# \#performance-metrics **URL:** https://forums.developer.nvidia.com/tag/performance-metrics/357.md [Latest](https://forums.developer.nvidia.com/latest.md) · [Categories](https://forums.developer.nvidia.com/categories.md) · [Tags](https://forums.developer.nvidia.com/tags.md) --- ## [RTX PRO 4000 Blackwell: FP16/FP8/FP6/FP4 throughput with FP32 accumulation?](https://forums.developer.nvidia.com/t/rtx-pro-4000-blackwell-fp16-fp8-fp6-fp4-throughput-with-fp32-accumulation/382903)
**Author:** [@misiekuba](https://forums.developer.nvidia.com/u/misiekuba)\ **Replies:** 1\ **Last updated:** [September 11, 2026, 11:07am UTC](https://forums.developer.nvidia.com/t/rtx-pro-4000-blackwell-fp16-fp8-fp6-fp4-throughput-with-fp32-accumulation/382903 "2026-09-11T11:07:40Z")
Hi, I am trying to determine the actual Tensor Core throughput of the RTX PRO 4000 Blackwell for mixed-precision training, specifically when lower-precision inputs are accumulated into FP32. The main question is: What… --- ## [Reason for warp stall not clear or unknown](https://forums.developer.nvidia.com/t/reason-for-warp-stall-not-clear-or-unknown/376233)
**Author:** [@marc.clasca](https://forums.developer.nvidia.com/u/marc.clasca)\ **Replies:** 0\ **Last updated:** [July 9, 2026, 3:46pm UTC](https://forums.developer.nvidia.com/t/reason-for-warp-stall-not-clear-or-unknown/376233 "2026-07-09T15:46:30Z")
Hey there, I have recently done a CUDA programming course where we were guided through some example CUDA code, like matrix multiplication. When analyzing the kernels with Nsight Compute, I noticed that ncu reports a com… --- ## [GB10 really does hit ~1 PFLOP NVFP4 (2:4 sparse) — measured, with an open-source tool to reproduce it](https://forums.developer.nvidia.com/t/gb10-really-does-hit-1-pflop-nvfp4-2-4-sparse-measured-with-an-open-source-tool-to-reproduce-it/373618)
**Author:** [@adg1](https://forums.developer.nvidia.com/u/adg1)\ **Replies:** 22\ **Last updated:** [June 25, 2026, 6:21am UTC](https://forums.developer.nvidia.com/t/gb10-really-does-hit-1-pflop-nvfp4-2-4-sparse-measured-with-an-open-source-tool-to-reproduce-it/373618 "2026-06-25T06:21:15Z")
GB10 really does hit ~1 PFLOP NVFP4 (2:4 sparse) — measured, with an open-source tool to reproduce it TL;DR. I built a small CLI, nvfp4bench, to measure peak achievable NVFP4 tensor-core throughput on DGX Spark (GB10, s… --- ## [ASUS Ascent GX10 — Second ConnectX-7 PCIe slot wired at Gen5 x2 instead of x4 (50% bandwidth loss vs DGX Spark FE)](https://forums.developer.nvidia.com/t/asus-ascent-gx10-second-connectx-7-pcie-slot-wired-at-gen5-x2-instead-of-x4-50-bandwidth-loss-vs-dgx-spark-fe/367853)
**Author:** [@RafaelPinilla](https://forums.developer.nvidia.com/u/RafaelPinilla)\ **Replies:** 2\ **Last updated:** [May 6, 2026, 11:39am UTC](https://forums.developer.nvidia.com/t/asus-ascent-gx10-second-connectx-7-pcie-slot-wired-at-gen5-x2-instead-of-x4-50-bandwidth-loss-vs-dgx-spark-fe/367853 "2026-05-06T11:39:23Z")
ASUS Ascent GX10 — Second ConnectX-7 PCIe slot wired at Gen5 x2 instead of x4 (50% bandwidth loss vs DGX Spark FE) Summary The ASUS Ascent GX10 is marketed as functionally equivalent to the NVIDIA DGX Spark FE, with ASU… --- ## [How to correlate range profiling metrics with a certain kernel?](https://forums.developer.nvidia.com/t/how-to-correlate-range-profiling-metrics-with-a-certain-kernel/340811)
