Inquiry Regarding Throughput and Latency Calculation in FlooNoC #97
Comments
I am not sure if I understand your question correctly, but essentially we count the ratio of cycles where a handshake occurs on the link and the total number of cycles. With a smaller load, you will also have fewer handshakes and a smaller utilization in the end.
This maximum throughput is derived from a frequency of 1.26GHz and a link width of 512-bit. So 1.26GHz x 512-bits = 645Gbps. However, this is maximum throughput. The effective bandwidth depends a lot on the burst length of a transfer or the number of bursts/transfers you can issue in the system. The 1.26GHz is the frequency achieved in typical conditions when we integrated FlooNoC into a compute tile (see the paper for more information). In this case study, FlooNoC was not on the critical path, so potentially the frequency could be even higher. I hope this helps, but let me know if it is not clear |
|
Thanks, I got your answer.
I just have one more question is there any particular reason to close the
frequency at 1.26Ghz? Can we go to frequencies higher than 1.26 GHz?
…On Fri, Feb 7, 2025, 2:13 PM Tim Fischer ***@***.***> wrote:
I would like to confirm the strategy used by FlooNoC to calculate the
throughput of the NoC. Specifically, I am interested in understanding the
behavior of FlooNoC under different test scenarios and how these scenarios
impact throughput and latency metrics. The goal is to verify how the NoC
performs under various conditions and to understand the relationship
between the system's traffic patterns and its throughput and latency
calculations.
Could you please provide insights on how FlooNoC calculates throughput and
how it adjusts under different load conditions, burst behaviors, and any
other relevant factors and approach that may affect these performance
metrics?
I am not sure if I understand your question correctly, but essentially we
count the ratio of cycles where a handshake occurs on the link and the
total number of cycles. With a smaller load, you will also have fewer
handshakes and a smaller utilization in the end.
FlooNoC has mentioned in its research paper that they have achieved a
maximum throughput of 645Gbps per link, how is it actually calculated. Can
you please guide so that I can repeat these steps and calculate the
throughput?
This maximum throughput is derived from a frequency of 1.26GHz and a link
width of 512-bit. So 1.26GHz x 512-bits = 645Gbps. However, this is
*maximum* throughput. The effective bandwidth depends a lot on the burst
length of a transfer or the number of bursts/transfers you can issue in the
system.
The 1.26GHz is the frequency achieved in typical conditions when we
integrated FlooNoC into a compute tile (see the paper for more
information). In this case study, FlooNoC was not on the critical path, so
potentially the frequency could be even higher.
I hope this helps, but let me know if it is not clear
—
Reply to this email directly, view it on GitHub
<#97 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AXTS27BY4ID53J4DEJQFV232OR2MNAVCNFSM6AAAAABWL3TOBOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNBSGM3TAMZQG4>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
|
1.26GHz is just the maximum frequency we achieved for the compute tile. But there is not a particular reason why you couldn't go higher than this. It just depends on the design |
|
Got it. Thanks for the answer!
…On Fri, Feb 7, 2025, 2:37 PM Tim Fischer ***@***.***> wrote:
1.26GHz is just the maximum frequency we achieved for the compute tile.
But there is not a particular reason why you couldn't go higher than this.
It just depends on the design
—
Reply to this email directly, view it on GitHub
<#97 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AXTS27G7QWXW2NSXKDDRM432OR5GVAVCNFSM6AAAAABWL3TOBOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNBSGQZDIMBXGI>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
I would like to confirm the strategy used by FlooNoC to calculate the throughput of the NoC. Specifically, I am interested in understanding the behavior of FlooNoC under different test scenarios and how these scenarios impact throughput and latency metrics. The goal is to verify how the NoC performs under various conditions and to understand the relationship between the system's traffic patterns and its throughput and latency calculations.
Could you please provide insights on how FlooNoC calculates throughput and how it adjusts under different load conditions, burst behaviors, and any other relevant factors and approach that may affect these performance metrics?
FlooNoC has mentioned in its research paper that they have achieved a maximum throughput of 645Gbps per link, how is it actually calculated. Can you please guide so that I can repeat these steps and calculate the throughput?
The text was updated successfully, but these errors were encountered: