Upgrading 2009 MP: frustrating choices

[666sheep]

There are no reasons except that my computer feels slow. I used to change my mac pros and power macs every 3-4 years, and now I've had this one for close to five. It works fine, just getting slow.

Slow in what exactly? What do you use your MP for?
Post how much RAM do you have and OS version you're running.
I have W3570 currently in my 4,1 and 16GB 1066 MHz (single channel) RAM + overclocked AMD 7870XT 2GB.
If you'd me want to do some benchmarks/tests, PM me or I can post results here if you prefer.

 

[Loa]

Hello,

I know that Moore's law doesn't equate with actual performance, but for a long long while we've seen a lot of great performance gains. Maybe I'm wrong, but I'm not seeing it anymore. Multiple cores is, firstly, somewhat akin to cheating if we equate it to a "single" computer's performance; secondly, nothing we couldn't have done years and years ago; and thirdly not useful for most types of tasks. The fact that we now *need* to do it is a symptom that single thread performance just isn't rising as fast as it used to. Many people won't ever benefit from more than 2 cores, and very few from more than 4. On the other hand, *everybody* would benefit from a 6GHz CPU.

Could we use geekbench 2 to plot the evolution of single thread performance ever since the dawn of, lets say, Power Macs? My *guess* (it's really a feeling) is that the exponential growth curve (of actual performance, not MHz/GHz) is slowing down.

---

666sheep: my Mac serves a very broad array of jobs and tasks, from usual basic computer stuff (iwork, web, music, etc...), photography stuff (12mpx raw edition), some video (imovie), some handbrake encoding, and some other things.

With 16GB RAM (very few page outs), a vertex 4 512GB on sata3 card, a RAID0 for my main data and a 6870, I don't see anything else (except CPU) that would make a significant difference.

Thanks for the offer of testing, but there isn't a single specific thing I really want to speed up. I'm looking for a general boost, and I think that a 2.66 -> 3.2 should do the trick. (I'm also planning on doing a 100% fresh install as soon as I have the time.)

Loa

 

[Inconsequential]

There is no guessing.

IPC gains of the past are gone. You get small improvements each generation, in the region of 10%. AVX and the suchlike help specific tasks with larger performance jumps.

You can't just wack up the clock speed as physics gets in the way.

The bottom line is, what you want simply can't happen without a step change in tech, one that isn't going to happen the foreseeable future.

The purple line in the graph you posted is the key to single thread performance.

I just wish Lightroom was OpenCL enabled and it's CPU usage is quite frankly a joke too.

 

[Loa]

what you want simply can't happen

Happens often, that... ;-)

It's looking pretty clear that my inexpensive upgrade to 3.2GHz is the only thing that will help me wait for the MP 7,1's (hopefully) 4GHz few-cores base model. (Luckily for me there's still a lot of performance headroom in my car, so the money I could have spent on the nMP will go there instead...)

Thanks for the help.

Loa

 

[deconstruct60]

IPC gains of the past are gone. You get small improvements each generation, in the region of 10%. AVX and the suchlike help specific tasks with larger performance jumps.

The single IPC is actually a highly coupled property of software not hardware. Hardware can mask/ignore how much Instruction Level Parallelism (ILP) get out of the software but not going to go past what is present in the software.

The fact that levels out at 10 is not surprising at all. There were a old Digital Western Research Lab report ILP that looked at max ILP and it didn't go much past 10 on average.

You can't just wack up the clock speed as physics gets in the way.

Rapidly executing no-ops doesn't do much anyway. ( If leave the rest of the system behind there isn't going to be productive work to do. )

The purple line in the graph you posted is the key to single thread performance.

Correct. What this is plotting is increasingly a smaller subset of the overall CPU package. People don't use subsets of a CPU packages... they use the whole thing.

 

[Loa]

Hello,

Is there a line in the graph that illustrates / explain why the i5 3.4GHZ is faster at single thread than the nMP's 3.7GHz? If not, what's the explanation? Is it a matter of design decisions on Intel's part? I'm just amazed that a lower clocked i5 is faster than a higher-clocked current Xeon...

Loa

 

[deconstruct60]

Multiple cores is, firstly, somewhat akin to cheating if we equate it to a "single" computer's performance;

It is not. The days of single tasking, single thread OS and Apps ( DOS and the like) are long over. The norm now are apps with mulitple threads and multiple apps all running at the same time.

If pop open activity monitor you'll find that processes with just one thread the minority. The majority are going to be low level OS tasks that are highly limited. Active worker threads are low but the notion that the bulk of apps are all still "completely single file" is stale and disconnected from reality.

The straightforward productive way to engaged the increased number of transistors available is to given these multitasking instruction streams (yes plural) their own instruction queue when possible. If leave one core and mulitple tasks basically taking time away from individual tasks because they'll need to be swapped out to wait their turn.

The fact that we now *need* to do it is a symptom that single thread performance just isn't rising as fast as it used to.

