Document Version: 1.0 Analysis Date: January 2026 Analyst: Claude
Code (Opus 4.5) Source Repository: /home/ubuntu/oshun/serwaa/
Executive Summary#
This document establishes the performance baseline for the Serwaa codebase
(future Psyche domain), covering:
- Avatar rendering performance characteristics
- End-to-end response latency targets
- Conferencing integration latency requirements
- Quality metrics and validation thresholds
- Monitoring and benchmarking infrastructure
| Metric |
Target |
Warning |
Critical |
| Avatar Rendering |
30 FPS @ 1024x1024 |
<27 FPS |
<20 FPS |
| STT Latency |
100 ms |
150 ms |
200 ms |
| TTS Latency |
100 ms |
150 ms |
200 ms |
| LLM Inference |
200 ms |
300 ms |
400 ms |
| End-to-End Response |
500 ms |
750 ms |
1000 ms |
| A/V Sync Offset |
20 ms (excellent) |
50 ms |
100 ms |
1.1 Resolution and FPS Configuration#
Source: /services/avatar-engine/src/avatar_engine/config.py
| Quality Preset |
Resolution |
Target FPS |
Use Case |
| LOW |
512x512 |
60 |
Mobile, low bandwidth |
| MEDIUM |
768x768 |
45 |
Standard video calls |
| HIGH (default) |
1024x1024 |
30 |
HD conferencing |
| ULTRA |
1024x1024 |
Max quality |
Recordings, presentations |
Configurable Range:
- Width/Height: 256 - 2048 pixels
- FPS: 15 - 120 frames per second
1.2 3D Gaussian Splatting (3DGS) Parameters#
| Parameter |
Default |
Range |
Impact |
| Max Gaussians |
100,000 |
10K - 500K |
Quality vs VRAM |
| SH Degree |
3 |
0 - 4 |
View-dependent effects |
| Tile Size |
16 px |
8 - 32 |
Rasterization granularity |
| Background |
White |
RGB |
Compositing |
1.3 Rendering Features Configuration#
Temporal Anti-Aliasing (TAA)#
- Enabled: Default
- History Weight: 0.9 (temporal stability)
- Jitter: 0.5 pixels
- Alpha Threshold: 1e-3
- History Clamping: Enabled
Dynamic Resolution Scaling#
- Enabled: Optional (disabled by default)
- Min Scale: 0.5x
- Max Scale: 1.0x
- Scale Step: 0.05
- EMA Alpha: 0.2
- Cooldown: 5 frames
Level of Detail (LOD)#
- Enabled: Default
- Min Pixel Radius: 0.0
- Max Gaussians Cap: Unlimited (0)
Culling Optimizations#
- Depth Sorting: Enabled
- Frustum Culling: Enabled
1.4 GPU Memory Estimates by Resolution#
Based on 3DGS implementation with 100K Gaussians:
| Resolution |
Estimated VRAM |
Recommended GPU |
| 512x512 |
~2-3 GB |
RTX 3060 (8 GB) |
| 768x768 |
~3-4 GB |
RTX 3070 (8 GB) |
| 1024x1024 |
~4-6 GB |
RTX 3080 (10 GB) |
| 1536x1536 |
~6-8 GB |
RTX 4080 (16 GB) |
| 2048x2048 |
~8-12 GB |
RTX 4090 (24 GB) / A100 |
Note: Actual VRAM usage depends on:
- Number of active Gaussians
- TAA history buffers
- Additional feature buffers (AO, etc.)
