AMD CPU speculation... and expert conjecture

[Blandge]

^^ xeon phis are more performance efficient than gpus? i thought kepler-based teslas were more efficient and phis' main attraction was the ease of progammability.
edit: disclaimer - lack in-depth knowledge about hpc and supercomputers.

This one implementation of Xeon + Xeon Phi achieved better efficiency than the Opteron + K20x Titan (And all other supercomputers at the time). That doesn't necessarily mean that Xeon Phi is always more efficient than Nvidia's offering. There's too many factors here to make a blanket statement about one component to the supercomputer.

 

[Blandge]

By efficiency I did mean performance per watt. The #1 top supercomputer using AMD X86-Opterons plus Nvidia cards offers 20 petaflop using 9 MW of power. Assuming linear scaling 200 petaflop would use 90 MW of power (the power consumption will be higher because is nonlinear). The European project, using ARM CPU + ARM GPU wait to obtain 200 petaflop using only 10 MW.

http://hexus.net/tech/news/cpu/48193-new-eu-based-supercomputer-arm-based/

Imagine what AMD could offer us with its future ARM-Opteron chips.

You should never assume linear scaling.

 

[8350rocks]

I didn't mean you were bragging, that the article was...perhaps that was worded poorly.

Either way...the lack of GPGPUs is where the power consumption difference falls squarely.

Would you not consider Xeon Phi GPGPU? If not it's pretty damn close. Certainly close enough to have similar power and performance characteristics.

Either way, you are wrong. Yes GPUs use more power than CPUs, but they increase efficiency on HPC workloads. This means that overall they use less power per FLOP on highly parallel workloads. That's why the top (x86) Green HPC devices all have some GPGPUs (Or Xeon Phi), because it is more efficient. Otherwise the whole toplist would be nothing but Opterons and Xeons.

That depends on the % ratio of CPUs to GPGPUs...it also depends on the workload, some workloads will use more of the CPUs and some will use more of the GPGPUs. The power draw from the 2 sources can very depending on the benchmark (often dramatically).

GPGPUs may be able to perform more FLOPS than CPUs because of their parallel nature, but they still draw more power. When you're talking about 17.59 petaflops...things are on a different level.

 

[Blandge]

That depends on the % ratio of CPUs to GPGPUs...it also depends on the workload, some workloads will use more of the CPUs and some will use more of the GPGPUs. The power draw from the 2 sources can very depending on the benchmark (often dramatically).

GPGPUs may be able to perform more FLOPS than CPUs because of their parallel nature, but they still draw more power. When you're talking about 17.59 petaflops...things are on a different level.

They do not draw more peak power per FLOP when processing the massively parallel workloads that they are designed for, otherwise nobody would use them. What is the point of putting 20000 K20x in Titan if they are not as efficient as CPUs? Why not just add 40000 Opterons and increase efficiency.

The benchmark used in the LINPACK Benchmark is to solve a dense system of linear equations. For the TOP500, we used that version of the benchmark that allows the user to scale the size of the problem and to optimize the software in order to achieve the best performance for a given machine. This performance does not reflect theoverall performance of a given system, as no single number ever can. It does, however, reflect the performance of a dedicated system for solving a dense system of linear equations. Since the problem is very regular, the performance achieved is quite high, and the performance numbers give a good correction of peak performance.Source

GPUs are good at solving linear equations.

 

[griptwister]

Kabini comes without Windows XP support
http://www.fudzilla.com/home/item/31314-kabini-comes-without-windows-xp-support

AMD's Answer to GeForce GTX 700 Series: Volcanic Islands
http://www.techpowerup.com/183704/amds-answer-to-geforce-gtx-700-series-volcanic-islands.html

Mother of mothers, mothers' Fathers... that's insane.

 

[esrever]

Volcanic Islands looks freaking amazing. I wonder if they will have full x86 or ARM cores in it or is it going to have limited x86 decode logic and work mostly with micro ops along with the gpu.

Nvidia won't any something like this until volta?

 

[JAYDEEJOHN]

It my understanding early Maxwells wont have this, later they will.

 

[Cazalan]

So can write your thesis statements on a cellphone? Really do tell us all about that...

What's to prevent you from writing a document on an Android powered cellphone?

OliveOffice, Google Docs, there must be hundreds of word processors available.

Bluetooth keyboard. Cheap as heck. http://www.newegg.com/Product/Product.aspx?Item=N82E16875982435

Or use one of the many voice to text apps.

