A REAL 64 Core CPU - For SCIENCE!

[Emily Young]

Forget Threadripper and Core i9 - Xeon Phi is the go-to for x86 core count. But how does it actually perform?

 

 

More info on the Supermicro SuperServer 5038K-I: http://geni.us/g82JV
Buy a Xeon Phi processor on Amazon(?): http://geni.us/z2IB4Og

[3 Lions]

I bet there is one guy in the world who actually has bought one of these with his own money.

[Daniel644]

1 minute ago, Name Taken said:

Wonder how fast it would mine Monero.

probably slower then a GPU at the same power draw and a FRACTION the price, CPU mining is stupid slow.

[BlueChinchillaEatingDorito]

Hey Linus, can I borrow that to do my MATLAB assignments? Solving 5000x5000 matrices took a good 1.5 hours to run on the remote Xeon workstation I had available. 

[Prysin]

26 minutes ago, GabenJr said:

Forget Threadripper and Core i9 - Xeon Phi is the go-to for x86 core count. But how does it actually perform?

 

 

More info on the Supermicro SuperServer 5038K-I: http://geni.us/g82JV
Buy a Xeon Phi processor on Amazon(?): http://geni.us/z2IB4Og

are you guys going to try get ahold of EPYC?

[flibberdipper]

16 minutes ago, LinusTechTipsFanFromDarlo said:

I bet there is one guy in the world who actually has bought one of these with his own money.

Well, one guy sure. Bet large datacenters or whatever have bought quite a few of these guys.

[Nicnac]

I don't get it. This thing is hella expensive and can't do anything too well. And the stuff it was designed for runs much much better on a 1080ti...?

[Anjelllo]

32 minutes ago, LinusTechTipsFanFromDarlo said:

I bet there is one guy in the world who actually has bought one of these with his own money.

 Some were actually on sale a while ago for 200$.

[SiverJohn17]

So fun fact, I am actually the customer for products like these. And I straight up don't care about Xeon Phis. They were interesting a while ago but from my perspective there is no reason to use one of those chips over your standard GPU. In fact my lab uses off the shelf Titan XPs for most of our simulations. For most of our purposes these cards are the best options with only the new Tesla V100s being more powerful. Standard disclaimer our simulations only utilize single floating point precision so pure FLOPS are fine. That being said I (being the odd man in our lab) want to get a V100 for a personal project.

TL;DR even for HPC applications Xeon Phis are mainly irrelevant.

[borisdarker]

So I noticed it has 4 threads per core. I've never seen that technology before. Are there any downsides. Also is there any chance that this will come to consumer or enthusest level CPUs in the near future? 

[Arby Bc]

Thank Dog for this video!  I've been waiting forever to see reviews of science-driven computers!

 

"unless you like to science for funsies."-Linus (6:45) ....   

 

C'mon Linus, some of us do science "for realsies". I couldn't give to sh--ts about video gaming, which seems to be the focus of this channel (sad poop emoji).  Please do more reviews on computers that are simulation focused like this!

 

[brandishwar]

1 hour ago, Nicnac said:

I don't get it. This thing is hella expensive and can't do anything too well. And the stuff it was designed for runs much much better on a 1080ti...?

 

It isn't an apples to apples comparison. The Phi is a CISC x64 processor, whereas the Pascal GPU is considered a RISC processor (though it doesn't follow any declared RISC standard). Yes there are major differences, and they are quite relevant depending on the task at hand.

 

And the Phi being a manycore x64 processor was the main selling point: existing software doesn't have to be significantly modified to run on it. And Intel's tools could help determine performance bottlenecks in multi-threaded applications to ensure performance scales with the core count, which is extremely important when talking about having these in a cluster.

 

One of the major reasons to have that many cores and threads is to spread out all the tasks a node will do. The Titan supercomputer uses 16-core Opteron processors with Tesla GPUs. The CPUs are not used for processing, but for driving the individual nodes, including getting data to and from the GPUs and communicating across the networking backbone, such as with the high speed storage or sending and receiving data from siblings and masters.

