Advanced Micro Devices
What Would it Take to Save Itanium?

Format for Printing

Format for printing

Request Reprints


By eachus
March 2, 2004

Posts selected for this feature rarely stand alone. They are usually a part of an ongoing thread, and are out of context when presented here. The material should be read in that light. How are these posts selected? Click here to find out and nominate a post yourself!

There has been a lot of talk in another thread about what Intel's plans are for future Itanium processors, and whether if Intel accomplishes its goals Itanium can replace Xeon. I think the answer is no. Not if this or if that, or assuming this much improvement in one or the other. Just no.

Let me explain. The economist's reasonable man will buy whichever computer system provides him with the most value for the money. Of course, if he has requirements which can't be met by System X, then its value to him is zero, whatever the price. If two systems both meet his requirements, he will look at the actual costs to him of each solution, and choose the system with the lower overall cost.

In the server market, greater than 32-bit address space is fast becoming one of those requirements. I say greater than 32-bits because we could spend hours defining which limits are architectural--part of the ISA--and which are specific to particular chip designs. But accept for now that Itanium, IA32e, and AMD64 all have limits that are similar to other existing 64-bit architectures like IBM Power, UltraSPARC, Alpha, and PA-RISC. A dozen years from now those differences may matter--but by then all of the still viable architectures will have evolved, eventually getting close to 64-bit physical and 64-bit virtual limits if needed.

So once Intel starts shipping Nocona, Xeon will get that magic checkmark. Now let's assume you are a CIO, and a rational man, and you need to move your current server farm to 64-bit. Let's just say for grins, that you can buy Intel Xeons with performance similar to your existing x86 server farms, or Itanium3 systems with ten times the performance and one-tenth the cost. Of course, you still buy the Xeons.

Why? A part of the costs you are comparing are software costs. For most server customers today, software costs are greater than hardware costs. Think about that for a second. Even if the Itanium hardware was free, the cost of software would make it uneconomic to replace existing x86 servers. The current software has been developed, licensed, or paid for, and that is a sunk cost. If you replace the current servers you don't get it back. Of course this may not apply to new systems to support new applications, but then you run into the left hook...

The expected cost of interface problems between the current systems (and software) and the new systems also has to be evaluated. Even in a well-managed shop, with excellent software management capabilities, the costs of interfacing between different applications can exceed the cost of the applications. Of course, the key word there is "can." Sane IS professionals have long since learned from bitter experience, if no other way, how painful those costs can be. So they run experiments if necessary and get a good handle on the potential interfacing costs before making a purchasing decision. Obviously if the cost of interfacing to System X exceed the cost of adding a compatible member of the current software suite on your existing systems, it doesn't matter what the asking price for System X is. It is too high.

The risk of migration problems is also a cost you have to put a number on, and this is often the showstopper. If transitioning to the new servers is going to put your company at risk of bankruptcy, it takes either an IS team with belief that they have godlike powers to propose it, or a situation where not replacing the current system is worse. (Obviously more than a few companies found themselves in this position with Y2K. But Itanium missed that potential boat.)

Do I need to say more? AMD invested a major part of the development of Hammer into guaranteeing that they could run existing software without change. By now that promise has be thoroughly tested in the marketplace, and found to hold true for existing 32-bit operating systems, and for 32-bit applications running under 64-bit Unix. The final release of 64-bit Windows is not ready yet, and there are still some driver availability issues, but it seems likely that yawn, yes all existing 16 and 32-bit software will run without change on Hammer systems. Will Intel do the same for Xeon? They had better, or IA32e will be stillborn. (A more interesting question may be whether Intel will make that assertion for IA32e and AMD64 or not. Right now it looks like they won't, and I expect that to hurt Intel more than AMD. In this area, right now Intel is not the 600-pound gorilla.)

Can Intel promise the same for Itanium? Of course not. To me the major complaint about the current hardware emulation capability from real users is not that it is slow, but that it is not perfectly compatible with x86. The new software emulation layer may be much better, but there is a lot of room between much better and perfect, and I don't know that Intel will spend the necessary money not only to make the emulation perfect, but to prove it to users. "Can you install Windows 2003 Server(x86 version) from CD-ROM? No? Come back when you can."

So there are these three tests which Itanium has to pass to make it into the server rooms at most companies. When Xeon was 32-bit only, even if Itanium failed one or more of these tests, it might be considered if there was no way to expand the existing servers to meet new requirements. But that changed last month, and it is fast becoming clear that Itanium was failing one or more of these criteria--but being considered anyway. But now, even coming close to failing one of these tests is enough to rule Itanium out of consideration. Good bye Itanium, we hardly knew ye.

Notice that I am discussing corporate server systems above. Nothing I said there applies to the cost effectiveness of Itanium in HPC applications, or on the desktop. Today there are HPC applications where Itanium wins on cost effectiveness--and those where it doesn't. Itanium workstations seem to have a tough row to hoe, but they will stand or fall mostly on (hardware) cost and performance, plus software availability. Similarly, there may be other application fields where Itanium can win if fast enough and inexpensive enough. But now we run into another gotcha for Itanium.

The price of almost anything manufactured can be separated into the fixed costs of manufacturing and the marginal cost of making just one more--or one less. In simple terms:

        Total_Cost = Fixed_Costs + Variable_Costs * N    (1)

Where the number of units manufactured is N. If we divide through by N, we get:

        Average_Cost = Fixed_Cost/N + Variable_Cost      (2)

The discussion above indicates that the demand for Itanium has recently decreased. This makes N smaller. I think that everyone with any sense now agrees that at any price, Itanium no longer has a chance to cover its fixed costs. Well, once you realize you can't make money, there is another equation to consider. Break fixed costs into sunk costs and sustaining costs:

        Average_Cost = Sunk_Cost/N +
               Sustaining_Costs/N + Variable_Costs       (3)

I broke it that way intentionally. If:

       Net_Income = Price * N -
               Sustaining costs - Variable_Costs * N     (4)

net income is negative, you lose money with each chip you sell.

So where is Itanium? I think that Intel hoped that this year or next, the net income would turn positive, even if there was no chance that the sunk costs could ever be recovered. Now I don't think the potential Itanium market is large enough that it can ever reach that point. Alpha and PA-RISC were both discontinued for exactly that reason, and their markets were significantly larger than the current Itanium market.

So I think that as soon as Craig Barrett retires Intel will announce the EOL plan for Itanium--if not sooner. I don't think they will cut it off cold, but probably choose an upcoming part to be the last new Itanium.

Become a Complete Fool
Join the best community on the web! Becoming a full member of the Fool Community is easy, takes just a minute, and is very inexpensive.