I got some emails over the last few days- some civil and some downright senseless. Merced and Athlon comparisions seems to have been the hot buttons of the last few days and here is an effort to clarify my thinking behind some of my recent postings. I am curious as hell as to know the psychographics of the nut-cases who send abusive trash to people for sharing their opinions on a public thread but that discussion would probably not belong on the AMD (or INTC) thread ...
My understanding of the server market is that it falls into at least three distinct segments:
- Massive I/O centric devices supporting hundreds or thousands of client applications and serving large databases
- CPU centric devices that are used in engineering environment to run large simulations, place-and-routes, graphics jobs and the like
- Simple devices that act as email or print servers in most organizations and also devices that people call servers because they are slightly beefier than the desktops they deploy to their employees (eg: marketing server - a place where they put all the marketing stuff; HR server - a place where they put all the documents on HR policies, etc.)
Now, the third one I will ignore. Print, email, application servers in corporate environments are rarely processor intensive. In most cases, the lowest speedgrade available in volume (K-6, Celeron, M2, you name it) is typically sufficient for this.
The server market that people generally bandy about is one of the first two. And, you can see that these two segments have very distinctive needs. The processing intensive ones are typically FP intensive and highly application dependent. Most of these high end applications today reside on Sparc, PA or Alpha platforms. The FP numbers of either Merced or Athlon derivatives that will be available a year from now are unlikely to cause a sea-change in this segment. I do not believe that we will see porting of these applications to an Merced or x86 platforms en masse for a long time to come. I can see them migrating to x86 platform because of the sheer momentum and cost effectiveness of the platform but even that will happen at a slow pace. The trickle of applications on the new platforms may increase to a stream but I do not see it being a major factor. In summary, I do not see this as a big segment for Merced or Athlon or PIII or Wilamette, in the overall server space.
That brings us to the very first segment: the segment that is IO centric. I think all the Internet servers that everyone in the industry talk about fit into this bucket. In corporate environment the server situation was very contained, MIS folks clearly knew what the ratio of servers to clients was and had a very good handle on the charecteristics of the server loads. Internet has dramtically changed the traffic pattern seen by these so-called servers. The theoritical client to server ratio on internet is infinite - though most MIS folks start turning away service requests after a set load factor.
In this environment, the system architecture that makes most sense to me is the one in which the processor and IO modules are decoupled (or loosely coupled). MIS can upgrade IO and processor modules independently and bring different modules on-line/off-line depending on the loading and servicing needs. Now, the question is what CPU architecture would do well in this case? I think the answer is - it depends. It depends on if the purpose of this "system" is to serve large databases or small individual applets (I say small applets because the current internet enviroment is clearly not suited for serving large application. Yes, I am aware of ASPs.).
The applications where large databases come into picture, the 64-bit space offered by Merced would be nice. But, remember this - there are players already serving this niche and Merced is a potential new entrant to this party (I say potential because it is not shipping in volumes in systems and it willnot for a long time to come). And, even when it starts shipping in volume it is going to take a VERY LONG time before it gathers the software base to become an effective competitor. Now, imagine what a 64-bit extension of Athlon could do in this application space. (I have not corroborated the K8 rumors over the last couple of days at this point but I do believe that it is possible for AMD to have a 64-bit chip [x86 or otherwise] in the market in roughly the same time frame as Merced. One year is LONG time in this industry. For all we know Transmeta could be doing something in that space though I hear that they are seriously behind. The time to compare this stuff is when the respective chip is close to production, so I will not muddy this discussion with further forays into what AMD may or may not be doing in the 64 bit space)
The applications where 64 bits are not necessary, I can see Athlon or its derivates doing extremely well against Merced or Alpha or PA or Sparc. From what I can tell, Athlon bus is superior to either Merced or Wilamette (unless it is being changed). The number of outstanding transactions that Athlon permits is of great use for these apps, and so is the point-to-point upto 400MHz bus that can do wonders in multiprocessing environment. The extended MESI protocol could do a lot of good depending on how big a deal coherency is for the system. As we all know, the integer benchmarks on Athlon are very good as it is and are likely to get a lot better with higher MHz, core tweaks, larger caches or on-chip caches. (I can see 8 or 16MB L2 cache doing wonders in this application space.) If AMD can keep up the current pace on execution, I would be surprised if AMD does not have the #1 crown in this area pretty soon. Remeber, we are talking about Athlon derivatives if and when Merced systems ship in volume.
So, all-in-all, I believe that Athlon is a very compelling platform for the server space and I do not believe Cascades will cut the mustard and will very quickly be relegated to the low-end/mid-range. I believe that Intel is aware of this and racing to protect its high-end by trying to move its customers to Merced platform. If I did not believe this were not the case, talking about Merced and Athlon in the same sentense would serve no purpose. This is the reason when Kap talked about Red-x campaign - I said I agrees with him in the server context.
Clear?
Chuck |
