AMD Zen supports CMT and SMT

[ShintaiDK]

My response was for ST, IPC alone doesn't mean anything.

So its as misleading as to claiming an Atom is not that far from Kaveri in ST.

But I guess you already knew that.

 

[monstercameron]

I think most Scientists / Researchers would disagree with you.

For scientific computing, most researchers seem to prefer all 8 projects running/finishing at a consistent speed (FX-8320e) -- versus 4 projects finishing quickly, with 4 projects dragging along like a slug on the "hyper" threads (Core i7).

My alma mater just purchased a cluster of water cooled FX's instead of i7's because the scientific computations were completed so erratically on the tested i7's.

Just saying.... Intel isn't the end all and be all that the fanboys on this forum make it out to be. CMT and SMT combined has the potential to be very good for many specialized tasks.

Can you provide soon proof.

 

[MiddleOfTheRoad]

This can not happen.
There is a thing called thread migration,each time a thread is beeing rescheduled it gets assigned to a new core(real or virtual) no matter how many or few threads you run each and everyone will get the same amount of cpu time so there is no way for them to finish at erratically different times.
Even if they went out of their ways to use affinity and force each process/thread to a different core,the scheduler would still give each thread an equal amount of cpu time.

Nice.... Maybe you should get a Ph.D so you can tell all the Math Ph.D's they don't know what they are talking about related to integer calculations. (Yeah, you know the people who actually tested the actual hardware from both companies before making the larger buying decision).

You clearly never ran scientific computing on both CMT and SMT setups -- because they do execute projects in a very different fashion.

 

[MiddleOfTheRoad]

Can you provide soon proof.

I wouldn't need to. Just install BOINC on both a SMT and CMT processor, select a single project (where all tasks are identical in size) -- you can visually see the difference looking at the progress bars.

A CMT finishes all the projects roughly in identical time -- the SMT is all over the map (with some projects finishing way faster than the CMT machine -- with several lagging behind).

 

[monstercameron]

I wouldn't need to. Just install BOINC on both a SMT and CMT processor, select a single project (where all tasks are identical in size) -- you can visually see the difference looking at the progress bars.

A CMT finishes all the projects roughly in identical time -- the SMT is all over the map (with some projects finishing way faster than the CMT machine -- with several lagging behind).

I'm talking about the orders.

 

[waffleironhead]

I wouldn't need to. Just install BOINC on both a SMT and CMT processor, select a single project (where all tasks are identical in size) -- you can visually see the difference looking at the progress bars.

A CMT finishes all the projects roughly in identical time -- the SMT is all over the map (with some projects finishing way faster than the CMT machine -- with several lagging behind).

This is very true in applications that actually use all of the processor resources.
For example running llr tests in primegrid, you are better off turning off HT on Intel processors as 8 tasks running on all 8 available threads in a 4c/8t system will take longer to complete that running 4 tasks plus 4 tasks with tht HT turned off.
On the other hand, running sieving tasks, you can leave HT on and it will be faster in the long run using the ht capability.

 

[AtenRa]

So its as misleading as to claiming an Atom is not that far from Kaveri in ST.

But I guess you already knew that.

You wanted to say that an ATOM ST performance is not that far way from Jaguar/Puma

 

[ShintaiDK]

You wanted to say that an ATOM ST performance is not that far way from Jaguar/Puma

No, I just cherry pick like you and its true.

But cherrypicking doesnt give any better products.

 

[MiddleOfTheRoad]

This is very true in applications that actually use all of the processor resources.
For example running llr tests in primegrid, you are better off turning off HT on Intel processors as 8 tasks running on all 8 available threads in a 4c/8t system will take longer to complete that running 4 tasks plus 4 tasks with tht HT turned off.
On the other hand, running sieving tasks, you can leave HT on and it will be faster in the long run using the ht capability.

Thank you.

CMT was designed to even out the workload to all the cores -- which does hold back single threaded performance. By contrast, SMT plays "favorites" and pushes certain threads much faster than others.

In heavily multithreaded programs with progress bars for each thread like BOINC (sitting my FX 8320 next to my i7 4770K) -- you can literally see how differently they process the data. Every project processes at the same speed on the FX -- the i7 rockets through some of the projects like lightning, while some projects kinda stagnant.

 

[AtenRa]

No, I just cherry pick like you and its true.

But cherrypicking doesnt give any better products.

AMD Puma IPC is higher than ATOM Baytrail, yet ST perf is very close because BayTrail reach higher Frequency.

Like it or not IPC alone means nothing.

 

[MiddleOfTheRoad]

I'm talking about the orders.

I was simply referring to what I've seen in real world performance in heavily multithreaded applications.

 

[monstercameron]

I was simply referring to what I've seen in real world performance in heavily multithreaded applications.

