I watched the WWDC Apple Keynote the other day. I am well aware that it is mostly marketing hype, but amongst the hyperbole there is stuff grounded in what is current. One thing they seem to be hinting at is that the speed of chips is levelling out.

It seems that 3GhZ is a limit which can’t be safely passed, and the speech talked up Apple’s multi core APIs which let developers multi thread their applications, but the OS itself organises them over the cores. Now, I am not sure exactly how this works, but the implication is that chip makers are turning to adding more cores because they can’t ramp the speed up any higher. They probably could, but there would be head dissipation problems. I’m not seeing ay 4GhZ chips being talked about, but I am hearing stuff about 16, 32 and even 128 core chips in development.

So, have we hit a limit for silicon, or is Apple talking up the current speed limitations to make Grand Central Despatch (their multi core system) more inviting?

I don't think so. All the Core i7's can overclock beyond 4ghz on air, and that's before the expected die shrink at the end of this year. As the manufacturing process shrinks, clock frequencies can go up without increasing heat out put. Obviously there is a limit to how fast things can run, but it's well above 3ghz.

There can be only one.

Mark

I think we are nearing the limit of transistor size and therefore clock speed.

As Levick said, the i7s overclock very nicely up to a very high speed but it's true that the rate in which CPU speed is increasing is decreasing.

I imagine we will have to wait for the new technology (i.e. new transistor materials etc...) before we see the next significant boost in CU speed.

Faster speeds are certainly possible as the OC abilities of several cpus demonstrate but Apple always sticks with stock speeds recommended by the chip manufacturers. The problem with OCing is that it can reduce the lifetime of the chips especially if they are running over their optimal working temp. With adequate cooling that isn't a problem.

It is however interesting that the top official speed of most chips is topping out at around the 3.2/3.3GHz mark and has done for quite some time. Still, I suppose you are getting more cores running at that speed with each iteration.

My first PC was 1MHz. It was only a few years before we hit 100MHz, then a few more to hit 1GHz. The pure clock speed increases have slowed down drastically since then, so yes - we're hitting the limits.

Even if they can make chips that run nicely at 10GHz, that would only be 3x faster. It doesn't compare to having 128 cores at 3GHz, which if you could actually use them all would be "equivalent" to 400GHz. There's no way silicon is going to break 100GHz, is there?

Whilst I think the idea that 3ghz is the limit for a safe clock speed, I think the idea that a faster single thread is better than two slower threads is quickly disappearing as programs are starting to make use of multiple threads. Grand Central sounds like a good way of doing things as does the use of OpenCL or CUDA to tap into the massively parallel potential of modern GPU's for things that can be split into smaller pieces. I have Badaboom running on Windows and the speed it chews through high definition video's is very impressive and puts my quad core CPU to shame. You can now buy a Leadtek add-in card with a Cell CPU on it for decoding video, even working in conjunction with your GPU to speed things up further. Yes, this is just video encoding - which is easy to split into pieces, but if you compare games like Far Cry 2 which is very well optimised for 4 cores to games like Crysis which only really uses two cores to any extent, the performance benefits are noticeable.

Trouble is, fast CPU's and GPU's end up moving the bottleneck to the hard disk.

Like people have said, overclocking disproves this.

And the AMD Phenom which hit 6.4ghz makes it look even worse.

I wasn’t really thinking about over clocking - I suspect that most people don’t do this, and the chips that leave the factory are rated below any theoretical over clocked speed.

But like JJ said, a I remember using an Acorn A3000 at home and that was a high spec computer at the time.

I was about 7 or 8 when I got it and it had an 8MHz CPU.

At high school maybe 4 years later I was using Pentium 133MHz CPUs and anything up to 166MHz.

At Uni my dad bought me a laptop with a 1 GHz (or something) CPU.

That's almost an increase of 150 times in the space of around 10 years.

Since then in the next 10 years we have had an increase of CPU speed of around 3 or 4 times.

Even with overclocking you can only get maybe 5 times increase in 10 years.

Unless you can overclock a current CPU to around 100-150GHz then you can't deny that we are approaching a barrier. Overclocking doesn't disprove anything. If anything it shows even more that we are struggling to increase CPU speeds.

The following is based on my understanding when read snippets of articles, hearing conversations, etc over the last couple of years.
When we had single core CPU's, Intel reached to a little over 3GHz, though AMD were achieve similar performance at lower speeds. Intel changed their architecture and improved the performance, so they could gently work their way up to creating similarly fast CPU's which would be far superior (core for core) than cores of old.

While you can ramp things up, is it green? In the last few years we've been constantly more and more aware of green computing.
If it is green, is it easy to ramp up the speed? Would people be more impressed if Intel;s new marketing said "Now even faster with 3GHz" or "Now perform more at the same time with six cores!"? Of course if may equate to the same thing, but for a company, what will sell comes into the equation.

As Gav said, the current thinking is to add more cores and program across them, as transistors become smaller, more cores can be fitted onto the same size (though why they can't make a new chip a little bigger, I've no idea)...

I suppose it's what the chip is for, one theory is more threads per core is better than more speed per core (give the analogy: You're moving house from London to Edinburgh, would you rather more a little at a time very fast in a Ferrari, or a lot at a time more slowly in a pick-up truck?). Experiencing this it tends to have poor performance under light load, but it has the same performance under high load too, whereas just more speed will give excellent performance under light load, but will be crippled under high load.

I remember reading an article on the old boards which highlighted a group who had created a chip which ran at something like 500GHz on a water cooled system, and they could get it up to something like 100GHz on air....

And is speed everything? Remember the PowerPC Macs, they were comparable in performance to PC's of the same time, yet they ran much lower clocked CPU's...

It's largely about the rejection rate of large chips.

The silicon wafers used to make a chip must be a flawless single crystal. Most wafers have some flaws, so if there's say 10 flaws and you can get 100 chips from the wafer then you get 90 that work and 10 that don't. If the chip is so big you only get 20, then half of them won't work. If you can only get 10 or fewer, then you'd be lucky to get a single one that worked.

One way around this is redundancy. If you make a processor with 4 cores, a flaw might mean you can still use 3 of them. Or if the flaw is in the cache, then you might still use half the cache.

I had read about this but what I've often wondered though, is how they test them in batches?

My grandma used to test chips in the 1970’s. They’d take a percentage of chips from a batch. If those chips failed, then the batch was rejected. If the batch passed, then the batch passed. She used to bring me back failed chips to look at. I spent ages with a magnifying glass looking at them.