**Author:** [@tianyao\_zhou](https://forums.developer.nvidia.com/u/tianyao_zhou)\ **Replies:** 2\ **Last updated:** [January 2, 2026, 2:22pm UTC](https://forums.developer.nvidia.com/t/how-to-correlate-range-profiling-metrics-with-a-certain-kernel/340811 "2026-01-02T14:22:49Z")
Hi, I’m using the latest range profiling apis to get profiling metrics such as nvlrx\_bytes.avg, nvlrx\_bytes\_data\_user.avg. And I choose auto range & kernel replay, which regard one kernel as one range. After profiling, … --- ## [Performance counters of Bluefield3](https://forums.developer.nvidia.com/t/performance-counters-of-bluefield3/343696)
**Author:** [@hua.zhang.2106108](https://forums.developer.nvidia.com/u/hua.zhang.2106108)\ **Replies:** 1\ **Last updated:** [October 10, 2025, 2:53am UTC](https://forums.developer.nvidia.com/t/performance-counters-of-bluefield3/343696 "2025-10-10T02:53:38Z")
Bluefield3 provides several performance counters via mlx\_pmc driver under /sys/class/hwmon. I reviewed the latest hardware monitoring (hwmon) documentation:https://docs.nvidia.com/networking/display/nvidia-bluefield-bsp… --- ## [RTX cards benchmarks - how to compare](https://forums.developer.nvidia.com/t/rtx-cards-benchmarks-how-to-compare/339862)
**Author:** [@bizmate](https://forums.developer.nvidia.com/u/bizmate)\ **Replies:** 2\ **Last updated:** [July 22, 2025, 7:21pm UTC](https://forums.developer.nvidia.com/t/rtx-cards-benchmarks-how-to-compare/339862 "2025-07-22T19:21:43Z")
I have looked around and cannot find a definitive lamers simplified comparison on AI specific cards benchmarks. By AI I mean RTX cards that can executed AI specific jobs like running Ollama models, Yolo/Darknet etc and h… --- ## [How to Get the Exact Amount of Resources the GPU Uses at the Moment (e.g., Used Tensor Cores) Regardless of the Running Process](https://forums.developer.nvidia.com/t/how-to-get-the-exact-amount-of-resources-the-gpu-uses-at-the-moment-e-g-used-tensor-cores-regardless-of-the-running-process/319410)
**Author:** [@rsmrsmeeman](https://forums.developer.nvidia.com/u/rsmrsmeeman)\ **Replies:** 5\ **Last updated:** [January 13, 2025, 2:42pm UTC](https://forums.developer.nvidia.com/t/how-to-get-the-exact-amount-of-resources-the-gpu-uses-at-the-moment-e-g-used-tensor-cores-regardless-of-the-running-process/319410 "2025-01-13T14:42:22Z")
As the Nsight tool relies on defining a process or executable to attach to, and it only shows resources allocated to a specific process, it seems like there is no hardware feature from Nvidia to monitor totally what reso… --- ## [Very low Recall for HG002 germline benchmark](https://forums.developer.nvidia.com/t/very-low-recall-for-hg002-germline-benchmark/299045)
**Author:** [@kuhnsa3](https://forums.developer.nvidia.com/u/kuhnsa3)\ **Replies:** 1\ **Last updated:** [July 24, 2024, 1:56pm UTC](https://forums.developer.nvidia.com/t/very-low-recall-for-hg002-germline-benchmark/299045 "2024-07-24T13:56:20Z")
Hello there, Has anyone benchmarked the parabricks germline workflow with GIAB sample HG002 using hap.py? Here are the results of my comparison using different options related to best performance and hardware configura… --- ## [Seeking Insights on the Performance Lifecycle of NVIDIA H100 GPUs](https://forums.developer.nvidia.com/t/seeking-insights-on-the-performance-lifecycle-of-nvidia-h100-gpus/294554)