Cellphones have HDMI outputs for monitors or get a phone with a docking bay like the Asus Padfone. I wouldn't suggest editing on a 4" screen but you can plug into your TV, monitor or soon Google Glass (VR headset).

http://www.zdnet.com/asus-padfone-2-review-7000014571/

Need some circuit simulation? EveryCircuit
Need a math app or calculator? Tons of them.

A modern cellphone craps all over the TI scientific calculator I had to use for Physics. And it isn't a whole lot more than the high end ones cost ($199) at the time.

 

[JAYDEEJOHN]

A motorcycle can be faster than most cars
A phone or PC shouldnt be confused with this
http://en.wikipedia.org/wiki/Sumlock_ANITA_calculator

 

[Cazalan]

Really need to look up the definition of a monopoly. There's a whole generation of kids that couldn't give a rats arse about the home PC. They're too busy on their tablets, smart phones and hand held gaming systems.

Everything I needed a PC for in college can be done on a cell phone now. There are specific market segments that Intel services better but that exclusivity becomes less every day.

By that you mean you can google search something for basic research but then lets also get into the myth that is "I can do everything on a cell phone".

1) Not all phones are equal, this is the biggest hurdle some phones don't even have proper high speed wireless internet or screens that are smaller and completely impractical to work off, to get better and bigger you need to spend money and by the time you are onto the Galaxy S3's ilk you are spending more than a laptop or PC so again phones are like justin bieber, a fab for now until they get boring ie: kik, whatsapp and chat apps really just save text costs but ramp up data costs until the kids figure that out they spend more money.

2) show me a phone that I can do a legal document on within time is of the essence and expediently and I will sell every system I have and only use a smartphone.

3) when a phone is no longer a phone it dilutes the purpose, I don't believe in smartphones

Believe in it or not, but Moore's law pretty much says the smartphone of tomorrow is the supercomputer of yesterday.

"For example, a tweaked Motorola Droid can hit 52 Mflop/s which is more than 15 times faster than the CPUs used in the 1979 Cray-1."

Cray-1 weighted 5.5tons and used 115kWatts of power.

That is in your pocket today, and soon in your watch. Just imagine what you'll get at the 5nm node.

 

[JAYDEEJOHN]

Let see, so everyone on a couch playing a console is actually playing a PC?
We can blur the line, but until it has become irrelevant its relevant

 

[Cazalan]

So instead of smartphones, they will be called smartGlasses

Or just Glass made by Google. It's a dual core ARM processor running Android.

Which rather looks like a Stick PC attached to some glasses.

http://www.makeuseof.com/tag/what-is-an-android-stick-computer-and-how-can-you-use-it/

 

[Cazalan]

Let see, so everyone on a couch playing a console is actually playing a PC?
We can blur the line, but until it has become irrelevant its relevant

PC is just Personal Computer. The form factors and inputs/outputs are changing but that happens all the time. A PS3 console modded to run Linux very much is a PC.

 

[Cazalan]

A motorcycle can be faster than most cars
A phone or PC shouldnt be confused with this
http://en.wikipedia.org/wiki/Sumlock_ANITA_calculator

And there are cars with motorcycle engines, and motorcycles with jet engines. All will get you from A to B and are personal transportation devices.

http://jalopnik.com/5852692/the-ten-most-awesome-motorcycle+engined-cars/

 

[Cazalan]

Except no. Smart phones are not killing / replacing desktops. Smart "glass's" will not replace smart phones. Each of these devices integrates and compliments the other. Smartphones let you communicate on the go while maintaining some connection to the internet with serviceable applications. The desktop is when you actually need to get something done, they compliment each other (smartphone docking stations SUCK HARD, I've used them first hand). In that same vein wearable computing devices aka "Smart Glass's" will connect to your smart phone and provide a HUD to display information from that same smart phone device. Utilizing PAN technology your glass's would be in constant communication with your communicator (proper name for a smart phone) and could display incoming information or even augmented reality. Remember glass's must be as light and small as possible, this drastically limits the hardware that can actually exist in them, yet your communicator has more physical space and isn't as limited.

And yes x86 doesn't scale down very well. Via's done their best but there is a hard limit to how low you can go due to the complex front end decoders, instruction schedulers and required caching system.

The biggest limitation of mobile tech is the battery that has relatively gone unchanged. It can even contribute more heat than the processor does.

