As you can see, a platform developed around the AMD Opteron has a very ideal scenario whereby each processor acquires a separate memory array to work on while having their own built-in memory controllers and symmetrical multiprocessing controlled within the processor. On top of that, with vast amounts of I/O bandwidth provided by the royalty-free HyperTransport links to cater to the other subsystems such as PCI-X, Infiniband and 10G Ethernet, the AMD Opteron processor seems like a promising enterprise-class and High Performance Computing (HPC) processor (or platform, looking at what it controls) that is easily scalable upwards with increasing processor frequencies.
The AMD Opteron has also set a milestone in computing with their own AMD64 ISA, but unlike other 64-bit CPUs, the AMD64 ISA is an extension to the existing widely utilized 32-bit x86 ISA. As such, the AMD Opteron is the only x86-based architecture for the 64-bit computing platform. The Opteron can perform as a 32-bit platform as well as execute 64-bit applications since it does not emulate a 32-bit environment but rather, the core natively supports 32-bit computing. It was all the more easier for the Opteron since the base of the AMD64 ISA is from the existing 32-bit x86 ISA.
The single major advantage of the AMD64 ISA in the Opteron is the ease of transition from 32-bit computing platforms to a 64-bit. Users could just initially invest on the Opteron, seamlessly use the 32-bit x86 infrastructure without loss of performance while enjoy the vast memory addressing memory capability of the 64-bit CPU when employed with a 64-bit operating system. Some 64-bit platforms that will support the Opteron are SuSE Linux and the forthcoming Microsoft Windows Server 2003. Meanwhile, system managers and staff can work on porting their existing 32-bit application and databases to a 64-bit one as and when they see fit; the Opteron can offer you full 64-bit computing whenever you are ready to migrate. Another possible reason for investing on the Opteron is to consolidate other 64-bit servers (which are of course non-x86 ISA) that mostly transact with many x86-based systems and convert them to a unified x86 platform using the Opteron and its AMD64 ISA. This would greatly lower IS-staff running costs to maintain all the various types of systems and the difficulties associated with such mixed architectures.
Although it is an inevitable evolutionary process to move from 32-bit computing platforms to a 64-bit one, how soon would everyone roll over and fully benefit using 64-bit applications is an interesting question. Since the Athlon 64 and possibly a 64-bit desktop operating system from Microsoft will be upon us later this year bringing 64-bit platform to all, it is a positive sign that it probably wont be long before 64-bit computing platforms would be mainstream.
However, the real question is when would 64-bit applications and tools be as readily available as today's 32-bit applications? Would they really perform better? If the AMD64 ISA is well received by all industry players, that answer would be resoundingly positive. During the launch, AMD shared a positive example where engineers were able to port a 32-bit application to run native in the Linux 64-bit OS within just two working days. In the developer world, that's quite an achievement especially that performance was nearly doubled as a result of the port. Now thats a jaw-dropping potential and advantage of the AMD64 ISA. But the actual take up rate in the workstation and server market isnt as predictable as it would be for the desktop market and only time will tell if the AMD64 ISA has made a mark as impressionable as the introduction of the AMD64 ISA itself. Remember that AMD tried hard to market the 3DNow! instruction set, but in the end, it went wayside. Although this isnt a fair comparison to the AMD64 ISA, it is still a reminder to the possibilities that may occur, no matter how easy it may be to implement.
Evaluating the AMD Opteron is no straightforward task either since 64-bit platforms have almost always been an exclusive one and is a much younger platform in contrast to the 32-bit x86 platform where many tools and applications widely existed for many years now. Also, benchmarking for the desktop systems is a lot easier because of the predictable usage pattern and tools developed to harness that. Server environments are not and they are subjected to a number of external factors that affect the overall effectiveness of the servers.
Looking at some of the industry benchmarks compiled by AMD, the Opteron seems to be having an upper hand on performance. But above them all, we consider the highly regarded TPC-C benchmarks as a more accurate measurement of server performance. According to this database server benchmark, a 4-way Itanium 2 server is still holding a significant 20% lead over the 4-way Opteron server system. Although the system parameters arent exactly alike, it should still give a relatively good gauge of where the products stand. However, the AMD Opteron 4-way server system is head over heels when compared against 4-way Xeon server systems. Yet another positive outcome from the TPC-C benchmarks is the measurement of system cost against performance gained and the AMD Opteron system scores a tremendous win in this department, which is definitely a substantial consideration for most corporate IT managers. Thus the AMD Opteron is a certain winner, from the cost perspective.
If you were to ask as our personal opinions of the possible pitfalls of the AMD Opterons design, we would probably say the high design costs involved in the tweaking of the Hypertransport and memory controllers for each technology advancement. Since they are built into the CPU, they can no longer be developed separately and pushing further speed envelopes may prove to be an uphill task for AMD. Perhaps a longer interval between the introduction of each CPU speed increment can also be expected due to the above reason. Additionally, we also have reservations on how much of a multiprocessing overhead may occur in an 8-way glue-less SMP setup since there's only this many ways to connect eight CPUs with three Hypertransport links in each of them. We think there'll be areas with potential bottlenecks and scalability with more processors may not be as rosy as they seem.
Looking further down the road, the AMD Opteron may require a different socket layout if it undergoes a drastic memory controller change, such as incorporating a DDR II controller and this is where we think there's potential inflexibility of this design. However, even Intels server class series is dotted with various CPU interface variants, so although this may seem fine, it would be advantageous to AMD if they can keep the same interface for some time to come. This is as far as we can speculate at this point of time, but if all the pieces fall in place perfectly for AMD, the Opteron could be the next big thing in the modern computing era.