It is true, Intel's Xeon can already address 64 GB (36-bit) and the Athlon can address 8000 GB (43-bit), but only by using tricks. A 64-bit CPU could address those 18 million terrabytes in one piece.

A 64-bit architecture features 64-bit general purpose registers and instructions for mathematical (ADD, MUL…) and logical (And, Or) operations on 64-bit integers. It should also handle double precision (64-bit) floating point faster.

Why would we need so much memory ? Corporate servers are already being shipped with well over 1 GB of RAM. And remember playing "Doom"? That 3D-game ran well in the 4 MB of my 486. Now, most games today need 64 MB to run well. In fact, a lot of gaming rigs feature 128 MB and more. In other words, in a period of only 6 to 7 years, the amount of memory in your average PC has become 32 times larger! In about 6 to 7 years, we will be using 4 GB memory in our systems, the maximum of amount of addressable memory of the PIII and K6-III.

So both Intel and AMD want to bring a 64-bit processor to the market. Let us examine the different choices they made.

The first x86 CPU's that powered the PC, namely the 8086, 8088 and 80286 were 16 bit CPU's. Compared to the other CPU's of their time they were ridiculously slow. Intel extended the x86 instructionset to 32-bit, added 32-bit registers and voila… the 386 was born. Can't they do that again? Extend the x86 instruction set to 64-bit this time? What is so bad about x86? Well…a lot.

1. x86 instructions can be 2 bytes long (or short rather), but they can also be 15 bytes long. This makes the decoding stage of the x86 CPU overly complex. Both AMD and Intel have found a workaround: Convert those variable length instructions into fixed length instructions (ROP's). Which makes the CPU more complex and bigger.

2. x86 has only 8 general purpose registers. That is far too few for a modern superscalar CPU like the Athlon, which can issue up to 9 instructions in one clock cycle. Each number of a calculation needs a register and you've got to write the results of your calculation in a register. There is again a workaround: Modern CPUs contain extra, "secret" registers. The x86 program works with only 8 registers but CPU assigns the numbers to other registers. Which makes the CPU more complex and bigger.

3. If you compare the Xeon with its competitors like the SUN UltraSparc II and the Alpha 21264, you will notice that the Xeon has the slowest FPU of the bunch. The reason for this is that x87 instructions use an operand stack. Simply said, it means that you have to make sure that the value you want to work with is always on the top of the stack. Before the FPU can even start crunching those floating point numbers, you've got to shuffle the required values to the top. Such a waste.