If this 3D battery tech takes off there will be a major paradigm shift.

http://www.extremetech.com/computing/153614-new-lithium-ion-battery-design-thats-2000-times-more-powerful-recharges-1000-times-faster

Docking may suck for phones but is it any worse than the docking stations for laptops used to be? At least now there is the wireless capability and Bluetooth auto pairing.

 

[jdwii]

Its always nonsense when people say things like the desktop market is dying, i've been hearing it for years and what's funny is my college and work still all use desktops, workstations use them smartphones are for fun only and not even enjoyable fun more like 5min on a game. What's funny is PC gaming is coming back and sales prove this to be true and i doubt this generation of consoles will be as popular this time around. Also i still fail to realize the point of a tablet who would want one? My mom had one and now she is back on her desktop. Tablet is just setting their. There's always going to be a market for the most powerful stuff, its never going to change i just find it funny that businesses think they know people more than people do.

 

[anxiousinfusion]

Its always nonsense when people say things like the desktop market is dying, i've been hearing it for years and what's funny is my college and work still all use desktops, workstations use them smartphones are for fun only and not even enjoyable fun more like 5min on a game. What's funny is PC gaming is coming back and sales prove this to be true and i doubt this generation of consoles will be as popular this time around. Also i still fail to realize the point of a tablet who would want one? My mom had one and now she is back on her desktop. Tablet is just setting their. There's always going to be a market for the most powerful stuff, its never going to change i just find it funny that businesses think they know people more than people do.

This compliments your argument:

People need to understand that new form factors go on to exist alongside of current designs. Did the laptop ever replace the desktop in all the years since they where introduced? No. Rather, they compliment one another. The industry's marketing divisions (especially Microsoft's) need to take a good hard look at that prospect.

 

[8350rocks]

That depends on the % ratio of CPUs to GPGPUs...it also depends on the workload, some workloads will use more of the CPUs and some will use more of the GPGPUs. The power draw from the 2 sources can very depending on the benchmark (often dramatically).

GPGPUs may be able to perform more FLOPS than CPUs because of their parallel nature, but they still draw more power. When you're talking about 17.59 petaflops...things are on a different level.

They do not draw more peak power per FLOP when processing the massively parallel workloads that they are designed for, otherwise nobody would use them. What is the point of putting 20000 K20x in Titan if they are not as efficient as CPUs? Why not just add 40000 Opterons and increase efficiency.

The benchmark used in the LINPACK Benchmark is to solve a dense system of linear equations. For the TOP500, we used that version of the benchmark that allows the user to scale the size of the problem and to optimize the software in order to achieve the best performance for a given machine. This performance does not reflect theoverall performance of a given system, as no single number ever can. It does, however, reflect the performance of a dedicated system for solving a dense system of linear equations. Since the problem is very regular, the performance achieved is quite high, and the performance numbers give a good correction of peak performance.Source

GPUs are good at solving linear equations.

http://www.advancedhpc.com/gpu_computing/tesla_gpu_s2050_s2070.html

4 GPGPUs used in that HPC consume 900W power...while 4 opterons consume an average of 600W.

Why do you think the most efficient HPC on the "green list" doesn't have them...?

The answer is DENSITY...you can do far more double precision FLOPS with GPGPUs bundled together in a tighter space. It would be more efficient to run the same number of Opterons in terms of power consumption in a raw form. However, depending on the performance you're looking for, GPGPUs become a clear advantage to get to massive numbers, at the cost of power consumption.

If you want to get into FLOPS/watt...that's a different animal, and GPGPUs will do more FLOPS/watt...but...your power consumption increases linearly as you increase performance.

As I said, it depends on the workload, if GPGPUs were so clearly more efficient, then why have a system with CPUs at all? Why not use some low level coded OS in a DOS type format coded specifically for GPGPUs?

Because GPGPUs are less efficient at some workloads than CPUs. Different benchmarks will stress different components differently.

EDIT: Xeon Phi is not a GPGPU, either...that would be like saying Kaveri is a GPGPU, it's not.

 

[palladin9479]

Except no. Smart phones are not killing / replacing desktops. Smart "glass's" will not replace smart phones. Each of these devices integrates and compliments the other. Smartphones let you communicate on the go while maintaining some connection to the internet with serviceable applications. The desktop is when you actually need to get something done, they compliment each other (smartphone docking stations SUCK HARD, I've used them first hand). In that same vein wearable computing devices aka "Smart Glass's" will connect to your smart phone and provide a HUD to display information from that same smart phone device. Utilizing PAN technology your glass's would be in constant communication with your communicator (proper name for a smart phone) and could display incoming information or even augmented reality. Remember glass's must be as light and small as possible, this drastically limits the hardware that can actually exist in them, yet your communicator has more physical space and isn't as limited.

