you want to spend. BTW, AMD is NOT an Intel clone and I do not think they ever had that image (at least internationally) if not in India). They are in this much longer than Intel and there are many firsts by them. For example the use of MMX by creating vector registers for enhancing better graphics was first done by AMD and Intel followed suit. One can use these vector registors for scientific computing too by using SSE and SSE2 instruction sets.
BTW all the above processors discussed above are 'scalar' processors despite their (small) vector registers. True vector processors were/are built by Cray Inc and NEC. They are really really very expensive (prohibitively high), virtually nobody is buying large installations of them today except perhaps entities like the US Defense Dept. I have had extensive experience in these sort of machines from 1985 till 1992. The world has chosen to go along a more efficient way of using many moderately potent processors at a time connected by a high bandwidth/low latency interconnects (generally termed as parallel computing) than having one or a few very very potent processors that can do a do-loop (or a while or for loop) involving variables with a very large dimension all at one go.
Regards.
Thanks for sharing. That was a very nice retrospective on the evolution of computing.
You're right about reliability. Even in commodity consumer processors, CPU reliability for both AMD and Intel has increased to the level that CPU failures are almost non-existent. In server CPUs as well, both companies have been steadily incorporating many of the
RAS features that were previously only found in high-end RISC based mainframes.
Along with a doubling of performance every 2 years, this is also one of the main reasons that the x86 architecture based CPUs from both Intel and AMD have been steadily taking over marketshare from other traditional CPU architectures that were mainly RISC based. This is why Cray revamped its strategy and adopted x86 in all its mainframes. Apparently, the DEC Alpha team was also responsible for AMD's performance jump compared to Intel as many of the design team moved to AMD when Compaq sold off Alpha to Intel.
In the server space, you're right, AMD has been trouncing Intel in anything bigger than a 2-CPU configuration server mainly because it scrapped the FSB early on and adopted the point to point Hypertransport interconnect technology that resulted in a huge increase in inter-CPU and CPU-memory bandwidth. Intel on the other hand stuck on to the ageing FSB for much longer than it should have. Recently though, Intel has caught up with a very similar technology called QPI or QuickPath Interconnect. For the record, QPI is not a copy of Hypertransport, and no, AMD was not forced to share this IP with Intel. QPI, earlier called CSI, was actually built from the ground up by Intel's design engineers teamed with.. surprise.. DEC Alpha's design engineers. Another interesting thing: QPI was supposed to first debut with a Xeon chip codenamed Whitefield which was actually named after Whitefield in Bangalore as the design team was based out of Bangalore. Unfortunately, the chip was scrapped before production and never saw the light of day. Due to this, it actually ended up debuting many years later.
Now, as the technology pendulum has swung, as it often swings in the tussle between Intel and AMD, Intel actually has superior processors and a superior interconnect technology. QPI for example has much higher bandwidth than Hypertransport, and lower latency to boot. Intel's floating point AND integer performance is about 1.5-2 times more than AMD's comparable chip. Today, AMD competes by essentially selling 4 cores for an equivalently priced 2 core Intel CPU in order to remain competitive. In some benchmarks, a 2 core Intel actually beats a 4-core AMD, but only in some. Another creative thing that AMD is doing to stay competitive is that it is pricing 4-way server (which earlier used to command huge premiums) at only slightly higher than an equivalent 2-way Intel server. For example, if you take a look at the top supercomputers today, you will see a large number of Xeon 5 series, basically Nehalem servers connected with QPI.
Of course, for the end consumer, especially for the price conscious cosumer, AMD still offers much better bang for the buck, both in terms of price-performance and in terms of motherboard price and features. The only thing Intel probably takes away from AMD is the fact that an Intel CPU+Intel motherboard still offers rock solid stability which many of the AMD motherboards don't, unless one chooses wisely.
Just to reiterate, the point of me telling all this is not to give a biased view or to promote any particular company. I'm just trying to set the record straight, and share some other facts and tidbits that I can remember. As I have said before, the technology pendulum invariably swings once every few years, often changing the whole technology equation.
For example, if you want to look at pure price-performance today for high performance computing or numerical analysis or scientific computing, nothing, and I repeat, nothing can touch the floating point performance of a GPGPU. A single commodity GPU in a high end graphics card by nVidia or ATI today is capable of close to 2 teraflops. This is a ridiculous amount of performance! No CPU on earth can come close to this kind of performance. This basically means that I can theoretically put just 512 graphics cards or nVidia Tesla chips together to get a petaflop of performance, which is what some of the top supercomputers are capable of! Only, they used tens of thousands if not hundreds of thousands of CPUs to get this kind of performance. I wouldn't be too surprised if we see more supercomputers based on GPGPU rather than CPU.
. The market is changing. In a few years time, only hard core techies (and of course corporates upto a certain extent) will have PCs in their homes. The portable market is already being ruled by the likes of ARM and intel has a long way to go. Intel cannot survive on just server computers, its biggest market is still desktops and home user, which is pretty much going to disappear. With the home user going away from MS and intel, the server market probably will also partly move away from intel based servers to other systems. Microsoft is not going to do a hell of a lot better than these guys, it will also face a huge crunch. Their portables OSs are still junk. Office is still too huge a monster for portables. 
