Would a 10GHz CPU be as fast as a 2.5 GHz quad core CPU?

[The_Absolute]

Would a single 10GHz CPU be as fast as a quad-core CPU with each core running at 2.5 GHz?

I would imagine that the caches on the Four processor cores would give it a performance advantage over the 10GHz single core.

The way CPU manufacturers went after the days of highly-clocked processors (Pentium 4) was to add more processors (cores) onto a single die, instead of having just One really fast core. Running a CPU at a speed of 10 GHz would be impossible with conventional cooling, and would probably require liquid nitrogen.

Does adding more cores onto a CPU die increase thermal output? Is this why the more cores, the smaller the die, and lower stock clock speeds? Such as the next generation 6-core processors from AMD and Intel being 32nm instead of 45nm like the quad cores.

 

[Borek]

I don't think these are directly comparable. A lot depends on the task and how easily it can be parallelized. On a general level I think one core has to spend much more time on switching tasks, so I would expect it to be slower, but I can be wrong.

 

[dE_logics]

If multithreading is used, the quad core will win. If not, you'll get 1/4th the performance.

 

[Borek]

If multithreading is used, the quad core will win. If not, you'll get 1/4th the performance.

Where did you get 1/4th from? OP asked about 10 GHz single core vs 2.5 GHz quadcore - in terms of raw GHz that's the same. Sure, it is not the same, but 1/4th seems wrong.

 

[Blenton]

Well without any multi-threading your quad core will only run as fast as its fastest core, so you're really only working with one 2.5GHz core (1/4).

 

[Borek]

Ah, OK. For some reason all the time I was thinking about a multithreading environement, not about a single thread.

 

[dE_logics]

Ah, OK. For some reason all the time I was thinking about a multithreading environement, not about a single thread.

Yeah, I was comparing the 2 you know.

 

[Negatron]

Does adding more cores onto a CPU die increase thermal output?

It decreases thermal output compared to equivalent serial execution. This is one of the big reasons for parallel computing. Power output scales linearly with the number of features, but has, if I understand the problem correctly, cubic scaling with feature clock. In the best case scaling would be quadratic. This is a result of the voltage/power law plus linear power scaling with frequency, assuming that voltage would need to double when doubling frequency to maintain a sufficiently low noise floor.

Maximally clocked devices would still be very useful for running the main application thread so behemoth CPU systems will always have their purpose, but the vast majority of a modern work load will run on massively parallel devices.

 

[The_Absolute]

It decreases thermal output compared to equivalent serial execution. This is one of the big reasons for parallel computing. Power output scales linearly with the number of features, but has, if I understand the problem correctly, cubic scaling with feature clock. In the best case scaling would be quadratic. This is a result of the voltage/power law plus linear power scaling with frequency, assuming that voltage would need to double when doubling frequency to maintain a sufficiently low noise floor.

Maximally clocked devices would still be very useful for running the main application thread so behemoth CPU systems will always have their purpose, but the vast majority of a modern work load will run on massively parallel devices.

Most of the above quote is greek to me, so what you're saying that a Quad-core 2.5 GHz CPU running the same application fully multi-threaded throughout all of it's cores, would have equal, or more performance than, let's say a 10 GHz single core CPU? With producing low enough levels of heat to be cooled with a simple heatsink and fan?

The heat produced by a processor running at 10 GHz would need to be cooled with liquid nitrogen.

 

[edpell]

I will venture my guess. Processors are fast but off-chip memory access is slow. One way to deal with this is to have two threads running on your core. When one stalls (needs info from external memory) the other thread quickly swaps in and continues doing useful work. Of course it take circuits and space and power to keep the state of the second thread local and alive. More threads to remember more hardware more space, power, delay, etc.

Let's say the 4 cores running at 2.5GHz has two threads per core. Let's say the 10GHz core has eight threads then the two would do approximately the same amount of computation per unit time. If having eight copies of every register did not so increase the layout area and hence interconnect distances that things slowed down.

 

[edpell]

The heat produced by a processor running at 10 GHz would need to be cooled with liquid nitrogen.

power goes as frequency x capacitance x voltage squared x number of cores

so 1 core 10GGHz should have the same power as 4 cores 2.5GHz

 

[Negatron]

Quad-core 2.5 GHz CPU running the same application fully multi-threaded throughout all of it's cores, would have equal, or more performance than, let's say a 10 GHz single core CPU?

Disregarding the subtleties of the problem, it can be said to be more or less equal, yes. But the quad core as mentioned would do the task far more power efficiently.

But it's important to keep in mind that this isn't at all a statement in favor of multi-core processors. If I were to wager, I'd say the consumer multicore race will go nowhere and for the most part it's already dead. CPUs are excellent single threaded general purpose execution units, but for massively parallel applications they fail terribly with their general inefficiency. Even multi-threading and SSE are a joke compared to GPUs.

Although 4 cores are more efficient than 1 core four times their speed, it will nevertheless be of much value for consumer applications to have one thread which executes as quickly as possible, where 10GHz units would be very helpful, and all parallel work will be offloaded by the main thread for hardware which is better fit for that purpose. This makes a whole lot more sense to the developer in the sense that rather than having to program their parallel code for two different types of devices and somehow distribute the work between them, they will only program for the devices which handles the work most efficiently. This is why multi-core processors are underutilized in modern games and applications, despite having been around for many years now. Developers want one application thread and one logical work unit. This is the programming paradigm a combination of 1 CPU core and an arbitrary number of GPUs deliver. The inefficiency of making one incredibly fast CPU core instead of multiple slower cores is overcome by the fact that the GPU will be the dominant consumer of power, so it becomes justifiable to sacrifice a bit of power efficiency to maintain an elegant programming model. Intel has a lot of resources invested in keeping the multi-core programming paradigm intact, primarily by driving false consumer demand, to the detriment of the world, but they can only sustain this for so long.

 

[kote]

The performance difference is in the partitioning. A four core CPU can only devote 25% of its total power to any single process. A single core CPU can devote up to 100% of its power to anyone process, or it can devote 25% each to four processes etc. It takes a lot of programming and optimization to split up a process into equal parts, so there's a lot of development involved in trying to get software to take advantage of all available cores.

 

[digitalblggr]

http://digitalblggr.blogspot.com/2010/01/on-parallelization.html