throughput gains of a "typical" single thread

[br111an]

I've read that PC architecture (Intel or AMD) clock rates have not risen at the historical rate since about 2005 so I'm wondering what single threaded throughput gains there have been over the past 7 years and how that compares to the past 10 or 15 years. Specifically, I'd be interested in seeing typical application throughput (say, Windows and a single threaded benchmark) graphed and seeing if there's a kink around 2005.

I would guess there would be gains even in the past 7 years due to:

1. memory rates, example: DDR3 being faster than DDR2
2. faster bus rates on the motherboard
3. SSD
4. additions to the x86-64 instruction set
5. slight gains in the clock rate, albeit not as great as in the 1990s

I'm not sure what else. Of course multi-threaded applications and multi-threaded operating system operations to take advantage of multiple cores are potentially a big variable and in my mind doesn't really deserve to be classed in the same category as any of the above because they are actually software improvements more than anything else. Hmmm... the audacity to say adding 3 cores to 1 is a software improvement. A little ironic, but really adding cores is sort of like moving the goal posts and it still doesn't work without software changes (which might not be feasible for some things).

I'd appreciate any perspective on this or if you've seen published examinations on this question. I did read one essay from a Microsoft compiler expert on this subject but that's fairly dated now. At the time that article came out the main point was throughput gains are harder to come by if you don't re-write for multiple threads. I wonder if things have been as bad as people used to expect.

[randy1]

I've read that PC architecture (Intel or AMD) clock rates have not risen at the historical rate since about 2005 so I'm wondering what single threaded throughput gains there have been over the past 7 years and how that compares to the past 10 or 15 years. Specifically, I'd be interested in seeing typical application throughput (say, Windows and a single threaded benchmark) graphed and seeing if there's a kink around 2005.

I would guess there would be gains even in the past 7 years due to:

1. memory rates, example: DDR3 being faster than DDR2
2. faster bus rates on the motherboard
3. SSD
4. additions to the x86-64 instruction set
5. slight gains in the clock rate, albeit not as great as in the 1990s

I'm not sure what else. Of course multi-threaded applications and multi-threaded operating system operations to take advantage of multiple cores are potentially a big variable and in my mind doesn't really deserve to be classed in the same category as any of the above because they are actually software improvements more than anything else. Hmmm... the audacity to say adding 3 cores to 1 is a software improvement. A little ironic, but really adding cores is sort of like moving the goal posts and it still doesn't work without software changes (which might not be feasible for some things).

I'd appreciate any perspective on this or if you've seen published examinations on this question. I did read one essay from a Microsoft compiler expert on this subject but that's fairly dated now. At the time that article came out the main point was throughput gains are harder to come by if you don't re-write for multiple threads. I wonder if things have been as bad as people used to expect.

Multicore processors are a massive hardware improvement. And them the software that takes advantage of the new architecture. Also look into the benefits of ever decreasing circuitry which provides power benefits and less heat dissapation which is also a benefit.

[br111an]

Ummm... it looks like you didn't read the question. Multi threading makes for an apples to oranges comparison. Low power and less heat don't help with throughput.

[bryant.english]

Multicore processors are a massive hardware improvement. And them the software that takes advantage of the new architecture. Also look into the benefits of ever decreasing circuitry which provides power benefits and less heat dissapation which is also a benefit.

[robertlempriere]

I think it is fair to say that during the past few years single threaded performance has not been the focus of the industry.
As randy1 pointed out already, multi-core is the way to go.
Just look at the recently announced Intel MIC / Knights Corner / Xeon Phi - this beast is running 50+ Pentium 1 cores on a single expansion board.

Further, and this is nothing new, performance improvements don't only come clock speed increments, look at SIMD, MMX etc etc, new cleverness within the CPU which in turn requires a new compiler to take advantage of said enhancements.

If you spend 5 minutes and look at the Intel site, you can easily see at a high level the major differences between the 2nd and 3rd generations of the Core i7 CPU for instance;

* Larger Cache
* Reduced Power consumption
* movement from 32nm process to 22 !!!!!! this is freaking huge
* PCI 3 from 2.0
* Smaller die!!!!


Then, look at the difference in launch dates:
Launch Date Q4'11 Q2'12

That's what has changed in 3 quarters in the Core i7 .. I'd say that's pretty impressive wouldn't you?

[randy1]

I think it is fair to say that during the past few years single threaded performance has not been the focus of the industry.
As randy1 pointed out already, multi-core is the way to go.
Just look at the recently announced Intel MIC / Knights Corner / Xeon Phi - this beast is running 50+ Pentium 1 cores on a single expansion board.

Further, and this is nothing new, performance improvements don't only come clock speed increments, look at SIMD, MMX etc etc, new cleverness within the CPU which in turn requires a new compiler to take advantage of said enhancements.

If you spend 5 minutes and look at the Intel site, you can easily see at a high level the major differences between the 2nd and 3rd generations of the Core i7 CPU for instance;

* Larger Cache
* Reduced Power consumption
* movement from 32nm process to 22 !!!!!! this is freaking huge
* PCI 3 from 2.0
* Smaller die!!!!


Then, look at the difference in launch dates:
Launch Date Q4'11 Q2'12

That's what has changed in 3 quarters in the Core i7 .. I'd say that's pretty impressive wouldn't you?

Exactly right.

[mid_gen]

Do your own homework :P

[robertlempriere]

Do your own homework :P

I fear you may be right in assuming this is homework, anyway, it's a topic I'm interested in

[mid_gen]

Yes I'd be interested to know the answer, but not enough to do the research for someone else too lazy to do it themselves

Ok I lied, I was curious enough. First hit on google :

A Look Back at Single-Threaded CPU Performance

[br111an]

That's a good link. Thanks mid_gen.

Robertlempriere, I never denied multi-threaded is the way to go. I'm interested in single threaded performance, much of which spills over to multi-thread. And as for listing the enhancements, well, I would just say I'm more interested in the throughput plot versus time because that's quantified and it's not the kind of thing I would expect Intel to put on their website.