It is not raising as fast in part because the hardware isn't doing as big of a job of masking out the ILP in the software. Now that the hardware is more competent there aren't going to be as many low-hanging-fruit gains.

Many people won't ever benefit from more than 2 cores, and very few from more than 4.

Horse hooey. This is a fun house mirror view of viewing through single x86 distortion funnel. There is far more to systems than just x86 cores.

On the other hand, *everybody* would benefit from a 6GHz CPU.

Not. Just cranking one subcomponent doesn't necssarily bring broad , effective benefits. A data straved single core isn't going to do much. Second, most folks have multiple dimensions of requirements than just drag racing speed.

Could we use geekbench 2 to plot the evolution of single thread

Nobody uses geekbench to get anything useful done except for the spec porn crowd. It isn't a system benchmark. It is a subcomponent benchmark. People use systems, not subcomponents.

My *guess* (it's really a feeling) is that the exponential growth curve (of actual performance, not MHz/GHz) is slowing down.

The users/humans are far limiting than the hardware. Folks are at least as much griping about cheaper than faster.

 

[Loa]

There is far more to systems than just x86 cores.

Yes, but it doesn't change the fact that *most* people won't benefit from going from 4 to 6+ cores.

A data straved single core isn't going to do much.

Indeed, but that is a bad argument: you're making your point from the worst case scenario. Most systems are *not* in that situation. In order to consider a 6GHz as a negligible gain over what we have today, everything else in the system would have to be maxed out. It's not.

Faster GHz is a gain all around, while more cores is not a benefit for everybody.

If I could fit a 64 core, 2.66GHZ CPU in my current set-up, it would make an insignificant performance increase (for me). But if I could fit a 6GHz quad in my system, it would make a very significant performance increase in a vast majority of tasks.

I know that it would not make a 225% performance increase across the board as there are other bottlenecks in my system (mainly SSD, RAM and all the interconnects).

But to argue that a 6GHz CPU would not equate to a significant performance increase is a specious argument.

Loa

 

[deconstruct60]

Yes, but it doesn't change the fact that *most* people won't benefit from going from 4 to 6+ cores.

Addition more graphics cores will make more folks happy. Especialy when "Moorse's Law" picks up again and the pixel density on their screens goes up.

You are mired in a mindset of looking just at the x86 core. That isn't even close to what the whole system is composed of.

Indeed, but that is a bad argument: you're making your point from the worst case scenario. Most systems are *not* in that situation. In order to consider a 6GHz as a negligible gain over what we have today, everything else in the system would have to be maxed out. It's not.

Not sure what alternative universe you are living in but simply not the case. Vast majority of systems can flip on SMT(symteric multithreading what Intel calls hyperthreading) with little to no impact. That is precisely because the function units inside the cores are UNDERUTILIZED. If they were fully utilize there wouldn't be room for a 2nd (or 3rd , etc) thread's execution stream.

So with cores with commonly underutilized function units and crank the clock higher the underutlization just goes up. Not down.

Everything else in most systems is at the rate at which they are at. Their progress rates aren't going to be that much faster than the CPU. In the real world have to deal with what you got. Right now (and over previous decades) the CPU is been a bit ahead of the rest of the system elements in terms of speed/throughput. Often the CPU has to wait. Cranking the CPU higher when it is ALREADY waiting periodically doesn't do squat but make a better space heater.

 

[Loa]

Last attempt.

Everybody would benefit from doubling the clock speed of all the cores they currently have on their system. (That benefit wouldn't be 100%, but it would be significant.)

NOT everybody would benefit from doubling the amount of cores they currently have on their system.

Simple as that.

I'm not saying that many cores is bad: it is good. But as you say, it's just one aspect of the system. There are other things we can improve, on the CPU, other than adding more cores.

Loa

 

[Thessman]

Faster GHz is a gain all around, while more cores is not a benefit for everybody.

Loa

While that statement is undoubtedly correct, what the people here intend to convey is that since the physical limits of upping the clock have been reached (you can always supercool your CPU and overclock it to 5.5-6 GhZ) until we have a completely new technology, we can utilize multicore technology to accelerate most if not all of the modern day computing tasks by multithreading them.

 

[Loa]

While that statement is undoubtedly correct, what the people here intend to convey is that since the physical limits of upping the clock have been reached (you can always supercool your CPU and overclock it to 5.5-6 GhZ) until we have a completely new technology, we can utilize multicore technology to accelerate most if not all of the modern day computing tasks by multithreading them.

I know, but I think it's pretty sad, and that multiple cores is a compromise.

I'm still wondering why a i5 with a lower clock speed is faster at single thread performance than the nMP's base quad.

Loa

 

[deconstruct60]

Last attempt.

Everybody would benefit from doubling the clock speed of all the cores they currently have on their system. (That benefit wouldn't be 100%, but it would be significant.)

You can keep repeating it but it isn't the case. Haswells' adding of support for trace cache decode and tranactional memory is likely a far more effective transistor utilization than than the simplistic cranking of clock speed and boosting increasingly ineffective caches to larger sizes.