- Model weights in memory
1.5 Cold Start Time Components#
| Component |
Estimated Time |
Notes |
| PyTorch/CUDA Init |
2-5 seconds |
First CUDA call |
| Model Loading |
1-3 seconds |
3DGS weights |
| FLAME Model Init |
0.5-1 second |
Face model |
| Preprocessing Pipeline |
0.5-1 second |
MediaPipe init |
| First Frame Render |
100-500 ms |
JIT compilation |
| Total Cold Start |
5-10 seconds |
First request |
1.6 Expression Transition Latency#
Face Detection Configuration:
- Detector: MediaPipe (default)
- Min Face Size: 128 pixels
- Detection Confidence: 0.9
- Target Processing FPS: 30
Expression Pipeline:
- Landmark Detection: ~10-20 ms
- FLAME Parameter Extraction: ~5-10 ms
- Gaussian Update: ~5-10 ms
- Render: ~16-33 ms (30-60 FPS)
- Total Expression Transition: ~35-70 ms
Based on benchmarking infrastructure in /scripts/benchmarks/gpu_benchmark.py:
| GPU Tier |
Example |
Expected FPS @ 1024x1024 |
VRAM |
Notes |
| Entry |
RTX 3060 |
20-25 FPS |
8 GB |
LOW/MEDIUM quality |
| Mid |
RTX 3080 |
30-40 FPS |
10 GB |
HIGH quality |
| High |
RTX 4090 |
50-60 FPS |
24 GB |
ULTRA quality |
| Datacenter |
A100 |
60+ FPS |
40/80 GB |
Multiple streams |
| Datacenter |
H100 |
80+ FPS |
80 GB |
Maximum throughput |
1.8 Quality Validation Thresholds#
Source: /services/avatar-engine/src/avatar_engine/utils/metrics.py
| Metric |
Minimum Acceptable |
Target |
Excellent |
| PSNR |
25.0 dB |
30.0 dB |
35.0+ dB |
| SSIM |
0.85 |
0.90 |
0.95+ |
| LPIPS |
0.15 (max) |
0.10 |
0.05 |
| Identity Similarity |
0.70 |
0.85 |
0.95+ |
2. End-to-End Response Latency (Task 24.1.3.2)#
2.1 Voice Pipeline Latency Budget#
Source: /packages/common/src/common/metrics/performance.py
| Component |
Target |
Warning (1.5x) |
Critical (2x) |
| Audio Capture |
20 ms |
30 ms |
40 ms |
| Audio Preprocessing |
10 ms |
15 ms |
20 ms |
| STT Inference |
100 ms |
150 ms |
200 ms |
| Intent Detection |
20 ms |
30 ms |
40 ms |
| Entity Extraction |
20 ms |
30 ms |
40 ms |
| Context Retrieval |
30 ms |
45 ms |
60 ms |
| RAG Retrieval |
50 ms |
75 ms |
100 ms |
| LLM Prompt Prep |
10 ms |
15 ms |
20 ms |
| LLM Inference |
200 ms |
300 ms |
400 ms |
| TTS Synthesis |
100 ms |
150 ms |
200 ms |
| Audio Postprocessing |
10 ms |
15 ms |
20 ms |
| Lipsync Generation |
20 ms |
30 ms |
40 ms |
| Avatar Rendering |
33 ms |
50 ms |
66 ms |
| Network Egress |
20 ms |
30 ms |
40 ms |
| End-to-End Total |
500 ms |
750 ms |
1000 ms |
2.2 LLM Provider Latency Expectations#
| Provider |
Model |
Expected Latency |
Notes |
| Anthropic |
Claude 3.5 Sonnet |
150-300 ms |
Streaming first token |
| Anthropic |
Claude 3 Opus |
300-500 ms |
More reasoning |
| OpenAI |
GPT-4 Turbo |
200-400 ms |
Streaming |
| OpenAI |
GPT-4o |
150-250 ms |
Optimized |
| Local |
Llama 3.1 70B |
100-200 ms |
A100 required |
Timeout Configuration:
- LLM Timeout: 60 seconds (env:
LLM_TIMEOUT_SECONDS)
- RAG Timeout: 5 seconds (env:
RAG_TIMEOUT_SECONDS)
- Long-term Memory Timeout: 5 seconds (env:
LONG_TERM_MEMORY_TIMEOUT_SECONDS)
2.3 Memory Retrieval Latency#
| Memory Type |
Expected Latency |
Storage |
| In-Context (Working) |
1-5 ms |
Redis |
| Short-term (Session) |
5-15 ms |
PostgreSQL |
| Long-term (Archival) |
20-50 ms |
Qdrant (vector) |
| Knowledge Base (RAG) |
50-100 ms |
Qdrant + reranking |
Database Performance Thresholds:
- PostgreSQL P95: 50 ms (regression threshold)
- Redis P95: 10 ms (regression threshold)
2.4 Audio Quality Monitoring#
Source:
/services/voice-engine/src/voice_engine/audio/quality_monitoring.py
| Metric |
Acceptable |
Poor |
Critical |
| Clipping Ratio |
< 0.1% |
< 1% |
< 5% |
| DC Offset |
< 0.02 |
< 0.05 |
< 0.1 |
| RMS Level |
> -45 dBFS |
> -55 dBFS |
- |
| SNR |
> 10 dB |
> 5 dB |
- |
| Dropout Ratio |
< 5% |
< 10% |
- |
2.5 Voice-to-Voice Response Time Breakdown#
Target: < 500 ms total
text
┌─────────────────────────────────────────────────────────────────────┐
│ Voice-to-Voice Pipeline │
