Electronics > Manufacturing & Assembly

Compute module alternatives or purchasable ARM chips?

<< < (5/6) > >>

NorthGuy:

--- Quote from: loki42 on June 18, 2023, 12:49:24 pm ---I also have had some trouble finding benchmarks that compare single core floating point performance (only performance thing I care about) between different options especially comparing Celeron / pentium gold to big ARM core A7x families.

--- End quote ---

The benchmark will not tell you how the CPU performs in your particular application. Create a small program which represents the operations you want to do and run it on several different CPUs. This will be your own benchmark. This will give you much more accurate estimate. But even then it also depends on many other things beyond CPU, such as memory.

The 64-bit mode (Aarch64) has SIMD registers for floating points. If you can use SIMD sensibly, it'll give you much better performance than scalar operations. A code which uses SIMD is likely to perform better on A53 than the scalar version of the same code on a similarly versed A72.

Also, you may consider an FPGA. It can do many operations in parallel and therefore will outperform CPU in doing floating operations. So, if you can port your algorithm to FPGA, you may be able to get your performance at lower price.

loki42:
Cheap and FPGA isn't something I'd thought of together but there's obviously a point where it makes sense.  The algorithm is already heavily SIMD on the A53, but obviously getting examples of 10 different chips to run it will take a while so an approximate benchmark would help.  Should I expect any of the ARM chips to be faster in single core floating point performance vs low end x86 if neon and SSE are being used 64 bit OS?

Smokey:

--- Quote from: Doctorandus_P on June 09, 2023, 10:37:18 pm ---You can try Olimex. Olimex has some special long term deal with Alwinner, and Olimex is big enough to buy enough chips so that Allwinner is willing to start up the factory to make a new batch of chips especially for Olimex. I think they have to buy 10.000 chips or so as a minimum order quantity for this deal. (Although at the moment both the A20 and A64 loose IC's seem to be "out of stock".

On the plus side,
Olimex has several different Linux boards and SOM modules, and in different configurations (inclusive industrial temperature range), and they have also made the complete KiCad projects of a lot of their boards available via gitlab, so it's quite easy to make a custom variant of them.

--- End quote ---

Link: https://www.olimex.com/Products/SOM/

loki42:
AMD xilinx meeting tuned out to be all them pushing FPGA chips which seem like they'll be pretty annoying to develop for. Couldn't get them to tell me about embedded ryzen. Intel has some decent options and has part volume available at sensible prices. 

NorthGuy:

--- Quote from: loki42 on June 18, 2023, 10:26:47 pm ---Cheap and FPGA isn't something I'd thought of together but there's obviously a point where it makes sense.  The algorithm is already heavily SIMD on the A53, but obviously getting examples of 10 different chips to run it will take a while so an approximate benchmark would help.  Should I expect any of the ARM chips to be faster in single core floating point performance vs low end x86 if neon and SSE are being used 64 bit OS?

--- End quote ---

I don't know. It's very difficult to predict performance of OOB CPUs. You don't need to buy 10 different chips. You can buy a few of what you feel are good candidates for you.

With FPGA, you don't need a physical part to estimate performance. The tools will tell you how fast your design works.

Navigation

[0] Message Index

[#] Next page

[*] Previous page

There was an error while thanking
Thanking...
Go to full version
Powered by SMFPacks Advanced Attachments Uploader Mod