No about the school that ordered the fx 9590s

 

[mrmt]

I wouldn't need to. Just install BOINC on both a SMT and CMT processor, select a single project (where all tasks are identical in size) -- you can visually see the difference looking at the progress bars.

Then BOINC on shoestring clusters is the ultimate scientific application? Is this your use case for CMT processors?

 

[Phynaz]

Thank you.

CMT was designed to even out the workload to all the cores -- which does hold back single threaded performance. By contrast, SMT plays "favorites" and pushes certain threads much faster than others.

In heavily multithreaded programs with progress bars for each thread like BOINC (sitting my FX 8320 next to my i7 4770K) -- you can literally see how differently they process the data. Every project processes at the same speed on the FX -- the i7 rockets through some of the projects like lightning, while some projects kinda stagnant.

You still don't get that BOINC is a framework, and isn't very thread friendly at all.

 

[AtenRa]

Then BOINC on shoestring clusters is the ultimate scientific application? Is this your use case for CMT processors?

At least he gave an example where CMT is better, I havent seen your case where SMT is better than CMT

 

[MiddleOfTheRoad]

No about the school that ordered the fx 9590s

They didn't -- they bought FX-8320e processors. I seriously doubt anyone is going to want to run 9590's for 24 hours a day doing math calculations.... Those things generate way too much heat / power draw to economically cluster them IMO.

When the nation's second largest university (UCF) buys 6 desktops for math research, it doesn't exactly make headlines. I don't believe there was any media coverage.

 

[MiddleOfTheRoad]

You still don't get that BOINC is a framework, and isn't very thread friendly at all.

BOINC is a middleware system for grid and volunteer computing. Specific projects can be weighted heavily towards integer or floating point performance -- The people who run Intel processors generally join the projects that are floating point intensive (SETI), the guys with AMD stuff generally favor the integer projects (World Community Grid).

I've been a member for 4 years -- I think I have pretty good understanding by now.

It depends on which project you decide to join, some projects are very thread friendly.
Last time I looked there were over 50 different projects available to join (Math, Astrology, Medical Research, etc).

AMD still does well in this general discipline -- the Second Most Powerful SuperComputer in the World is powered by
AMD Opteron 6274's. (It's a Cray XK7 named Titan). CMT does make a lot of sense for Supercomputers.

http://en.wikipedia.org/wiki/TOP500

 

[MiddleOfTheRoad]

Then BOINC on shoestring clusters is the ultimate scientific application? Is this your use case for CMT processors?

Why don't you describe your personal experience on the topic? Because that was mine.

 

[mrmt]

Why don't you describe your personal experience on the topic? Because that was mine.

Our company has to deal with a lot of research projects:

- The upstream guys work with lots geological and seismic data for mapping prospective oil basins, a lot of CAD simulations of deep water projects, and CAD design for engineering components.

- Then there are the downstream guys working on building and improving the design of refineries (CAD and chemical process simulations), then the lubricants guys (basically chemical process simulations) and the supply chain guys.

Basically there are a few thousands of people working with R&D in our company, and we need this R&D to succeed in order to make money. To put things in perspective, we spend an amount close to what TSMC spends in R&D every year. All these research projects use Xeon or Xeon + Tesla for doing the heavy lift, and there's a lot of guys working on non-R&D CAD activities that uses Xeon + Quadro or at least Quadro, and these servers/clusters/workstations have to pass a ROI test before going for a certain configuration.

This amount of money you are talking about (around $5.000) is pocket change for even the smallest companies out there. The fact that $5.000 is all your friend got means that he is being paid to learn, not generate big results, so by no means you should consider it a baseline scenario for R&D deployments.

 

[MiddleOfTheRoad]

Our company has to deal with a lot of research projects:

- The upstream guys work with lots geological and seismic data for mapping prospective oil basins, a lot of CAD simulations of deep water projects, and CAD design for engineering components.

- Then there are the downstream guys working on building and improving the design of refineries (CAD and chemical process simulations), then the lubricants guys (basically chemical process simulations) and the supply chain guys.

Basically there are a few thousands of people working with R&D in our company, and we need this R&D to succeed in order to make money. To put things in perspective, we spend an amount close to what TSMC spends in R&D every year. All these research projects use Xeon or Xeon + Tesla for doing the heavy lift, and there's a lot of guys working on non-R&D CAD activities that uses Xeon + Quadro or at least Quadro, and these servers/clusters/workstations have to pass a ROI test before going for a certain configuration.

This amount of money you are talking about (around $5.000) is pocket change for even the smallest companies out there. The fact that $5.000 is all your friend got means that he is being paid to learn, not generate big results, so by no means you should consider it a baseline scenario for R&D deployments.

So the bottom line -- is you just wrote a bunch of off topic stuff because you couldn't come up with a single way SMT is superior to CMT.

