Defining operand precisions

[sjoks93]

How do we define the precision (data width) and sparse compressed representation of each operand of a problem, and at different memory levels (e.g compressed output features at higher levels, operands uncompressed in higher levels)?

Is it possible to define them in a flexible architecture, where memory is completely shared, instead of simply defining the datawidth of the storage?

[angshuman-parashar]

That's not possible today. It's a very useful feature though, so if you are interested in contributing I can guide you. Should be a ~3-day project.

[siddharth-joshi]

Hi Angshu, actually, I'd like to contribute to this. It's something on the critical path for our own projects, so I'd be interested. Might take me longer than three days though, just to get up to speed!

[angshuman-parashar]

@siddharth-joshi are you thinking of per-operand precision (data-width) or compression formats?

[siddharth-joshi]

Both, first per-operand precision and then per-operand compression. I think we have the latter somewhat hacked in somewhere, but I can let @pooria-taheri comment on that more

On Mon, Oct 16, 2023 at 13:59 Angshuman Parashar @.***> wrote:

@siddharth-joshi https://github.com/siddharth-joshi are you thinking of per-operand precision (data-width) or compression formats?

— Reply to this email directly, view it on GitHub https://github.com/NVlabs/timeloop/issues/213#issuecomment-1764997554, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAJT7QZ63R5EAEOQHRBR26TX7VYZHAVCNFSM6AAAAAA3C3EFW6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONRUHE4TONJVGQ . You are receiving this because you were mentioned.Message ID: @.***>

[vkb27]

Hi, I have to analyse performance of implementation of different numerical representations, like POSIT. Is this something possible in Timeloop (I'm just starting out on the tool, gone through the tutorials) by working on the source code? If yes, are there any online resources around it?

[angshuman-parashar]

@vkb27 If you specify the number of bits required by each numerical representation you want to evaluate, Timeloop will model the energy cost of moving those bits across the hardware. This works very well for fixed-width data. For variable bit-widths, you may have to provide an average and convince yourself that the resultant modeling is a reasonably accurate proxy of the actual variable-width behavior. You will also have to provide (via Accelergy) the cost of accessing each arithmetic unit and each SRAM structure on your hardware. If you would like to discuss this further I suggest creating a new issue.

[angshuman-parashar]

Both, first per-operand precision and then per-operand compression. I think we have the latter somewhat hacked in somewhere, but I can let @pooria-taheri comment on that more

@poant Could you confirm if per-operand compression format is already supported?

[siddharth-joshi]

Hi Folks, just following up on this?