can you have more than 1 Xeon Phi on a PC?

It's available installed in a PCIe slot, so several cards if you want.

Release date was last year. A 60 core/8GB 5110P-based card will set you back thousands of dollars, in the same range as a higher end Tesla accelerator. It's not really a consumer product.
 
Any1 use Office 2007 or 2010 or 2013 that does a lot of Excel iteration calculation and notice a major gain?

I know Excel should benefit from multiple core, but Microsoft wasn't very clear whether it support multi-core on co-processor such as the Xeon Phi.

It did say it support multi-core on your main CPU.

In short, for that kind of $, I want to be very sure
 
Excel does not scale with more and more cores. I know people misuse spreadsheets for many purposes, but if you're doing heavy parallelize-able calculations, a dedicated application is probably better.
 
pxc said:
Excel does not scale with more and more cores. I know people misuse spreadsheets for many purposes, but if you're doing heavy parallelize-able calculations, a dedicated application is probably better.
Click to expand...

^^ This.

The Phi is also meant to be used for enterprise and scientific purposes and not for home use. It's going to be expensive and compete with Tesla and Firepro cards.
 
Well, the re-iteration calculation is for scientific engineering purpose. Having said that, what if I add a Tesla, would it help on Excel 2007 or 2010 for those iteration calculation?
 
No. You would need to write a custom program, using CUDA, OpenCL or DirectCompute (/C++ AMP) to exploit GPGPU. Excel doesn't know what to do with anything other than the main CPU*.

* you could write some kind of add on, but that's just silly and inefficient.
 
Happy Hopping said:
Well, the re-iteration calculation is for scientific engineering purpose. Having said that, what if I add a Tesla, would it help on Excel 2007 or 2010 for those iteration calculation?
Click to expand...

Why are you using Excel if you're contemplating spending >$1000 on a co-processor that won't benefit from parallelisation? If you're constrained by time and need more performance, and you're dishing out that much money, then you need to rethink what software you're using.
 
pelo said:
Why are you using Excel if you're contemplating spending >$1000 on a co-processor that won't benefit from parallelisation? If you're constrained by time and need more performance, and you're dishing out that much money, then you need to rethink what software you're using.
Click to expand...

it's not me. It's my client, he only knows how to use excel. Mirosoft did say on their office web page that their software supports multi-thread and can benefit by multi-core processors.

http://msdn.microsoft.com/en-us/library/office/bb687899.aspx

Microsoft Office Excel 2007 was the first version of Excel to use multithreaded recalculation (MTR) of worksheets. You can configure Excel to use up to 1024 concurrent threads when recalculating, regardless of the number of processors or processor cores on the computer.

If the computer has multiple processors or processor cores, the operating system takes responsibility for allocating the threads to the processors in the most efficient way.
Click to expand...

so the above appears to contradict what you people are saying?
 
Happy Hopping said:
so the above appears to contradict what you people are saying?
Click to expand...
The level of parallelism that can be extracted is limited by dependencies between the cells on the sheets*. Look at task manager when you recalculate a large spreadsheet. You could make a contrived example that allows more threading, but it's not going to be representative of virtually any real life data.

Xeon Phi and GPGPU resources are not be available to current versions of Excel. With C++ AMP available, it's possible future versions may support such hardware.

* for example, if there are 10 totally independent sets of data that need to be manipulated before other calculation(s) involving those cell(s) (because those rely on the results of another calculation first), etc and repeated if the calculation implementation is iterative
step 1: launch 10 separate threads on 10 ranges
step 2: after that is finished, launch 1..n threads to do next dependent steps
step 3: etc

Even if you're running millions or billions of calculations, if there are bottlenecks on dependencies, nothing magical will launch 1K threads just because there's a lot of data. If the calculation is one big ball of yarn, you may see little threading at all, and that's may be a sign that the strategy used is a poor one. It's a real problem on massively parallel programming, and optimizing algorithms can be tricky, especially when the answer needs to be correct (which is why sometimes simpler, less efficient algorithms are used since it's earlier to validate for correctness).

tl;dr: Excel isn't magical and does not use a Xeon Phi or GPGPU. To get the maximum performance out of Excel, you'll need to minimize dependency bottlenecks, and maybe even rethink your algorithms.

*edited*
 
Last edited:
You must log in or register to reply here.
Back
Top