 

People who talk computers often forget the very diverse roles CPUs take on. One of which is handling hardware interrupts.

 

Such as those coming from networking equipment. Such as the Infiniband and 100GbE cards used in supercomputers. One of the reasons Threadripper's PCI-E lane count is eye-catching is the ability to support up to four (4) PCI-E 3.0x16 cards at full x16 speed across its 64 lanes. The Phi supports 36 lanes by comparison, still enough to handle a GPU compute card and an Infiniband card. The Phi does support hexa-channel memory configurations, compared to Threadripper's quad-channel.

 

Though it is interesting they made the Phi a primary processor instead of a coprocessor on a daughter card.

[SiverJohn17]

1 hour ago, borisdarker said:

So I noticed it has 4 threads per core. I've never seen that technology before. Are there any downsides. Also is there any chance that this will come to consumer or enthusest level CPUs in the near future? 

To give the standard answer, depends on what your work load is. If we are talking about gaming or something that needs high performance of a single core then yes. There are huge drawbacks, because you basically have four threads competing for one core. However for certain applications if you are intelligent on how you do your threads you could get some respectable speed ups.

 

Edit: Forgot to answer the last part of the question, no I don't see this coming to consumer any time soon. It'd be unnecessary for most applications I can think of.

[SiverJohn17]

20 minutes ago, brandishwar said:

 

It isn't an apples to apples comparison. The Phi is a CISC x64 processor, whereas the Pascal GPU is considered a RISC processor (though it doesn't follow any declared RISC standard). Yes there are major differences, and they are quite relevant depending on the task at hand.

 

And the Phi being a manycore x64 processor was the main selling point: existing software doesn't have to be significantly modified to run on it. And Intel's tools could help determine performance bottlenecks in multi-threaded applications to ensure performance scales with the core count, which is extremely important when talking about having these in a cluster.

 

One of the major reasons to have that many cores and threads is to spread out all the tasks a node will do. The Titan supercomputer uses 16-core Opteron processors with Tesla GPUs. The CPUs are not used for processing, but for driving the individual nodes, including getting data to and from the GPUs and communicating across the networking backbone, such as with the high speed storage or sending and receiving data from siblings and masters.

 

People who talk computers often forget the very diverse roles CPUs take on. One of which is handling hardware interrupts.

 

Such as those coming from networking equipment. Such as the Infiniband and 100GbE cards used in supercomputers. One of the reasons Threadripper's PCI-E lane count is eye-catching is the ability to support up to four (4) PCI-E 3.0x16 cards at full x16 speed across its 64 lanes. The Phi supports 36 lanes by comparison, still enough to handle a GPU compute card and an Infiniband card. The Phi does support hexa-channel memory configurations, compared to Threadripper's quad-channel.

 

Though it is interesting they made the Phi a primary processor instead of a coprocessor on a daughter card.

I apologize if this sounds condescending as I am unaware of your knowledge base. However while the statement that the code doesn't have to be significantly modified to run on it is true. If you wanted to get the most of the performance out of it you'd definitely have to rewrite it in order for it to be more vectorizable. Which has its own challenges. Of course if you've already written your code for that purpose then sure slapping one of these in versus a GPU is a huge difference. 

 

That being said I agree with your comment about it is interesting they made it a primary processor instead of a coprocessor. Honestly, that is most of the reason I've lost interest. Also the comment from Linus about it being good for NN is a bit strange considering Intel announced the Nervana Chip to handle that end of the market.

[brandishwar]

56 minutes ago, SiverJohn17 said:

I apologize if this sounds condescending as I am unaware of your knowledge base. However while the statement that the code doesn't have to be significantly modified to run on it is true. If you wanted to get the most of the performance out of it you'd definitely have to rewrite it in order for it to be more vectorizable. Which has its own challenges. Of course if you've already written your code for that purpose then sure slapping one of these in versus a GPU is a huge difference. 

 

Not condescending at all. You merely clarified what I said, given I mentioned the Intel analysis tools but didn't mention really why you'd want to use them. Intel advertised that toolset for better multithreading performance regardless of whether you use the Phi coprocessor or not.

