UPGRADE YOUR BROWSER

We have detected your current browser version is not the latest one. Xilinx.com uses the latest web technologies to bring you the best online experience possible. Please upgrade to a Xilinx.com supported browser:Chrome, Firefox, Internet Explorer 11, Safari. Thank you!

cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Observer moshiko
Observer
813 Views

Peak INT8 DSP numbers

Hi,

 

VU9P device specifications show 6840 DSPs and "Peak INT8 DSP (TOP/s)" = 21.3.

 

What does "INT8 DSP" operation refer to ?  Do you mean multiply ? multiply-accumulate or another operation?

 

I am familiar with Xilinx white paper on performing 2 INT8 MACC ops per DSP.

Assuming this use-case, then 6840 DSPs give 13680 ops/cycle. 400MHz clock frequency gives 5.4 TOP/s. Even with 800MHz clock, the throughput is still far from the spec.

 

Please explain how this 21.3 Top/sec is calculated.

 

Thank you

-Moshe

0 Kudos
5 Replies
Observer moshiko
Observer
748 Views

Re: Peak INT8 DSP numbers

Can anyone from Xilinx respond to this question?

0 Kudos
Observer jeremyward
Observer
640 Views

Re: Peak INT8 DSP numbers

No response?

0 Kudos
Xilinx Employee
Xilinx Employee
433 Views

Re: Peak INT8 DSP numbers

Hello @moshiko,

 

We are looking into this and we will get back to you very soon.

 

Thanks,
Chris

0 Kudos
Xilinx Employee
Xilinx Employee
419 Views

Re: Peak INT8 DSP numbers

@moshiko, apologies for the delay responding, here's a breakdown of the calculation for you.

 

For VU9P:

#DSPs 6840, for a -3 device FMAX = 891 MHz

 

Using the pre-adder in the DSP48 slice you get a standard GMACs performance of 12188 GMACs.

Applying the method in Xilinx WP486, i.e. multiplication factor of 1.75. You get 12188 x 1.75 = 21.3 INT8 TOPs.

 

Without the pre-adder included you would get approximately 10.7 INT8 TOPs.

 

Best regards

 

 

0 Kudos
Observer moshiko
Observer
413 Views

Re: Peak INT8 DSP numbers

Hi,

 

Can we continue this discussion offline by mail? 

I have a few questions to further clarify this calculation.

 

Thank you,

-Moshe

0 Kudos