**Author:** [@jesse61](https://forums.developer.nvidia.com/u/jesse61)\ **Replies:** 1\ **Last updated:** [May 30, 2024, 4:42am UTC](https://forums.developer.nvidia.com/t/seeking-insights-on-the-performance-lifecycle-of-nvidia-h100-gpus/294554 "2024-05-30T04:42:50Z")
Hello NVIDIA Developer Community, I’m evaluating an investment and need insights into the performance lifecycle of the NVIDIA H100 Tensor Core GPUs. Specifically, I am interested in: Lifecycle: Expected performance li… --- ## [Cannot use model-analyzer on ONNX classification model with dynamic input](https://forums.developer.nvidia.com/t/cannot-use-model-analyzer-on-onnx-classification-model-with-dynamic-input/291886)
**Author:** [@simon.renuart](https://forums.developer.nvidia.com/u/simon.renuart)\ **Replies:** 2\ **Last updated:** [May 28, 2024, 7:49am UTC](https://forums.developer.nvidia.com/t/cannot-use-model-analyzer-on-onnx-classification-model-with-dynamic-input/291886 "2024-05-28T07:49:57Z")
• RTX 3060 • DeepStream Version 6.4 - Docker • Cannot run model-analyzer on a model I am currently profiling several models with model-analyzer. I can’t manage to do this with one of my models and I’d like to have mo… --- ## [What is PCIe to BAR Requests mean?](https://forums.developer.nvidia.com/t/what-is-pcie-to-bar-requests-mean/287581)
**Author:** [@yunie9447](https://forums.developer.nvidia.com/u/yunie9447)\ **Replies:** 0\ **Last updated:** [March 27, 2024, 9:30am UTC](https://forums.developer.nvidia.com/t/what-is-pcie-to-bar-requests-mean/287581 "2024-03-27T09:30:25Z")
Hello Guys, I was using Nsight Graphics GPU trace mode to capture GPU performance metrics. In my case, I noticed that there are heavy PCIe throughput during rendering and there are a lot of PCIe to BAR Requests. FYI, My … --- ## [Which foundational libraries do the high-frequency GPU metrics in Nsight Systems come from?](https://forums.developer.nvidia.com/t/which-foundational-libraries-do-the-high-frequency-gpu-metrics-in-nsight-systems-come-from/285827)
**Author:** [@slei.chu](https://forums.developer.nvidia.com/u/slei.chu)\ **Replies:** 4\ **Last updated:** [March 14, 2024, 1:31am UTC](https://forums.developer.nvidia.com/t/which-foundational-libraries-do-the-high-frequency-gpu-metrics-in-nsight-systems-come-from/285827 "2024-03-14T01:31:26Z")
Regarding the high-frequency GPU metrics(such as GPC clock frequency/SM active and etc.) collection in Nsight Systems, which interfaces do these metrics originate from? To my knowledge, the Profiling API in CUPTI recomme… --- ## [GPU metrics not working in Nsight System / Compute](https://forums.developer.nvidia.com/t/gpu-metrics-not-working-in-nsight-system-compute/275646)
**Author:** [@antoinerichermoz](https://forums.developer.nvidia.com/u/antoinerichermoz)\ **Replies:** 2\ **Last updated:** [December 12, 2023, 9:29am UTC](https://forums.developer.nvidia.com/t/gpu-metrics-not-working-in-nsight-system-compute/275646 "2023-12-12T09:29:57Z")
GPU performance metrics are broken in a weird way when using Nsight System / Compute. The timeline is mostly empty, only showing metrics on a few locations. On some runs it is a bit better, but overall it is pretty unu… --- ## [The meaning of KPI metrics of multi-class classification models](https://forums.developer.nvidia.com/t/the-meaning-of-kpi-metrics-of-multi-class-classification-models/248139)
