PC General Discussion Thread

[Darth Alec]

I play games, so I'd rather have few with tons of power.

But what does most people use? That's really the question. Theres always a marked for dedicated gaming machines, so I believe that mainstream will go with more cores.

[E6600]

Originally Posted by Darth Alec

I play games, so I'd rather have few with tons of power.

But what does most people use? That's really the question. Theres always a marked for dedicated gaming machines, so I believe that mainstream will go with more cores.

That's kinda like what my COE professor said, but he also claimed that more cores with less power is a better option than less cores with more power. I'm not sure I can agree with him, but then again it really depends how you use your PC, but from a marketing perspective I think more GHz sells.
If you consider how Supercomputers are made, they don't use a single CPU rated at 999PHz, instead they depend on parallelism, thousands of average CPUs working together.

[Darth Alec]

True, but super-computers also do tons of computing. I believe it's more practical to have several cores once each core reaches a certain speed.

[Power666]

Originally Posted by E6600

Which would you rather have?

8 cores each clocked at 500 MHz or 4 cores each clocked at 1 GHz? Given that both of them will cost the exact same price and will have the same amount of cache.

There is not right or wrong answer because it entirely depends on the task at hand. Additionally, other subsystems like the memory controller, coherency traffic, IO etc. play a huge role in which design philosophy can scale for a given task.

For 'general' tasks and performance, clock speed is the still the major factor. Multithreading doesn't lend itself to every task. Even with code that can be multithreaded, there are often parts dependent on a serial portion. This serial portion becomes the limiting factor for further speed gain.

Originally Posted by E6600

Now the argument is, which direction should the industry follow. The numbers I put are just an example, you can assume 4 cores each clocked at 4GHz vs 80 cores each clocked at 200 MHz.

There is no argument here. The reason the industry is going with multi-core chips is simply because they ran into barrier dealing with the physics of running at high clock speeds. Intel hasn't passed the 4 Ghz mark yet and their closest attempt was a Pentium 4 running at 3.8 Ghz released nearly three years ago.

There is a point where putting more cores into a system will not get any improvement in speed. A system here at home has 112 threads spread over 48 processes (both applications and the OS). An 80 processor system at 200 Mhz would be utterly useless for that environment. Only 5 out of those 48 processes are actively running and the number of concurrently running threads is far less than 80. At 200 Mhz, performance would be pitiful despite a task being able to monopolize an entire CPU. Four processors at 4 Ghz would have the overhead of switching between tasks but that is a minor concern. The tasks that I run would actually get more processing power allocated than what is possible on an 80 core, 200 Mhz system.

[Spindel]

I realized as I wrote this that this entire post is more or less only aimed at Power

Shouldn't the industry take a different turn instead of just boosting clock rates and number of cores?

I mean if we look at the standard desktop market and go back to the P4, the P4 is idle 90-92% (don't remember the exact fraction but it was somewhere around this) of the time when working because of cache errors or misses. This was the price for reaching the high numbers in frequency (since for a long time the only sale argument for a new CPU was the frequency).

This was the reason (afaik) why AMD abandoned the MHz war with the Athlon XP CPU line and started to use the performance rating (PR) system instead. They tried to reduce the idle time (while under load) for the CPU to gain performance in that way. (and afaik this is what intel have started to do with the Core series also).

Now my guess is that you probably know more of this than I do Power so please correct me if I'm totally off. But I'm pretty sure it is somewhere along the lines of what I said.

But also afaik there is a limit to how much you can reduce this on the x86 architecture because of the way it simply is set up.

So my thinking is more along the lines of: Shouldn't the PC industry try to move away from the 30 years old x86 architecture in favour of a newer more efficient architecture?

[chedabob]

Currently if you go above 2 cores, you won't really see the benefit. Not that many applications support dual core, let alone quad. There is only so much the O/S can do, the rest has to be written into the program itself. Some of the higher-end applications such as Maya and 3D Studio Max support quad core, but the majority of applications rely heavily on the O/S to divide the processing up.

[E6600]

Originally Posted by Power666

There is not right or wrong answer because it entirely depends on the task at hand. Additionally, other subsystems like the memory controller, coherency traffic, IO etc. play a huge role in which design philosophy can scale for a given task.

