Is l1tex__data_bank_reads.sum*128bytes the L1 traffic？
Hello,
Thank you for using Nsight Graphics and your question on L1TEX. Please have a look at the user guide here to see if this provides you more details on L1TEX Advanced Learning :: Nsight Graphics Documentation. Also, see the details about L1 Throughput in the same user guide in the section here Advanced Learning :: Nsight Graphics Documentation.
I’ll check with the engineering team to get more info on the specific metric
Hello,
I got more details from our engineering manager for your question above. See below.
The standard way to measure L1TEX bandwidth for tagged memory is l1tex__t_sectors * 32. Access to this metric requires the Pro version of Nsight Graphics.
Tagged memory includes local, global, texture, surface. It does not include groupshared or 3D attributes.
The l1tex__data_bank_reads counter is not necessarily proportional to requested or achieved bandwidth. It is more of an internal activity indicator.
@dwoods Can you share a bit more about what l1tex__data_bank_reads actually represents? My hypothesis is that it should be proportional to the number of data banks accessed (both as shared memory or L1 cache), but this hypothesis is at odds with the profiling results of my kernel (more details here), so I would really like to learn about the definition of this counter. Thanks!
Hello,
I believe you have received feedback on this question via your post here High shared memory usage but low l1tex__data_bank_reads - #3 by Curefab. Is that correct?
Yes, thanks @dwoods !