 

My knowledge base, to fill in that detail, comes from about 20 years experience with programming, including writing multi-threaded applications, though I have never written anything intended to target the Phi. I've been considering playing around with the various parallelization options available -- just a matter of finding the time and motivation.

 

Plus slapping in a Phi versus using a GPU means not having to rewrite code that's already targeting the x64 instruction set. If you're building something new, however, you're likely much better off targeting a GPU platform -- whether using NVIDIA's library directly or going through OpenCL -- than targeting any CPU platform, though there may be use cases wherein the CPU or a hybrid approach is better than pure GPU computing.

 

1 hour ago, SiverJohn17 said:

That being said I agree with your comment about it is interesting they made it a primary processor instead of a coprocessor. Honestly, that is most of the reason I've lost interest.

 

Not to mention the setup that is needed to power it. I don't think you could've just plugged the daughter card into any mainboard. It had to meet certain requirements with regard to its BIOS, making it inaccessible as an option to most people. Which is half the genius behind CUDA and OpenCL: you can use consumer-grade hardware out of the gate.

[antoipod]

4 hours ago, Nicnac said:

 This thing is hella expensive and can't do anything too well

Wrong, it makes a cool looking paperweight. But in all seriousness, I think your right, It gets outperformed by other xeons in single core and by gpus in multicore, I can't think of why someone wouldn't just buy a gpu instead.

[brandishwar]

8 minutes ago, antoipod said:

Wrong, it makes a cool looking paperweight. But in all seriousness, I think your right, It gets outperformed by other xeons in single core and by gpus in multicore, I can't think of why someone wouldn't just buy a gpu instead.

 

GPUs and CPUs fit different use cases, and not every use case can be accelerated with a GPU.

[captain_to_fire]

I wonder how this would fare with Folding@Home? 

[SiverJohn17]

Just now, hey_yo_ said:

I wonder how this would fare with Folding@Home? 

It would definitely be beneficial, how beneficial depends on exactly how that code is implemented, something I am ashamed to say I am ignorant of (considering it is close to my field of interest).

[antoipod]

5 minutes ago, hey_yo_ said:

I wonder how this would fare with Folding@Home? 

Probably very poorly, I doubt folding at home can utilize 256 threads. 

[captain_to_fire]

1 minute ago, SiverJohn17 said:

It would definitely be beneficial, how beneficial depends on exactly how that code is implemented, something I am ashamed to say I am ignorant of (considering it is close to my field of interest).

 

1 minute ago, antoipod said:

Probably very poorly, I doubt folding at home can ustilize 256 threads. 

Linus could’ve tested it with F@H. 

[CalhounTyphoon]

Love these more out there and non traditional stuff. Definitely my favorite kinds of videos! 

[SiverJohn17]

4 minutes ago, antoipod said:

Probably very poorly, I doubt folding at home can utilize 256 threads. I'd like to try it out but don't have a powerful enough power supply or windows server 2016.

Actually, and I didn't realize when I made my first comment, it most certainly can. According to Wikipedia at least (I had trouble finding this on the F@H website) Folding@Home utilizes GROMACS which is known for its strong scalability. As well as its optimization so much so that it gets relatively minor improvements from going from CPU to GPU, making it a perfect application if someone has a couple grand they want to just spend for the hell of it.

[Drak3]

2 hours ago, antoipod said:

Wrong, it makes a cool looking paperweight. But in all seriousness, I think your right, It gets outperformed by other xeons in single core and by gpus in multicore, I can't think of why someone wouldn't just buy a gpu instead.

 

7 hours ago, Nicnac said:

I don't get it. This thing is hella expensive and can't do anything too well. And the stuff it was designed for runs much much better on a 1080ti...?

PHI brings something(s) to the table that GPUs can't. More instruction sets.

[NumLock21]

Note the thumbnail with Linus having the Xeon Phi at the bottom of the cpu cooler. No, the thermal paste isn't super glued to the cpu. That's how it actually is, cpu is first attached to the heatsink with that black bracket piece. Then the entire thing goes onto the socket, with metal rods as guides.