For 'general' tasks and performance, clock speed is the still the major factor. Multithreading doesn't lend itself to every task. Even with code that can be multithreaded, there are often parts dependent on a serial portion. This serial portion becomes the limiting factor for further speed gain.



There is no argument here. The reason the industry is going with multi-core chips is simply because they ran into barrier dealing with the physics of running at high clock speeds. Intel hasn't passed the 4 Ghz mark yet and their closest attempt was a Pentium 4 running at 3.8 Ghz released nearly three years ago.

There is a point where putting more cores into a system will not get any improvement in speed. A system here at home has 112 threads spread over 48 processes (both applications and the OS). An 80 processor system at 200 Mhz would be utterly useless for that environment. Only 5 out of those 48 processes are actively running and the number of concurrently running threads is far less than 80. At 200 Mhz, performance would be pitiful despite a task being able to monopolize an entire CPU. Four processors at 4 Ghz would have the overhead of switching between tasks but that is a minor concern. The tasks that I run would actually get more processing power allocated than what is possible on an 80 core, 200 Mhz system.

I get your point, their is an issue of how much a process can be multithreaded and how many threads will run at any given time, I agree with you on this, but imagine if we could design an OS and programs that can achieve a high level of a parallelism such that no core is idle and their is a minimum dependency on serial tasks. I know current OSs and programs are not, most of them run only a couple of threads and at some point while running the threads will be waiting for a specific action to finish.
Think back to the supercomputers of the 80s and 90s, instead of focusing on raw speed, what if we focus more on achieving parallelism at the software and hardware level, increase the throughput per clock cycle by using vector processing (a la Cray-2).

I'm not arguing against your point, I agree with you on it, to me I'd rather have less core with more power, I've thrown this argument to see if any agree with my professor and think there is another way for processors to evolve.

Originally Posted by Spindel

So my thinking is more along the lines of: Shouldn't the PC industry try to move away from the 30 years old x86 architecture in favour of a newer more efficient architecture?

Actually the x86 evolved and isn't the same old architecture, if anything the new x86 processors are more RISC than CISC, since every CISC instruction is actually translated into a number of RISC instructions. The x86 isn't pure RISC due to backward compatibility constraints, but every feature in the RISC design is in the x86 architecture one way or another.

What I really want to see, is vector processing, I know the x86-64 processors and even before have some special purpose vector instructions (SIMD), there's also the CELL which focuses on vector processing.

The thing is, it's like what power666 said, the physical limit is almost reached, after a decade or so they wont be able to shrink the transistors any more, they are already approaching atomic level, an atom has a diameter of 0.06 to 0.6 nanometers.

Intel scientists find wall for Moore's Law:

http://news.zdnet.com/2100-9584_22-5112061.html

[Gnu]

Originally Posted by chedabob

Currently if you go above 2 cores, you won't really see the benefit. Not that many applications support dual core, let alone quad. There is only so much the O/S can do, the rest has to be written into the program itself. Some of the higher-end applications such as Maya and 3D Studio Max support quad core, but the majority of applications rely heavily on the O/S to divide the processing up.

There are far more applications out there that support multi threading than you might think...

And if you do a lot of multitasking, you'll certainly see the benefit of having more than two cores.

When I first got a dual core computer a couple years ago, people told me I was wasting my money, and using the same argument that "most applications don't support multiple cores", yet looking at my hardware performance monitor, I could see both cores pretty much always being utilized quite well. I don't really know where this same old argument keeps coming from...

[E6600]

Originally Posted by Gnu

There are far more applications out there that support multi threading than you might think...

And if you do a lot of multitasking, you'll certainly see the benefit of having more than two cores.

When I first got a dual core computer a couple years ago, people told me I was wasting my money, and using the same argument that "most applications don't support multiple cores", yet looking at my hardware performance monitor, I could see both cores pretty much always being utilized quite well. I don't really know where this same old argument keeps coming from...

who can say no to more? I don't claim to have the fastest CPU in the market, but every now and then my CPU is fully utilized I suffer from delay when running multiple programs, like playing a movie while rendering in Mental Ray.

[chedabob]

Originally Posted by Gnu

I don't really know where this same old argument keeps coming from...

I had an article a while back that showed dual core vs. quad core performance, and most of the time Quad core was less than dual. I've posted it a few times, I think at least once in this thread.