And yes x86 doesn't scale down very well. Via's done their best but there is a hard limit to how low you can go due to the complex front end decoders, instruction schedulers and required caching system.

The biggest limitation of mobile tech is the battery that has relatively gone unchanged. It can even contribute more heat than the processor does.

If this 3D battery tech takes off there will be a major paradigm shift.

http://www.extremetech.com/computing/153614-new-lithium-ion-battery-design-thats-2000-times-more-powerful-recharges-1000-times-faster

Docking may suck for phones but is it any worse than the docking stations for laptops used to be? At least now there is the wireless capability and Bluetooth auto pairing.

No Just no.

Smartphones design limitations are what prevents them from being efficient at those tasks. Did you miss the part where I said I'm living in the Cell Phone capital of the world? I get to see and play with all sorts of cool prototypes, most of which will never be seen in the English speaking world. I've seen and used a proposed desktop conversion kit for a Galaxy S3, the station had a HDMI connector to a HDTV along with the required USB connectors for KB / Mouse. The results sucked hard. Neither the OS nor the HW are designed with that implementation in mind and thus it becomes incredibly clunky to use. Not to mention how utterly weak the hardware is, composing an email was excruciatingly painful.

Now they could introduce a true preemptive multitasking OS, increase the hardware power while providing larger storage, but then nobody would buy it was it's no longer mobile enough to be a smartphone.

That's the catch of the entire thing, every advancement in miniaturized computing is first done on the desktop level. Every update, advancement or capabilities increase is a result of a desktop innovation. This means that the desktop form-factor will always be more advanced then the handheld. Which in turn means there will always be things you can do on the desktop which won't work on a handheld. So seriously, stop thinking that handhelds are replacing desktops, that reality only exists in your imagination or in the propaganda that pass's for advertising these days.

 

[Cazalan]

Its always nonsense when people say things like the desktop market is dying, i've been hearing it for years and what's funny is my college and work still all use desktops, workstations use them smartphones are for fun only and not even enjoyable fun more like 5min on a game. What's funny is PC gaming is coming back and sales prove this to be true and i doubt this generation of consoles will be as popular this time around. Also i still fail to realize the point of a tablet who would want one? My mom had one and now she is back on her desktop. Tablet is just setting their. There's always going to be a market for the most powerful stuff, its never going to change i just find it funny that businesses think they know people more than people do.

When they say dieing, what they really mean is it's growth is dieing or is negative. Many of the companies I work with replaced their desktops with laptops some time ago. The desktop isn't going to disappear but the need for it certainly has diminished greatly.

In my office we really only use desktops for test machines because it's a lot easier to swap drives around and boot multiple system configurations. That part is still a pain to do with laptops.

When kids go to college now they don't want a desktop, they want a laptop or a tablet they can carry to class and take notes with. Even so, notebook sales are down 20%.

 

[Blandge]

That depends on the % ratio of CPUs to GPGPUs...it also depends on the workload, some workloads will use more of the CPUs and some will use more of the GPGPUs. The power draw from the 2 sources can very depending on the benchmark (often dramatically).

GPGPUs may be able to perform more FLOPS than CPUs because of their parallel nature, but they still draw more power. When you're talking about 17.59 petaflops...things are on a different level.

They do not draw more peak power per FLOP when processing the massively parallel workloads that they are designed for, otherwise nobody would use them. What is the point of putting 20000 K20x in Titan if they are not as efficient as CPUs? Why not just add 40000 Opterons and increase efficiency.

The benchmark used in the LINPACK Benchmark is to solve a dense system of linear equations. For the TOP500, we used that version of the benchmark that allows the user to scale the size of the problem and to optimize the software in order to achieve the best performance for a given machine. This performance does not reflect theoverall performance of a given system, as no single number ever can. It does, however, reflect the performance of a dedicated system for solving a dense system of linear equations. Since the problem is very regular, the performance achieved is quite high, and the performance numbers give a good correction of peak performance.Source

GPUs are good at solving linear equations.

http://www.advancedhpc.com/gpu_computing/tesla_gpu_s2050_s2070.html

4 GPGPUs used in that HPC consume 900W power...while 4 opterons consume an average of 600W.

Why do you think the most efficient HPC on the "green list" doesn't have them...?

