Let me ask you: if your wildest dreams were realised, how many cores per CPU would you have in your servers, workstations and power desktops right now? How much Level 2 cache memory would you have in each core, or would you rather it be shared amongst the cores? Would you rather have memory controllers for each pair of cores that access a set-aside block of memory, or one memory controller that sees the entire address space?
Oh, so that’s the sound of one million pages turning. I’ll wager that the majority of my certifiably savvy readership hasn’t given such nuts-and-bolts matters much thought. Well, Intel and AMD are giving it lots of thought right now, and I imagine the burning question at AMD is this: what do we do after AMD and Intel are matched at eight cores?
If I had a vote, I’d have both vendors stop at four cores and focus on fat and fast buses that give those cores something to fill instead of something to wait for. AMD and Intel both face bus bottlenecks, and that’s the bane of multi-core. Presently, dual-core CPUs from both chip-makers have to share memory and I/O buses that were designed to serve a single CPU core. Every time you add a core, contention for access to memory and peripherals rises, lowering the genuine performance benefit of the extra core.
Even AMD’s laudable design can’t deliver near-linear performance gains from additional cores. AMD is revamping AMD64’s total design for quad-core so even when cores get stuck in contention, the buses run so fast that the traffic clears quickly. AMD is taking a run at getting third-party vendors signed up to place their peripherals directly on its hypertransport serial bus.
If AMD can make that work, then, potentially, every core can have direct access to system peripherals. That would be a quantum leap for x86.
What worries me is that Intel might just go nuts bumping the number of cores, clock speed, cache and front-side bus speed while more or less hewing to today’s core micro-architecture in terms of key factors like bringing its memory and bus controllers on-chip.
Let’s say that AMD wants to hold fast at quad or octo-core and work on perfecting the total system architecture so it can keep up. But what if Intel keeps cranking up on cores and gigahertz?
While AMD is sledding along with eight cores, non-uniform memory architecture and direct on-core bus links from everything to everything in the system, Intel might ship a 16-way core micro-architecture CPU/chipset that looks more or less like Cloverdale
4-way, except for more — you guessed it — cache, gigahertz and cores.
The question is, how will AMD react? The buying public isn’t yet fully educated on the advantages that AMD64 has over core micro-architecture. When both chip-makers talk 64-bit computing and virtualisation, my sense is that most buyers (not, I hope, my readers) take these to mean that AMD and Intel are in lock-step.
Not terribly long from now, AMD could find itself cornered into explaining to non-gearhead buyers why AMD64 stayed put at eight cores and 3.8GHz when Intel announces 16 core, five GHz core micro-architecture CPUs with 2.5GHz front-side buses.
I want to tell AMD to keep to the high road and trust that the commercial market will come around as it did with Opteron. But I know that before Intel runs its full-page ad in InfoWorld hawking its 16-way core micro-architecture CPU — the only CPU that can cram 32 cores into a 1U rack chassis, I’ll have to make sure you know what it all means. I don’t assume AMD will have the better story; I’ll lay it all out on the table, fairly, in terms you’ll understand.
(Shortly after this column was written Intel announced at the Intel Developers Forum that it plans to go up to 80 cores in five years’ time — Editor).