So, since the Intel head honcho Craig Barrett hinted at some possible slight differences, i.e. things that Intel CPUs may do that AMD ones may not, sufficient enough for software vendors to produce specific things, I started comparing the Intel "64-bit extended" architecture manuals vs AMD64 manuals, hoping to quickly find Intel's advantages and the unique features that Dr Barrett hinted at.
Finding the differences: a needle
in a haystack?
Well, it has turned out to be a very, very tough job, guys - just like trying to find differences in the genetic code between identical twins. Up to now, this stuff looks identical to me, and with it all the talk that "Itanium is for 64 bits, and Pentium/Xeon is fine with 32-bits" becomes moot. These chips are as AMD64 as AMD Athlons are X86 chips.
Also, the mantra that "64-bit enhanced" Nocona Xeon is good for 64-bit addressing, but not for 64-bit processing compared to Itanium (well, it and Opteron are just as good for 64-bit processing as the Itanium is) is to be thrown into the nearest paper recycler too. Simply AMD64 processors from both AMD and Intel are as 64-bit in terms of addressing and data chunk processing as Itanium or Alpha or POWER or SPARC are! It just happens they have, er, excellent 32-bit compatibility since the 64-bitness was added on top of the old, cumbersome X86 kludge. More interestingly, this 64-bit "extension" can equally well be added on top of the Pentium M core, too!
Comparing the horse(power)s
Let's assume Nocona rolls out fine a quarter from now, at a nice 3.6 GHz clock (at least it can then match the 3.2 GHz Xeon in the worst case) with 800 MHz FSB, and competes against a 2.4 GHz Opteron in both 32-bit and AMD64-bit applications. Which horse has got more power?
Well, Nocona might lead by 50% in clock frequency, bringing along a 50% higher peak FP rate, good for churning out supercomputer cluster nodes with high theoretical (I didn't say practical) FP speeds - 7.2 GFLOPs peak in double precision is not a joke compared to 4.8 GFLOPs of a 2.4 GHz Opteron, but still less than 2.2 GHz PowerPC 970FX machines from Apple and IBM, which will throw 8.8 GFLOPs peak per CPU, thanks to its 4 FP ops/cycle peak execution. Of course, the real maximum achievable rate depends on the CPU and system environment, as well as compiler and math library optimisation.
Integer-wise, I expect Nocona 3.6 GHz to be a bit slower than Opteron 2.4 GHz in most applications, whether 32-bit or 64-bit (ignoring, of course, any secret "elements" by Intel - Windows or otherwise). On the memory front, well yes, Nocona gets a boost from finally getting a 800 MHz FSB throughput, and using DDR2-400 memory. But hold on, didn't we see that those first DDR2-400 DIMMs have a bit higher latencies than top DDR-400 DIMMs? A Dual Opteron 2.4 GHz still has two 6.4 GB/s memory paths, one per CPU, plus a 6.4 GB/s HyperTransport link in between them. So, either CPU can access both its own and the other's CPU memory - at the same time, in parallel. In the Nocona case, well, the two CPUs still have to compete for one FSB - depending on the applications, your mileage may vary, but the Alpha-style Opteron memory architecture sure beats a shared FSB in elegance and scalability. That is valid especially in quad-CPU and eight-CPU systems, where, by incident, Opteron will have the whole "64-bit extended X86" for itself till Intel Potomac comes out.
The Battle's Just Starting
While AMD prepares its 90 nm process, and Intel settles the frequency and power/heat issues with its first 90 nm parts, these CPUs are just a preparation for the tougher battle starting late this year, where Nocona's Jayhawk followon will compete against 90 nm Opterons, and Potomac will probably make an early entry towards yearend to assault the Opteron 8XX series. Also, the battle is expected to move to the desktop, with Tejas CPUs in LGA775 sockets taking on the Athlon64 and 64FX. Or is it maybe some form of Pentium-M desktop version with 64-bit stuff in, too?
But, anyway, wait for a while - it will take me maybe a week to fully go thru the 1000+ pages of Intel and AMD books. Who knows, there may be a small difference somewhere? µ
Sign up for INQbot – a weekly roundup of the best from the INQ