**Author:** [@swka1043338](https://forums.developer.nvidia.com/u/swka1043338)\ **Replies:** 20\ **Last updated:** [April 17, 2023, 6:58am UTC](https://forums.developer.nvidia.com/t/the-meaning-of-kpi-metrics-of-multi-class-classification-models/248139 "2023-04-17T06:58:48Z")
In Tao-Toolkit-API AutoML multi-class classification notebook, there is a metric of models named KPI. What does it mean ? Does it mean classification accuracy ? --- ## [Gtx 1080 ti showing zero fan rpm on nvidia performance overlay](https://forums.developer.nvidia.com/t/gtx-1080-ti-showing-zero-fan-rpm-on-nvidia-performance-overlay/238872)
**Author:** [@shahbazsaif92](https://forums.developer.nvidia.com/u/shahbazsaif92)\ **Replies:** 2\ **Last updated:** [March 29, 2023, 3:56am UTC](https://forums.developer.nvidia.com/t/gtx-1080-ti-showing-zero-fan-rpm-on-nvidia-performance-overlay/238872 "2023-03-29T03:56:59Z")
need help does this gtx 1080 ti dont read fan speed or its driver bug am on latest driver by the way --- ## [Unable to view memory usage with nvidia-smi, "Not available in WDDM driver model"?](https://forums.developer.nvidia.com/t/unable-to-view-memory-usage-with-nvidia-smi-not-available-in-wddm-driver-model/244700)
**Author:** [@infinitay](https://forums.developer.nvidia.com/u/infinitay)\ **Replies:** 0\ **Last updated:** [March 2, 2023, 6:40pm UTC](https://forums.developer.nvidia.com/t/unable-to-view-memory-usage-with-nvidia-smi-not-available-in-wddm-driver-model/244700 "2023-03-02T18:40:06Z")
I can’t see how much GPU memory applications when I run nvidia-smi. The memory usage is listed as “N/A” for some reason. FWIW running nvidia-smi -q gives me Used GPU Memory: Not available in WDDM driver model. I’ve seen … --- ## [Jetson： Failed to start nvpmode1 server](https://forums.developer.nvidia.com/t/jetson-failed-to-start-nvpmode1-server/240948)
**Author:** [@waikiki](https://forums.developer.nvidia.com/u/waikiki)\ **Replies:** 10\ **Last updated:** [February 1, 2023, 7:49pm UTC](https://forums.developer.nvidia.com/t/jetson-failed-to-start-nvpmode1-server/240948 "2023-02-01T19:49:54Z")
Jetson： Failed to start nvpmode1 server。 GUI cannot be started at boot --- ## [Running perf inside a docker container / Docker seccomp profiles](https://forums.developer.nvidia.com/t/running-perf-inside-a-docker-container-docker-seccomp-profiles/240403)
**Author:** [@n.hiebl](https://forums.developer.nvidia.com/u/n.hiebl)\ **Replies:** 1\ **Last updated:** [January 27, 2023, 1:27pm UTC](https://forums.developer.nvidia.com/t/running-perf-inside-a-docker-container-docker-seccomp-profiles/240403 "2023-01-27T13:27:09Z")
Hi! I’m currently attempting to profile an application running inside a docker container on a Jetson TX2, using nvidia-docker. I would prefer to do this using the perf tool, as the stats it provides are exactly what I ne… --- ## [How is TOPS metric calculated with any AI model.?](https://forums.developer.nvidia.com/t/how-is-tops-metric-calculated-with-any-ai-model/237815)
**Author:** [@MJ\_SW](https://forums.developer.nvidia.com/u/MJ_SW)\ **Replies:** 1\ **Last updated:** [December 26, 2022, 2:36am UTC](https://forums.developer.nvidia.com/t/how-is-tops-metric-calculated-with-any-ai-model/237815 "2022-12-26T02:36:46Z")
Hello , As per the https://developer.nvidia.com/embedded/jetson-modules,the Jetson Xavier NX supports 21 TOPS in AI Performance For example if the use case is people detection, tracking and counting, how many TOPS … --- ## [Perf\_client (or perf\_analyser) unable to run on real input data](https://forums.developer.nvidia.com/t/perf-client-or-perf-analyser-unable-to-run-on-real-input-data/235578)
