Executive summary
altifigence_npu_top · Full bundle B
This stored analysis run brings together 12 artifacts. The report links each observation to the source artifact.
| Waveform capture B | Signals 148 |
|---|---|
| Schematic snapshot B | Instances 42 |
| seed-full-analysis-run | Kernels 6 |
| Path analyzer B | Paths 31 |
| Memory dump B | Regions 6 |
| Event timeline B | Events 11.1k |
| Perf chart B | MAC avg 44.2% |
| Timing report B | WNS -0.07 ns |
| PPA snapshot B | Power 21.3 W |
| Evidence bundle B | Claims 14 |
| Data flow B | Nodes 36 |
| 3D canvas B | Grid 32 × 32 |
Input lineage
| Module | altifigence_npu_top |
|---|---|
| Source | rtl/altifigence-v002/LLM/rtl/top/altifigence_npu_top.sv |
| Run | altifigence-npu-full-v003 |
| Named view | Full bundle B |
| Artifacts | 12 |
Methodology
The report reads the saved analysis bundle and reorganizes its metrics, detail rows, and notes into document sections.
Each artifact retains its own units and status. Compare neighboring metrics within the same artifact, then follow its source path to the stored evidence.
| Waveform | signal transitions and capture timing |
|---|---|
| Schematic | module connections and structural evidence |
| Roofline | compute throughput and arithmetic intensity |
| Path | critical paths and timing evidence |
| Memory | memory traffic and bandwidth |
| Events | recorded event activity |
| Perf | performance samples and utilization trends |
| Timing | timing margins |
| PPA | power, performance, and area estimates |
| Signoff | stored validation evidence |
| Data Flow | data movement between compute stages |
| 3D Canvas | stored spatial and placement evidence |
Findings register
Flagged metrics from the stored artifacts.
| Waveform · MAC util | 47.5% |
|---|---|
| Waveform · Stalls | 518 |
| Schematic · Inline | 1 |
| Schematic · Diagnostics | 2 |
| Path · Worst slack | -0.07 ns |
| Path · Logic levels | 14 |
| Path · Fanout | 26 |
| Memory · Reads | 1.4 GB/s |
| Memory · Writes | 1.6 GB/s |
| Events · MAC | 36% |
| Events · Stalls | 18% |
| Perf · MAC avg | 44.2% |
| Perf · L2 write | 39.6% |
| Timing · WNS | -0.07 ns |
| Timing · TNS | -0.41 ns |
| Timing · Failing | 3 |
| PPA · Power | 21.3 W |
| PPA · LUT util | 67% |
| PPA · DSP util | 79% |
| PPA · Freq | 950 MHz |
| Signoff · Passed | 10 |
| Signoff · Open | 4 |
| Signoff · Coverage | 78% |
| Data Flow · Edges | 63 |
| Data Flow · Max depth | 11 |
| Data Flow · Hot path | HP→DISP→MAC |
| 3D Canvas · Avg util | 41.7% |
| 3D Canvas · Peak PE | 99% |
| 3D Canvas · Idle PEs | 173 |
Report summary
| Waveform capture B | Signals 148 |
|---|---|
| Schematic snapshot B | Instances 42 |
| seed-full-analysis-run | Kernels 6 |
| Path analyzer B | Paths 31 |
| Memory dump B | Regions 6 |
| Event timeline B | Events 11.1k |
| Perf chart B | MAC avg 44.2% |
| Timing report B | WNS -0.07 ns |
| PPA snapshot B | Power 21.3 W |
| Evidence bundle B | Claims 14 |
| Data flow B | Nodes 36 |
| 3D canvas B | Grid 32 × 32 |
01WaveformWaveform capture B
4981200 cycles · 32 cores · higher stall density
Interpretation
The Waveform section reviews signal transitions and capture timing. It reorganizes 4 metric(s), 2 detail row(s), and 0 note(s) from the “Waveform capture B” artifact into report form.
The first observation is MAC util = 47.5%. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 2.
The Waveform section contains critical-level signals. Use the 2 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is MAC util = 47.5%, and source path is analyze/altifigence-npu-full-v003/waveform.json. The interpretation is linked to stored analysis output.
The “Waveform capture B” result in Waveform requires additional review. Compare MAC util = 47.5% against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Waveform requires careful interpretation. MAC util = 47.5% and 2 flagged metric(s) remain the main review points.