Net burst (Pentium 4) failed.
IBM Power is actually down from 5-6GHz over the last couple of iterations.

There is a substantial number of folks who don't want power ineffective improvements. In fact most of the market doesn't (including the substantive segments of the server market ). So your "everybody" is blanantly false. The CPU market has already demonstrated this a couple of times.

NOT everybody would benefit from doubling the amount of cores they currently have on their system.

Doubling the wrong kind cores? No it wouldn't. Doubling the appropriate ones sure. It is already happening if you would open your eyes and look at what is being deployed in 2014.

Simple as that.

Myopically hiding in a x86 cores are only thing that matters hole is simplistic; not simple.

There are other things we can improve, on the CPU, other than adding more cores.

And Intel is doing that. Some of that is bring more stuff closer ( and having substantially higher interconnect speed in the process ) and that most of the x86 core work has nothing to do with burping instructions down an over extended pipeline faster.

But after 15-20 years optimizing the x86 instruction processing process there aren't going to be many completely unoptimized segments left. It is old and frankly the constraints are going to build up about as much as going to compose workarounds for old ones.

 

[Inconsequential]

Hello,

Is there a line in the graph that illustrates / explain why the i5 3.4GHZ is faster at single thread than the nMP's 3.7GHz? If not, what's the explanation? Is it a matter of design decisions on Intel's part? I'm just amazed that a lower clocked i5 is faster than a higher-clocked current Xeon...

Loa

The purple IPC line.

Haswell has the ability to do more work per clock cycle than IB-E can, either by x86 core improvements or the way it handles data.

Hence why the iMac is the fastest single threaded mac ever.

I'm not completely up-to-speed with the inner workings of a CPU but that is generally the gist of it.

 

[666sheep]

666sheep: my Mac serves a very broad array of jobs and tasks, from usual basic computer stuff (iwork, web, music, etc...), photography stuff (12mpx raw edition), some video (imovie), some handbrake encoding, and some other things.

With 16GB RAM (very few page outs), a vertex 4 512GB on sata3 card, a RAID0 for my main data and a 6870, I don't see anything else (except CPU) that would make a significant difference.

3570 is the sweet spot if you don't need more than 4 cores. Unless you'd get a very good deal on W3580 (sometimes they go on eBay for closer to $200 than $300).

 

[Pjotr28]

Tired of waiting for a new Mac Pro I decided to pimp up my 2009 quad core last year. With a 6 core 3.33ghz, a 7950 vid card and a SSD it feels like new again. Shopping around at ebay it costed me around 800 euro's. Main use of the machine is video editing.

I'm happy with this decision for now. Better this than spend 4k on a new machine.

 

[Loa]

Thanks Concorde Rules.

I'll keep an eye on ebay deals, as I can't find a canadian store that sells these used CPUs.

Loa

 

[funwithdesign]

Thanks Concorde Rules.

I'll keep an eye on ebay deals, as I can't find a canadian store that sells these used CPUs.

Loa

I picked up a x5590 (same as w3580) for less than $200cdn from ebay.ca

 

[Loa]

Hello,

Can I put the X CPUs in a single CPU MP?

Loa

 

[funwithdesign]

Hello,

Can I put the X CPUs in a single CPU MP?

Loa

Yes. Dual chips can be put in single processor systems, not the other way around though. And I believe it's the '5' that makes it a dual chip. The X is related to TDP. Either way, no issues at all.

 

[Loa]

Thanks! more choices!

Loa

 

[ozbimmer]

Hi Loa,

I think the following list of CPU upgrade might be of help to your quest I got the info from OWC.

3.20GHz Quad-Core 'Nehalem' W3565 (4-Cores, QPI: 4.8GT/s)
3.20GHz Quad-Core 'Nehalem' (Intel W3570, 4-Cores, QPI: 6.4GT/s)
3.33GHz 6-Core 'Westmere' (Intel W3680, 6-Cores, QPI: 6.4GT/s)
3.46GHz 6-Core 'Westmere' (Intel W3690, 6-Cores, QPI: 6.4GT/s)
Two x 2.53GHz Quad-Core 'Westmere'(Intel E5630,8-Cores,QPI:5.86GT/s)
Two x 2.40GHz Quad-Core 'Westmere'(Intel E5620,8-Cores,QPI:5.86GT/s)
Two x 2.66GHz Quad-Core 'Westmere'(Intel E5640,8-Cores,QPI:5.86GT/s)
Two x 2.66GHz 6-Core 'Westmere' (Intel X5650, 12-Cores, QPI: 6.4GT/s)
Two x 2.93GHz 6-Core 'Westmere' (Intel X5670, 12-Cores, QPI: 6.4GT/s)
Two x 3.33GHz 6-Core 'Westmere' (Intel X5680, 12-Cores, QPI: 6.4GT/s)
Two x 3.46GHz 6-Core 'Westmere' (Intel X5690, 12-Cores, QPI: 6.4GT/s)