Coprocessor, why would you use this?

[SebastiKing]

I was looking around some websites and found this on a german website, idk what the use of it is. It says it is an coprocessor made by Intel?

 

https://www.it-market.com/en/hewlett-packard/hp-proliant-server/hp-proliant-accessories/hp-e2m34a2?gclid=CjwKCAjw8MD7BRArEiwAGZsrBdNMBCx_DGqMeP4uwOFhjZrH3rNPe7jVapYWqSV6rwOCyumoQrXcdhoCvEIQAvD_BwE

[emosun]

for special workloads that require it , basically if you dont know what it does then you dont need it

[Oshino Shinobu]

It's more or less a x86 based CPU in the form of a PCIe expansion card. 

 

It's a pretty specialised bit of hardware, not many applications can really utilise it. 

 

Xeon Phi was actually discontinued this year, so presumably no one was using it anymore.

 

https://en.wikipedia.org/wiki/Xeon_Phi

[RollinLower]

those cards are used in specialized machines that benefit from high parallel workloads that can benefit from all te cores.

think stuff like weather calculations or something.

 

not really a consumer could use for anything. 

[igormp]

The Xeon Phi is a manycore processor that was meant to run MIMD (multiple instruction, multiple data) scenarios. Whereas a GPU works with SIMD (single instruction, multiple data) (like multiplying an entire vector of different numbers with another), and CPUs are SISD (you should get the gist of it by now), running a single operation (add, multiply, move, w/e) over a single piece of data.

 

Compared to your regular i7 CPU, the xeon phi is more akin to a crappy atom core, but features tons of those, so applications that have a high number of stuff to do in a simpler way can benefit from that. Some of those can actually run windows or linux since they're x86 CPUs, but performance is awful and there's not much benefit by doing so, you're better off writing custom programs that can take advantage of it, or abstracting it through something like OpenCL.

[porina]

The chips in those are weak general x86 cores partnered with strong FP64 FPU. They kinda sit somewhere between a CPU and a GPU in concept. Cores are more CPU-like, but scaled more GPU-like.

 

FP64 is an area that has been disappearing in consumer hardware. In consumer GPUs they were nerfed heavily after the nvidia 500 series and AMD 200 series, so modern cards perform a lot less efficiently than those older ones in this area. Some of the pro GPUs might still have decent FP64 rate, but you're really paying for it. The Radeon VII is a recent oddity, since it was essentially a rebadged pro level card it still had decent FP64 at consumer price, and there are people still hunting them down for such use. In CPUs, Intel leads the way with AVX-512 offering up to double the throughput of normal cores. AMD don't have anything comparable so are only fractionally ahead of a regular Intel core, but can kinda offset that by having more cores.

[porina]

8 minutes ago, igormp said:

the xeon phi is more akin to a crappy atom core

It's not just akin to a crappy atom core, it is mostly a crappy atom core. It's just there to allow it to function like a CPU. The heavy work is done by the phat FPU that eats floating point work. It is kinda like a drag racing car. The atom core is just enough to hold the massive engine and allow the whole thing to move.

 

If you're mostly using the atom parts, you're doing it wrong.

[Sophia_Borjia]

If you have use for dozen old rigs running single task in cluster, this could do that with out taking up a shelving rack.  Similar idea to people connecting dozens of rasberry pi, except phi is x86 not arm and scales much better for pro vs hobby use.

 

You could put 8 of those in single case and have 8*72=576 slow cores, this was also in pre 64 core epyc days when 28 cores was price of a new economy car.

 

Not really any consumer use, maybe if you wanted to host many game servers for older titles or something very niche.

[NumLock21]

1 hour ago, porina said:

It's not just akin to a crappy atom core, it is mostly a crappy atom core.

Them Intel Atom core is a CPU.

My HP netbook ran on an Intel Atom, and the Xeon Phi is a bunch of Intel Atom cores. Intel did something a bit different with HT, on them. Normally it's 1 core, 2 threads. For these it's 1 core, 4 threads. So a 64 core Xeon Phi will have 256 threads.

[igormp]

3 hours ago, porina said:

It's not just akin to a crappy atom core, it is mostly a crappy atom core. It's just there to allow it to function like a CPU. The heavy work is done by the phat FPU that eats floating point work. It is kinda like a drag racing car. The atom core is just enough to hold the massive engine and allow the whole thing to move.

 

If you're mostly using the atom parts, you're doing it wrong.

If you're talking about the newest atom-based phis, IIRC those have just the regular FPU unit of a regular atom core. The fancy thing is the 2 AVX512 units present in each core, meaning that you can use those as a shitty GPU.

[porina]

51 minutes ago, igormp said:

The fancy thing is the 2 AVX512 units present in each core, meaning that you can use those as a shitty GPU.

That is the selling point. High throughput potential of FP64 operations. A feature lacking in consumer grade GPUs for a long time. Not that weak sauce FP32 rubbish.