Metrics
| Signals | 148 |
|---|---|
| Ticks | 4.98M |
| MAC util | 47.5% |
| Stalls | 518 |
Details
| Top scope | u_gemm_systolic_top |
|---|---|
| Playhead | cycle 90210 |
analyze/altifigence-npu-full-v003/waveform.json02SchematicSchematic snapshot B
Same hierarchy with fewer expanded inline blocks
Interpretation
The Schematic section reviews module connections and structural evidence. It reorganizes 4 metric(s), 0 detail row(s), and 0 note(s) from the “Schematic snapshot B” artifact into report form.
The first observation is Inline = 1. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 2.
The Schematic section contains warning-level signals. Use the 2 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Inline = 1, and source path is analyze/altifigence-npu-full-v003/schematic.json. The interpretation is linked to stored analysis output.
The “Schematic snapshot B” result in Schematic requires additional review. Compare Inline = 1 against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Schematic requires careful interpretation. Inline = 1 and 2 flagged metric(s) remain the main review points.
Metrics
| Instances | 42 |
|---|---|
| Nets | 186 |
| Inline | 1 |
| Diagnostics | 2 |
analyze/altifigence-npu-full-v003/schematic.json03Rooflineseed-full-analysis-run
Full bundle B · altifigence_npu_top
Interpretation
The Roofline section reviews compute throughput and arithmetic intensity. It reorganizes 4 metric(s), 0 detail row(s), and 0 note(s) from the “seed-full-analysis-run” artifact into report form.
The first observation is Kernels = 6. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 0.
Read the Roofline metrics alongside the stored evidence and neighboring artifacts.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Kernels = 6, and source path is analyze/altifigence-npu-full-v003/roofline.json. The interpretation is linked to stored analysis output.
The source artifact preserves the values shown in this report.
AI opinion
AI opinion: compare Kernels = 6 with neighboring artifacts before drawing a conclusion.
Metrics
| Kernels | 6 |
|---|---|
| Intensity | 11.8 |
| Achieved | 15.1 GOPS |
| Avg util | 41.7% |
analyze/altifigence-npu-full-v003/roofline.json04PathPath analyzer B
Weight dispatcher queue path is tighter
Interpretation
The Path section reviews critical paths and timing evidence. It reorganizes 4 metric(s), 2 detail row(s), and 0 note(s) from the “Path analyzer B” artifact into report form.
The first observation is Worst slack = -0.07 ns. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 3.
The Path section contains critical-level signals. Use the 3 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Worst slack = -0.07 ns, and source path is analyze/altifigence-npu-full-v003/path-analyzer.json. The interpretation is linked to stored analysis output.
The “Path analyzer B” result in Path requires additional review. Compare Worst slack = -0.07 ns against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Path requires careful interpretation. Worst slack = -0.07 ns and 3 flagged metric(s) remain the main review points.
Metrics
| Paths | 31 |
|---|---|
| Worst slack | -0.07 ns |
| Logic levels | 14 |
| Fanout | 26 |
Details
| Start | GEMM_weight_dispatcher/clk |
|---|---|
| End | mem_u_operation_queue/push |
analyze/altifigence-npu-full-v003/path-analyzer.json05MemoryMemory dump B
HP buffer pressure elevated
Interpretation
The Memory section reviews memory traffic and bandwidth. It reorganizes 4 metric(s), 2 detail row(s), and 0 note(s) from the “Memory dump B” artifact into report form.
The first observation is Reads = 1.4 GB/s. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 2.
The Memory section contains warning-level signals. Use the 2 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Reads = 1.4 GB/s, and source path is analyze/altifigence-npu-full-v003/memory.json. The interpretation is linked to stored analysis output.
The “Memory dump B” result in Memory requires additional review. Compare Reads = 1.4 GB/s against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Memory requires careful interpretation. Reads = 1.4 GB/s and 2 flagged metric(s) remain the main review points.
Metrics
| Regions | 6 |
|---|---|
| Reads | 1.4 GB/s |
| Writes | 1.6 GB/s |
| Watches | 3 |
Details
| Hot region | HP Buffer FIFO |
|---|---|
| Cursor | 0x1000_0140 |
analyze/altifigence-npu-full-v003/memory.json06EventsEvent timeline B
More systolic stalls in mid window
Interpretation
The Events section reviews recorded event activity. It reorganizes 4 metric(s), 2 detail row(s), and 0 note(s) from the “Event timeline B” artifact into report form.
The first observation is MAC = 36%. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 2.
The Events section contains critical-level signals. Use the 2 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is MAC = 36%, and source path is analyze/altifigence-npu-full-v003/events.json. The interpretation is linked to stored analysis output.
The “Event timeline B” result in Events requires additional review. Compare MAC = 36% against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Events requires careful interpretation. MAC = 36% and 2 flagged metric(s) remain the main review points.