The answer is DENSITY...you can do far more double precision FLOPS with GPGPUs bundled together in a tighter space. It would be more efficient to run the same number of Opterons in terms of power consumption in a raw form. However, depending on the performance you're looking for, GPGPUs become a clear advantage to get to massive numbers, at the cost of power consumption.

If you want to get into FLOPS/watt...that's a different animal, and GPGPUs will do more FLOPS/watt...but...your power consumption increases linearly as you increase performance.

As I said, it depends on the workload, if GPGPUs were so clearly more efficient, then why have a system with CPUs at all? Why not use some low level coded OS in a DOS type format coded specifically for GPGPUs?

Because GPGPUs are less efficient at some workloads than CPUs. Different benchmarks will stress different components differently.

EDIT: Xeon Phi is not a GPGPU, either...that would be like saying Kaveri is a GPGPU, it's not.

Of course I'm talking about FLOPS/Watt that's what Green500 list is. We are comparing GPUs to CPUs in the context of HPC peak Linpack performance. GPUs win at efficiency hands down.

Let's take your example of 4 Telsa K20x vs 4 Opteron 6274.

4 Telsa K20x
900w (As you said)
1.31 x 4 TFLOPS = 5240 GFLOPS
5240/900 = 5.82 GFLOPS/W or 5822MFLOPS/W
Source

4 Opteron 6274
600W (As you said)
0.292 x 4 TFLOPS = 1168 GFLOPS
1168/600 = 1.95 GFLOPS/W or 1947 MFLOPS/W
Source

The GPUs are roughly 3x more efficient by that math. Obviously if you are doing serial workloads with tons of branches the GPUs will be worthless, but we're talking about HPC workloads here.

Xeon Phi is MUCH more similar to a GPGPU than Kaveri, because Kaveri is a heterogeneous APU with a few powerful x86 cores for serial workloads and many GPU shaders, while Xeon Phi is a homogeneous array of relatively weak x86 cores that are meant to perform massively parallel workloads. Obviously if you think Xeon Phi is equivalent to Kaveri you do not understand what one of them is. I'm guessing Xeon Phi. Please go search Google and read up on Xeon Phi and return when you are more informed.

 

[Blandge]

That's the catch of the entire thing, every advancement in miniaturized computing is first done on the desktop level. Every update, advancement or capabilities increase is a result of a desktop innovation. This means that the desktop form-factor will always be more advanced then the handheld. Which in turn means there will always be things you can do on the desktop which won't work on a handheld. So seriously, stop thinking that handhelds are replacing desktops, that reality only exists in your imagination or in the propaganda that pass's for advertising these days.

Do you believe this to be true when it comes to system integration and wireless (3G/LTE) technology? I would tend to say that my cell phone is leaps and bounds ahead of my desktop in these categories.

 

[mayankleoboy1]

As I said, it depends on the workload, if GPGPUs were so clearly more efficient, then why have a system with CPUs at all? Why not use some low level coded OS in a DOS type format coded specifically for GPGPUs?

Because GPGPUs are less efficient at some workloads than CPUs. Different benchmarks will stress different components differently.

As far as Supercomputers are concerned, almost all of their workload is solving simoltanous linear equations, maybe of thousand or millions of variables. These equations are solved with matrixes, which are trivial to parallelize. Hence GPU's excel at this type of work. And this is the reason why supercomputers are trying to add as much of GPU to themselves. Or they have to do large FFT transforms. Which again are trivial to parallelize.
GPU's give much higher FLOPS , and much higher FLOPS/watt as well.

EDIT: Xeon Phi is not a GPGPU, either...that would be like saying Kaveri is a GPGPU, it's not.

Xeon phi does not have any serial processing. It also does not have any graphics pipeline or output. So its not a GPGPU. It is exactly what it says on the packaging : Its a math co-processor.

Kaveri has serial cores, and parallel cores. So it is a GPGPU.

 

[mayankleoboy1]

It my understanding early Maxwells wont have this, later they will.

links ?
Last i remember, CharlieD said that early Maxwell will be made on 28nm ,but later will be made on 20nm.

 

[mayankleoboy1]

AMD has special order of second batch from TSMC and this probably explains why Maxwell will only see the light of day closer to 2015.

I have poor memory. Source ?

I do expect volcanic Islands to be on average 2-3x faster than existing Southern Islands parts across the board with a universal "double em up" through all SKU's.

2-3x seems excessive. I would be happy with 1.6-1.8 perf in the top end parts.