**Author:** [@anil.choubey95](https://forums.developer.nvidia.com/u/anil.choubey95)\ **Replies:** 0\ **Last updated:** [November 28, 2022, 10:21am UTC](https://forums.developer.nvidia.com/t/perf-client-or-perf-analyser-unable-to-run-on-real-input-data/235578 "2022-11-28T10:21:50Z")
I am refering this doc I are trying to do latency and throughput evaluation using perf\_client with the same Triton’s “end-to-end” ensemble (see config.pbtxt) that It is working fine with random self generated input data… --- ## [TAO pretrained model performance metrics changed, documentation tables need clarified](https://forums.developer.nvidia.com/t/tao-pretrained-model-performance-metrics-changed-documentation-tables-need-clarified/222076)
**Author:** [@rsc44](https://forums.developer.nvidia.com/u/rsc44)\ **Replies:** 2\ **Last updated:** [August 16, 2022, 3:19am UTC](https://forums.developer.nvidia.com/t/tao-pretrained-model-performance-metrics-changed-documentation-tables-need-clarified/222076 "2022-08-16T03:19:49Z")
I recently checked the performance metrics for the pre-trained models, specifically peoplenet. I noticed that the stats have changed, and according to your charts the FPS is less than previously reported. This is what … --- ## [Difference between thread\_inst\_executed metrics](https://forums.developer.nvidia.com/t/difference-between-thread-inst-executed-metrics/217587)
**Author:** [@spraesi](https://forums.developer.nvidia.com/u/spraesi)\ **Replies:** 5\ **Last updated:** [June 27, 2022, 6:29pm UTC](https://forums.developer.nvidia.com/t/difference-between-thread-inst-executed-metrics/217587 "2022-06-27T18:29:12Z")
Hi there - I was wondering what the difference between the metrics smsp\_\_thread\_inst\_executed and smsp\_\_sass\_thread\_inst\_executed is. I am guessing it is just the same? --- ## [How to figure out the ratio of the number of GPU-cycles fp64 to the total number of cycles?](https://forums.developer.nvidia.com/t/how-to-figure-out-the-ratio-of-the-number-of-gpu-cycles-fp64-to-the-total-number-of-cycles/202104)
**Author:** [@Vasilyev\_v](https://forums.developer.nvidia.com/u/Vasilyev_v)\ **Replies:** 1\ **Last updated:** [April 25, 2022, 10:27pm UTC](https://forums.developer.nvidia.com/t/how-to-figure-out-the-ratio-of-the-number-of-gpu-cycles-fp64-to-the-total-number-of-cycles/202104 "2022-04-25T22:27:41Z")
Hello! I’d like to estimate the ratio of the number of GPU (AI100) cycles when FP64:{FADD, DMUL, DFMA} instructions is executed to the total numbers of GPU cycles is active. Something like to The percentage of GPU cycl… --- ## [What is the meaning of sm\_\_pipe\_fp64\_cycles\_active\[burst/sustained\]](https://forums.developer.nvidia.com/t/what-is-the-meaning-of-sm-pipe-fp64-cycles-active-burst-sustained/202391)
**Author:** [@Vasilyev\_v](https://forums.developer.nvidia.com/u/Vasilyev_v)\ **Replies:** 3\ **Last updated:** [February 9, 2022, 10:42pm UTC](https://forums.developer.nvidia.com/t/what-is-the-meaning-of-sm-pipe-fp64-cycles-active-burst-sustained/202391 "2022-02-09T22:42:40Z")
Hello! I’ve read the documentation about burst/sustained metrics. But, unfortunately, I didn’t understand the meaning the burst and sustained. In particular, in relation to the output (AI100): sm\_\_pipe\_fp64\_cycles\_act… --- ## [How to calculate the total count of bytes were loaded/stored from/to global device memory?](https://forums.developer.nvidia.com/t/how-to-calculate-the-total-count-of-bytes-were-loaded-stored-from-to-global-device-memory/202628)