Metrics
| Events | 11.1k |
|---|---|
| Cycles | 220 |
| MAC | 36% |
| Stalls | 18% |
Details
| Peak lane | SYSTOLIC_STALL |
|---|---|
| Marker | cycle 142 |
analyze/altifigence-npu-full-v003/events.json07PerfPerf chart B
Lower MAC util with heavier DMA write share
Interpretation
The Perf section reviews performance samples and utilization trends. It reorganizes 4 metric(s), 0 detail row(s), and 0 note(s) from the “Perf chart B” artifact into report form.
The first observation is MAC avg = 44.2%. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 2.
The Perf section contains warning-level signals. Use the 2 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is MAC avg = 44.2%, and source path is analyze/altifigence-npu-full-v003/perf.json. The interpretation is linked to stored analysis output.
The “Perf chart B” result in Perf requires additional review. Compare MAC avg = 44.2% against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Perf requires careful interpretation. MAC avg = 44.2% and 2 flagged metric(s) remain the main review points.
Graph
Metrics
| MAC avg | 44.2% |
|---|---|
| L2 read | 35.8% |
| L2 write | 39.6% |
| Samples | 256 |
analyze/altifigence-npu-full-v003/perf.json08TimingTiming report B
Post-impl summary — one failing path cluster
Interpretation
The Timing section reviews timing margins. It reorganizes 4 metric(s), 2 detail row(s), and 0 note(s) from the “Timing report B” artifact into report form.
The first observation is WNS = -0.07 ns. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 3.
The Timing section contains critical-level signals. Use the 3 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is WNS = -0.07 ns, and source path is analyze/altifigence-npu-full-v003/timing.json. The interpretation is linked to stored analysis output.
The “Timing report B” result in Timing requires additional review. Compare WNS = -0.07 ns against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Timing requires careful interpretation. WNS = -0.07 ns and 3 flagged metric(s) remain the main review points.
Metrics
| WNS | -0.07 ns |
|---|---|
| TNS | -0.41 ns |
| Failing | 3 |
| Endpoints | 18.2k |
Details
| LUT | 43,910 |
|---|---|
| DSP | 528 |
analyze/altifigence-npu-full-v003/timing.json09PPAPPA snapshot B
Higher power for lower achieved throughput
Interpretation
The PPA section reviews power, performance, and area estimates. It reorganizes 4 metric(s), 0 detail row(s), and 0 note(s) from the “PPA snapshot B” artifact into report form.
The first observation is Power = 21.3 W. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 4.
The PPA section contains critical-level signals. Use the 4 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Power = 21.3 W, and source path is analyze/altifigence-npu-full-v003/ppa.json. The interpretation is linked to stored analysis output.
The “PPA snapshot B” result in PPA requires additional review. Compare Power = 21.3 W against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: PPA requires careful interpretation. Power = 21.3 W and 4 flagged metric(s) remain the main review points.
Metrics
| Power | 21.3 W |
|---|---|
| LUT util | 67% |
| DSP util | 79% |
| Freq | 950 MHz |
analyze/altifigence-npu-full-v003/ppa.json10SignoffEvidence bundle B
Regression vs bundle A on stall and timing claims
Interpretation
The Signoff section reviews stored validation evidence. It reorganizes 4 metric(s), 0 detail row(s), and 1 note(s) from the “Evidence bundle B” artifact into report form.
The first observation is Passed = 10. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 3.
The Signoff section contains critical-level signals. Use the 3 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Passed = 10, and source path is analyze/altifigence-npu-full-v003/evidence.json. The interpretation is linked to stored analysis output.
The “Evidence bundle B” result in Signoff requires additional review. Compare Passed = 10 against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Signoff requires careful interpretation. Passed = 10 and 3 flagged metric(s) remain the main review points.
Metrics
| Claims | 14 |
|---|---|
| Passed | 10 |
| Open | 4 |
| Coverage | 78% |
Notes
Regression: timing closure and stall-rate claims diverged from bundle A.
analyze/altifigence-npu-full-v003/evidence.json11Data FlowData flow B
HP buffer back-pressure adds a feedback edge into the dispatcher
Interpretation
The Data Flow section reviews data movement between compute stages. It reorganizes 4 metric(s), 3 detail row(s), and 1 note(s) from the “Data flow B” artifact into report form.
The first observation is Edges = 63. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 3.
The Data Flow section contains critical-level signals. Use the 3 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Edges = 63, and source path is analyze/altifigence-npu-full-v003/dataFlow.json. The interpretation is linked to stored analysis output.
