Intel keeps mum about its new baby

I was very keen on the concept of hyper-threading when I first heard about it from Intel about a year ago.

The idea makes sense in that the majority of the current and future applications for which we need high processor performance are suitable for multiprocessing, as the work can be split into parts and then run in parallel on separate processors.

However, I had only considered the possible positive aspects of hyper-threading and hadn't considered the likely problems.

I'd asked Intel about one problem with a pipeline stall when you had two threads going through a single processor. I assumed that both pipelines would be flushed, causing a significant penalty. The answer is no: just one thread is affected. But I had seen this as a possible serious problem.

Several years ago, we had a processor test that exhibited far more pipeline flushes than it should and the memory of the number of hours I spent tracing the code checking for wrongly predicted branches and the like must be still very much with me.

So I was surprised when we decided to look into the performance characteristics of hyper-threading in workstations to find that Intel was being extremely cagey about this new technology and was actually discouraging its use in workstations, at least for the time being.

After many conversations and emails with the company, I think the reason for this reticence goes back to the experience it had with the Pentium 4.

This was designed for what Intel expected to be the performance characteristics of modern applications and it had taken the decision not to be concerned about the performance of applications like Microsoft Word.

However, even though most people would agree that we don't need any more performance for Office-type applications, when the Pentium 4 came out that was exactly what most people ran on the new chip. The resulting criticism stung Intel and it is wary about having a similar experience with hyper-threading.

I think Intel's fears are misplaced, because the sort of people likely to buy an Xeon hyper-threading workstation are also likely to be fairly computer literate.

However, there's a clear problem here. The PC world isn't getting any simpler and getting a clear message over - even assuming that companies like Intel dropped all the usual hype and so on - is becoming increasingly difficult.

The particular problem highlighted by hyper-threading is that two things are happening. First, as well as making processors simply run faster, the other way of increasing performance is to try and find ways of making it possible to do more than one thing at a time.

From Intel's perspective, the superscalar development that started with the original Pentium has been doing just this. However, these developments are bringing rapidly diminishing returns and are introducing new problems: this feature makes program A run a bit faster, but slows down program B.

Second, this makes it all the more important to know how particular software will react, as the performance of different applications is diverging more and more.

You can't say that one processor is faster than another unless you specify what they're running. Change applications and the differences might be reversed.

It's a PR nightmare, as these aren't the kind of issues that can be expressed simply to a mass market in a 30-second TV advert. No wonder Intel is so guarded about hyper-threading.

Our testing shows that responsibility for performance improvements is being pushed more and more onto the software developers. Ten years ago, if Intel had brought out a processor that ran code twice as fast as on the last system, then everybody else's code would also run about twice as fast.

But that's no longer the case and Intel is expecting that, for performance-critical applications, developers will pay increasing attention to the characteristics of the processor. Hyper-threading is the most extreme example of this yet.

But Intel will also need to change. It has very good profiling capabilities and I suspect it will be looking at the kinds of resources different threads are competing for in workstation systems.

Until such a time is reached, though, Intel is right and the main benefits of the hyper-threading technology will be in server systems.

Like this story? Spread the news by clicking below:

 del.icio.us     Digg this     reddit!