**Author:** [@Vasilyev\_v](https://forums.developer.nvidia.com/u/Vasilyev_v)\ **Replies:** 1\ **Last updated:** [February 7, 2022, 10:16am UTC](https://forums.developer.nvidia.com/t/how-to-calculate-the-total-count-of-bytes-were-loaded-stored-from-to-global-device-memory/202628 "2022-02-07T10:16:35Z")
Hello! I have the nv-nsight-cu-cli --set=full --details-all output. I need to caclulate the total count of bytes were loaded/stored from/to global memory. Is there any guidlines? Which metrics I should use? Thanks! --- ## [Difference sm\_\_cycles\_elapsed/smsp\_\_cycles\_elapsed and sm\_\_inst\_executed/smsp\_\_inst\_executed?](https://forums.developer.nvidia.com/t/difference-sm-cycles-elapsed-smsp-cycles-elapsed-and-sm-inst-executed-smsp-inst-executed/202101)
**Author:** [@Vasilyev\_v](https://forums.developer.nvidia.com/u/Vasilyev_v)\ **Replies:** 5\ **Last updated:** [February 2, 2022, 4:52pm UTC](https://forums.developer.nvidia.com/t/difference-sm-cycles-elapsed-smsp-cycles-elapsed-and-sm-inst-executed-smsp-inst-executed/202101 "2022-02-02T16:52:26Z")
Hello! I have a dumb question :), but I’m a little confused. I don’t understand clearly the difference between this metrics, and how to the cycles counters related to the HW architecture. My current understanding (AI10… --- ## [How to calculate the number of DADD instructions of a kernel on GPU?](https://forums.developer.nvidia.com/t/how-to-calculate-the-number-of-dadd-instructions-of-a-kernel-on-gpu/202103)
**Author:** [@Vasilyev\_v](https://forums.developer.nvidia.com/u/Vasilyev_v)\ **Replies:** 1\ **Last updated:** [February 2, 2022, 1:14pm UTC](https://forums.developer.nvidia.com/t/how-to-calculate-the-number-of-dadd-instructions-of-a-kernel-on-gpu/202103 "2022-02-02T13:14:05Z")
Hello! I need to calculate the number of DADD instructions executed on device AI100. I guess to use the expression: DADD = smsp\_\_sass\_thread\_inst\_executed\_op\_dadd\_pred\_on.sum.per\_cycle\_elapsed \* smsp\_\_cycles\_elapsed.av… --- ## [How to get current GPU usage percentage with C++ program?](https://forums.developer.nvidia.com/t/how-to-get-current-gpu-usage-percentage-with-c-program/198198)
**Author:** [@user97979](https://forums.developer.nvidia.com/u/user97979)\ **Replies:** 1\ **Last updated:** [January 11, 2022, 4:41am UTC](https://forums.developer.nvidia.com/t/how-to-get-current-gpu-usage-percentage-with-c-program/198198 "2022-01-11T04:41:10Z")
I have Cuda 11.5 toolkit installed. And I have found samples there named : deviceQuery. This will list out all the necessary information of the device but I am not able to find any api through which I can get current GPU… --- ## [Nsight Compute metrics value confused](https://forums.developer.nvidia.com/t/nsight-compute-metrics-value-confused/197032)
**Author:** [@user17130](https://forums.developer.nvidia.com/u/user17130)\ **Replies:** 1\ **Last updated:** [December 14, 2021, 12:26pm UTC](https://forums.developer.nvidia.com/t/nsight-compute-metrics-value-confused/197032 "2021-12-14T12:26:59Z")
Hi, I am testing MPS on Quadro RTX 6000. Whether MPS is enable or not, I start two processes, they all launch kernel on GPU(first process configuration:vector\_add\<\<\<1,1\>\>\>, second process configuration:vector\_add\<\<\<1,32\>\>… [Next page](https://forums.developer.nvidia.com/tag/performance-metrics/357.md?match_all_tags=true&page=1&tags%5B%5D=performance-metrics)