ARM takes on the server big boys

Well for those who say 32bit is all that is needed, your right unless you want a database in there somewhere. Data grows exponentially every year so the more of the tables you can keep in ram the faster the access... 4GB is very small footprint to do this. For your average home/Soho/medium business 32bit is fine but for someone like Google or Microsoft 64bit is necessary.

I am happy about this development though because it will just take away segment space from Intel so that they work even harder to please the people who buys it's kits and sell them for less.

@Bazza
I worked with Transputers in the early 90s (in Occam - not lower level), so I appreciate the nostalgia, but I don't expect the clock to be turned back to a simpler time. The world runs on "apps written in rubbish languages". I doubt businesses will pay to have them rewritten for a sake of a new little processor.

@jpf
PAE or any other kind of page table hacks, segmentation or base+offset scheme is dead wrong. Addressing beyond 32-bits demands a 64-bit ISA. ARM has an elegant instruction set and I'd be seriously disappointed if they did anything other than elegantly extend it. Had MIPS been the core of choice they'd be in a better position having had a 64 architecture from the start. I'm very interested to see what ARM come up with.

@Stuart
Agree. The Archemedes was a great machine.

If you define "server" as a standalone computer that performs a small, set task, then we already have ARM based servers today. However, if you're talking in the general case of a compute server, then you need to handle modern workloads. 4GB isn't a large amount of memory nowadays.

The server world is pretty much addicted to x86. As feature geometry shrinks and efforts to design for low power evolve, the complex x86 instruction decoder will be of little consequence. Eventually we'll have super efficient and low power cores that are fully x86. Depressing... isn't it...

I don't think Intel has anything to worry about. Nor do ARM though... they're doing great.

exactly why google acquired agnilux

@64bit ISA Paul:
You don't need a 64bit ISA to address more than 4GBs of ram, think of x86 PAE (http://en.wikipedia.org/wiki/Physical_Address_Extension) for example.

Some of us remember when ARM had the fastest CPU in a desktop PC and ran its OS (RISC OS) blistering fast.

Nice to see the possibility of it returning.

@Paul Gray:
Today's CPU architectures with AMD's Hypertransport and Intel's Quickpath are already mimmicking the Transputer. Transputers allowed for joining several together over serial links. All that a multi-core / multi-chip machine these days is are cores connected over serial links like Hypertransport. It's just that that fact is 'hidden' from you under a bodged veil of SMPness. OK, shared caches complicate that picture a bit, but not much.

If you write software in the style of that you had to adopt for networks of Transputers you can get tremendous performance from a modern CPU. This is because you're exploiting their underlying architecture rather than abusing it. Thing is though is that style doesn't really fit today's object orientated world. Look up Communicating Sequential Processes in some very old books...

@64bit ISA Paul:
You only need 64bitness if the entire workload for a machine demands it. So for a host with 10 virtual machines you'll need a shed load of RAM, and 64bit is necessary. But if you go for 10 smaller physical machines to do the work of those VMs instead, it's likely that none of them will need to be 64bit. For example, a process handling a single client's session on a website is very very unlikely to need access to more than 4GByte of RAM...

If the industry rumours are true, it looks like the clever boys have finally worked out that hulking great machines hosting many VMs running apps written in rubbish languages can never be as efficient as much smaller physical machines running well written resilient software. This could be very catastrophically bad news indeed for VMWware and Intel and AMD.

@Jason Goatcher
Good plan, I like it! However, I fear that it might be bad news on the whole. If the server world adopts ARM wholesale (and I think that is inevitable in the medium term) then a whole sector of the market will no longer be driving the development of x86. Business is moving wholesale over to small PCs for worker's desktops as most don't have a need for anything other than a word processor and a browser. The only sectors left will be home gamers and the very small supercomputing sector. There's not much profit in the home PC market as it is, certainly not enough to justify the £2-3billion R&D costs of a new CPU which is largely funded by server sales. Weirdly this could be very bad news for nVidia too - fast GPUs, but they still need a pretty meaty CPU to deliver the overall gaming / graphics experience.

@Everyone
Well, it's taken nearly twenty years. The world finally seems to be realising that most of the time there is no need for huge amounts of compute power in a single computing core. The whole of the past twenty years has been about increasing compute power and then getting that power to do more jobs simultaneously. Now it has been realised that for the majority it's better to spread jobs out over many cores instead.

As a long time parallel computing type (I even built Transputer based computers from scratch...) I'm pleased. But as a professional user of the fastest single thread performance that I can lay my hands on, I fear that the end of the road is nigh. We have nearly reached the pinnacle of single thread performance, and from there we can only go backwards. That might mean that there's things that can be done today that might not be possible in the future. Now that really would be weird in the world of computing!

Isn't a 64-bit instruction set necessary to play at the server table? I thought that ARM was 32-bit only...?

"East wasn't willing to name the chip foundries that ARM is working with, however he said that information should tip up in the next 12 months."

Please let it be Global Foundries, please let it be Global Foundries.

Why do I want it to be Global Foundries? Simple. Because AMD is part owner, and as long as they don't go bankrupt they'll be doing their best to breathe down Intel's neck as competitors. And THAT, my friends, means cheaper chips for the masses.

Although, I guess ARM might at some point start competing in general purpose computing themselves at some point. All I really want is at least 2 decent competitors competing for my computing(or gaming) dollars.

a 4 CPU processor, delivering 10,000 MIPS @ 2GHz and drawing just 1.9 W of power. That's about 3 times the performance of an Intel Atom N270 on 25% less power. The Atom is becoming the CPU of choice for SOHO class servers so there's no reason that ARM shouldn't go after that space themselves.

BTW, @Paul Gray, stacking CPUs on a single socket hits the laws of diminishing returns. You get limited by memory and IO bandwidth quite quickly.

If they made a simple 4 core multicore design and included a CPU level FSB say akin to what the old occam transputer did, thus enabling the multi cores to be linked.

This would enable stacked designs, of course some form of heatsink/pipe sandwich inbetween layers. This approach enables a easily expandable number of cores using the small of die cpu hits for the chip to chip talk but with cheaper production costs due to higher yeild for chips per wafer over cramming things into smaller and smaller designs. Of course there will be some careful design to cover this approach and things like having the cores in and out and layout on the chip so any heatspots are alternated per chip layer and or design of the flow of the heatpipe sandwich need to be carefuly thought out well in the early stages and work back.

Now if they did this, the world is there oyster as they would finaly be able to offer a cheap 1 socket expandable multicore range with reduced investment on the motherboards due to expandability of this appraochl; least if done right.

But we all have our dreams.

ARM has NEVER been about raw power. Maybe you missed the point of the article.

OMG up to 2 GHZ they'll rule the planet with that kinda power