The “Data flow B” result in Data Flow requires additional review. Compare Edges = 63 against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: Data Flow requires careful interpretation. Edges = 63 and 3 flagged metric(s) remain the main review points.
Metrics
| Nodes | 36 |
|---|---|
| Edges | 63 |
| Max depth | 11 |
| Hot path | HP→DISP→MAC |
Details
| Source node | axi_hp_weight_dma |
|---|---|
| Sink node | result_packer |
| Feedback | hp_buffer_full → dispatcher |
Notes
Back-pressure feedback edge present in this run; raises the longest dependency chain to 11.
analyze/altifigence-npu-full-v003/dataFlow.json123D Canvas3D canvas B
Hot cluster in the south-east quadrant from DMA write stalls
Interpretation
The 3D Canvas section reviews stored spatial and placement evidence. It reorganizes 4 metric(s), 2 detail row(s), and 1 note(s) from the “3D canvas B” artifact into report form.
The first observation is Avg util = 41.7%. This value should be read against the neighboring metrics in the same artifact; the current flagged metric count is 3.
The 3D Canvas section contains critical-level signals. Use the 3 flagged metric(s) to trace cause and scope.
Evidence basis
This section summarizes 4 metric evidence item(s). The representative metric is Avg util = 41.7%, and source path is analyze/altifigence-npu-full-v003/canvas3d.json. The interpretation is linked to stored analysis output.
The “3D canvas B” result in 3D Canvas requires additional review. Compare Avg util = 41.7% against neighboring artifacts and inspect the detailed path in the full viewer.
AI opinion
AI opinion: 3D Canvas requires careful interpretation. Avg util = 41.7% and 3 flagged metric(s) remain the main review points.
Metrics
| Grid | 32 × 32 |
|---|---|
| Avg util | 41.7% |
| Peak PE | 99% |
| Idle PEs | 173 |
Details
| Hottest tile | row 27 · col 29 |
|---|---|
| Cold band | cols 0–6 starved |
Notes
Utilisation skewed: a hot SE cluster coexists with a starved west band under heavier DMA writes.
analyze/altifigence-npu-full-v003/canvas3d.jsonTraceability matrix
| Waveform capture B | analyze/altifigence-npu-full-v003/waveform.json |
|---|---|
| Schematic snapshot B | analyze/altifigence-npu-full-v003/schematic.json |
| seed-full-analysis-run | analyze/altifigence-npu-full-v003/roofline.json |
| Path analyzer B | analyze/altifigence-npu-full-v003/path-analyzer.json |
| Memory dump B | analyze/altifigence-npu-full-v003/memory.json |
| Event timeline B | analyze/altifigence-npu-full-v003/events.json |
| Perf chart B | analyze/altifigence-npu-full-v003/perf.json |
| Timing report B | analyze/altifigence-npu-full-v003/timing.json |
| PPA snapshot B | analyze/altifigence-npu-full-v003/ppa.json |
| Evidence bundle B | analyze/altifigence-npu-full-v003/evidence.json |
| Data flow B | analyze/altifigence-npu-full-v003/dataFlow.json |
| 3D canvas B | analyze/altifigence-npu-full-v003/canvas3d.json |
Technical conclusion
The Perf artifact reports MAC avg = 44.2%, with two metrics flagged for review. Compare the heavier DMA write share against the Memory and Waveform artifacts.
| Waveform · MAC util | 47.5% |
|---|---|
| Waveform · Stalls | 518 |
| Schematic · Inline | 1 |
| Schematic · Diagnostics | 2 |
| Path · Worst slack | -0.07 ns |
| Path · Logic levels | 14 |
| Path · Fanout | 26 |
| Memory · Reads | 1.4 GB/s |
| Memory · Writes | 1.6 GB/s |
| Events · MAC | 36% |
| Events · Stalls | 18% |
| Perf · MAC avg | 44.2% |
| Perf · L2 write | 39.6% |
| Timing · WNS | -0.07 ns |
| Timing · TNS | -0.41 ns |
| Timing · Failing | 3 |
| PPA · Power | 21.3 W |
| PPA · LUT util | 67% |
| PPA · DSP util | 79% |
| PPA · Freq | 950 MHz |
| Signoff · Passed | 10 |
| Signoff · Open | 4 |
| Signoff · Coverage | 78% |
| Data Flow · Edges | 63 |
| Data Flow · Max depth | 11 |
| Data Flow · Hot path | HP→DISP→MAC |
| 3D Canvas · Avg util | 41.7% |
| 3D Canvas · Peak PE | 99% |
| 3D Canvas · Idle PEs | 173 |







