Profile · Profyle

Find every bottleneck before you spend a GPU-hour fixing it

Profyle captures Nsight and CUPTI traces on the target GPU, classifies each bottleneck and attributes its cost to specific kernels and layers. It is the front door of the Turbokern loop.

Profile summaryExample
Workload
Llama 3.1 70B decode
GPU
H100 SXM
Bound by
Memory, attention
Top opportunity
41% of step time
Next stage
Kernl, Quantra
Example. Values are illustrative.

Capabilities

What Profyle does

01

One-command profiling

Capture timeline and kernel traces on DGX, HGX and Jetson with a single run.

02

Bottleneck classification

Learned models label each hotspot as compute, memory, launch or serving bound.

03

Cost attribution

Latency and throughput cost assigned to individual kernels and layers.

04

Roofline and occupancy

Each kernel placed on the roofline, with a plain-language root cause.

05

Opportunity map

A ranked, machine-readable list the optimizers consume directly.

06

Regression watch

Re-profiles on every model or GPU change and flags drift.

Opportunity map

From trace to ranked opportunities

Profyle reads the timeline so a person does not have to. It attributes time to kernels and layers, classifies why each is slow, and estimates what fixing it is worth.

The output is a machine-readable map that Kernl, Quantra and Servex act on directly.

Decode step · H100 · one layerExample
Stream 0
Stream 1
Memcpy
CPU launch
0 µs350 µs700 µs
Attention (memory-bound)GEMM waiting on HBMCopiesOther kernels
Example opportunity map. Values are illustrative.
RankBottleneckClassShare
1Attention, decodeMemory-bound41%
2MLP GEMMsLow tensor-core use23%
3Launch gapsCPU-bound launch9%

How it works

Inside a Profyle run

  1. 01Minutes

    Capture

    Run the workload under Nsight Systems, Nsight Compute and CUPTI on the target GPU.

  2. 02Seconds

    Classify

    Trace features go through the bottleneck classifier.

  3. 03Seconds

    Attribute

    Cost is assigned to kernels and layers and ranked.

  4. 04Instant

    Hand off

    The opportunity map goes to Kernl, Quantra and Servex.

Specification

Inputs, outputs and hardware

ItemDetailStatus
InputModel or workload, target GPU, representative inputsPilot
Trace sourcesNsight Systems, Nsight Compute, CUPTIPilot
TargetsH100, H200 (pilot); Blackwell, Jetson OrinPilot · Planned
OutputOpportunity map (JSON), HTML reportPilot
Workload classesLLM inference first; vision and speech nextNext
Narrative summariesNatural-language root-cause write-upsPlanned

FAQ

Questions engineers ask

Do I need Nsight installed?

No. Profyle brings its own capture tooling for the GPU it runs on.

Can I use Profyle on its own?

Yes. Many teams will start with a profile to see what optimization could be worth before running anything else.

Does profiling slow production?

Profyle runs on a separate replica or on our GPUs, not on production traffic.

What if the bottleneck is outside the GPU?

Profyle reports CPU launch gaps, data loading and serving queues too, and says when the GPU is not the limit.

Next step

Profile one workload first

Start the pilot with Profyle and get a ranked opportunity map for your model on your GPU.