CAD is a GPU intensive workload where the CPU makes a marginal contribution at best.

You could probably perform those same tasks with a lowly i5 and a Quadro and most of your engineers probably couldn't tell the difference. But, I do want to thank you for telling me about your lubricants guys. Downstream? Lubricants? LOL

You do realize that a Ph.D teaching at a major university makes more money than you do. Call me crazy -- but I'd rather have the big paycheck than the "big results."

 

[mrmt]

So the bottom line -- is you just wrote a bunch of off topic stuff because you couldn't come up with a single way SMT is superior to CMT.

I provided a scenario where a multi-billion company has 0 AMD CPUs on R&D projects because they can't pass a ROI analysis, you either has a clue of this kind of environment and get that it's not only CAD or you don't, and then still thinks that science workloads is BOINC on $5.000 clusters. I won't proceed any further here, ask your PhD friend what an oil company can do, maybe he has a better clue of what oil companies research.

You do realize that a Ph.D teaching at a major university makes more money than you do.

I really doubt that many universities can compete against oil companies in terms of salary, especially if you go against the top 20 companies of the world, and those that do will not limit their relevant research projects to $5.000 grants. We don't have many people leaving our staff to work on universities, but there are plenty of PhD guys that quit their academic careers to work with us, and most of those guys didn't work on pocket change grants like that you are using as example. Those guys do not come to us because they love the industry, they come to us mainly because of salary.

 

[Enigmoid]

So the bottom line -- is you just wrote a bunch of off topic stuff because you couldn't come up with a single way SMT is superior to CMT.

CAD is a GPU intensive workload where the CPU makes a marginal contribution at best.

Advantages for SMT

-Greater singlethread performance
-Easier and less complex to design (easier to avoid bottlenecks)
-Very efficient
Threads share resources so a thread in a haswell core CAN access all for ALUs for instance if the other thread is stalled on a memory fetch.
-Takes very little extra die space

Disadvantages for SMT

-Going for a wide core is inherently difficult and inefficient
-Sometimes performance is lower when code is not hitting a lot of stalls or cache misses
-Can put a lot of pressure on specific resources


Advantages for CMT
-two weak cores are inherently more efficient for a lot of tasks than 1 strong core
-area, die and power savings (supposedly as AMD never delivered on this)

Disadvantages for CMT
-weak ST performance
-Shared resources must be 'bulked' up for full module operation meaning that they are larger and therefore less efficient for singlethread operation. This is similar to SMT in that two weak ALUs or FPUs generally can get more done than 1 strong ALU/FPU. Ie the front end for FPU in BD has to be completely power up for singlethread operation and large enough for full module operation.
-Shared resources do not scale like non-shared resources. As the module is larger than a simple core would be the performance/mm^2 of these resources are generally not as good.
-A single BD thread only ever has access to 2 ALUs; the non-shared units are unique

Add more if I forgot anything.

 

[MiddleOfTheRoad]

I won't proceed any further here, ask your PhD friend what an oil company can do, maybe he has a better clue of what oil companies research.

I really don't need to ask my PhD friend what an oil company can do -- I'm pretty aware of "what an oil company can do" since I was born and raised in FLORIDA.

This is totally something I'd brag about.

Not only does my friend make more money -- he does it in a far more respectable fashion.

The State of Florida thanks you for your service.

We only lost about $23 Billion in tourism and another $37 Billion cleaning up the mess. Not exactly an industry I'd brag about belonging -- but more power to you.


Off topic/thread derail/trolling
Markfw900

 

[mrmt]

I really don't need to ask my PhD friend what an oil company can do -- I'm pretty aware of "what an oil company can do" since I was born and raised in FLORIDA.

No, you don't. Being born in Florida and watching an oil spill doesn't make you an expert in the oil industry, just as meeting a PhD and using BOINC doesn't make you an expert in scientific applications. It's taking a unidimensional view of a complex subject and thinking you have some deep knowledge of something, Dunning-Kruger at its prime.

 

[MiddleOfTheRoad]

No, you don't. Being born in Florida and watching an oil spill doesn't make you an expert in the oil industry, just as meeting a PhD and using BOINC doesn't make you an expert in scientific applications. It's taking a unidimensional view of a complex subject and thinking you have some deep knowledge of something, Dunning-Kruger at its prime.

Hilarious development.

You are well known on this forum for being one of the harshest critics of AMD.

And then it turns out you work in the oil industry -- which is the king of the universe for EPIC FAILS.

An industry infamous for its colossal, monstrous, amazingly huge screw-ups that massively dwarf just about any other failure (Deepwater Horizon, Exxon Valdez) which end up making AMD's Bulldozer look like an award-winning blockbuster home run success by comparison.

So I just want to thank you for your contribution. You didn't just make my day -- you made my year.