├─────────────────────────────────────────────────────────────────────┤
│ │
│ Audio In ──► STT ──► NLU ──► LLM ──► TTS ──► Audio Out │
│ 20ms 100ms 40ms 200ms 100ms 20ms │
│ │
│ ├───────────────────── 480ms total ─────────────────────────────┤ │
│ │
│ + Avatar render (parallel): 33ms │
│ + Network latency: ~20ms │
│ │
│ Total with avatar: ~500ms │
│ │
└─────────────────────────────────────────────────────────────────────┘
2.6 Latency Monitoring Configuration#
Histogram Buckets (ms): [5, 10, 25, 50, 100, 250, 500, 1000, 2500, 5000,
10000]
Percentiles Tracked: P50, P75, P90, P95, P99
Aggregation Window: 60 seconds
Sample Rate: 1.0 (all requests)
3. Conferencing Integration Latency (Task 24.1.3.3)#
3.1 A/V Synchronization Targets#
Source: /services/voice-engine/src/voice_engine/sync/
| Quality Level |
Offset Threshold |
User Experience |
| Excellent |
≤ 20 ms |
Imperceptible |
| Good |
≤ 50 ms |
Acceptable |
| Acceptable |
≤ 100 ms |
Noticeable but tolerable |
| Poor |
> 100 ms |
Distracting |
| Critical |
> 200 ms |
Unusable |
3.2 Sync Measurement Configuration#
| Parameter |
Value |
Description |
| Measurement Window |
2.0 seconds |
Rolling window for sync calc |
| Sample Interval |
0.05 seconds |
20 samples/second |
| Outlier Threshold |
3.0 σ |
Statistical outlier removal |
| Min Valid Samples |
5 |
Minimum for valid measurement |
| Smoothing Factor |
0.3 |
Exponential moving average |
3.3 Sync Compensation Configuration#
| Parameter |
Value |
Description |
| Strategy |
GRADUAL |
Smooth corrections over time |
| Target |
VISUAL |
Adjust video timing |
| Min Correction Trigger |
10 ms |
Minimum offset to correct |
| Max Single Correction |
500 ms |
Maximum adjustment per step |
| Correction Rate |
0.1 |
Fraction corrected per update |
| Smooth Correction Time |
200 ms |
Duration to apply correction |
| Drift Detection Window |
5.0 seconds |
Window for drift calc |
| Drift Threshold |
5.0 ms/sec |
Max acceptable drift rate |
| Max Total Adjustment |
2000 ms |
Maximum cumulative adjustment |
| Platform |
Video Injection |
Audio Sync |
Notes |
| Zoom |
50-100 ms |
±20 ms |
Virtual camera API |
| Google Meet |
80-150 ms |
±30 ms |
Chrome extension |
| Microsoft Teams |
60-120 ms |
±25 ms |
NDI/Virtual camera |
| Webex |
70-130 ms |
±25 ms |
Virtual camera |
3.5 Frame Rate and Drop Monitoring#
Source: /packages/common/src/common/metrics/performance.py
| Metric |
Target |
Threshold |
| Video Capture FPS |
30 |
Variance ≤ 10% |
| Video Output FPS |
30 |
Variance ≤ 10% |
| Avatar Render FPS |
30 |
Variance ≤ 10% |
| Display FPS |
60 |
Variance ≤ 10% |
Frame Drop Detection:
- Dropped Frame Threshold: 50 ms (frame exceeds expected interval by 50ms)
- Measurement Window: 1.0 second
- Max Frame History: 1,000 frames
3.6 Audio Quality Score (MOS Estimation)#
Mean Opinion Score Categories:
| MOS Score |
Quality |
Description |
| 4.3 - 5.0 |
Excellent |
Indistinguishable from original |
| 3.6 - 4.3 |
Good |
Perceptible but not annoying |
| 2.6 - 3.6 |
Fair |
Slightly annoying |
| 1.6 - 2.6 |
Poor |
Annoying |
| 1.0 - 1.6 |
Bad |
Very annoying |
MOS Computation Weights:
| Factor |
Weight |
Excellent Value |
| Codec Quality |
15% |
Opus: 4.5 |
| SNR |
25% |
≥ 35 dB |
| Packet Loss |
25% |
0% |
| Jitter |
15% |
≤ 20 ms |
| Bitrate |
10% |
Sufficient |
| Speech Clarity |
10% |
High |
3.7 Network Requirements#
| Quality |
Min Bandwidth (Up) |
Min Bandwidth (Down) |
Latency |
| Audio Only |
64 kbps |
64 kbps |
< 150 ms |
| SD Video (480p) |
1 Mbps |
1 Mbps |
< 100 ms |
| HD Video (720p) |
2.5 Mbps |
2.5 Mbps |
< 80 ms |
| Full HD (1080p) |
4 Mbps |
4 Mbps |
< 60 ms |
4. GPU Benchmarking Infrastructure#
Source: /scripts/benchmarks/gpu_benchmark.py
Quick Benchmark Preset#
python
warmup_iterations = 10
measured_iterations = 50
matmul_sizes = [1024, 2048]
conv2d_shapes = [(8, 64, 128, 56, 56, 3)]
bandwidth_test_size = 256 MB
Standard Benchmark Preset#
python
warmup_iterations = 20
measured_iterations = 200
matmul_sizes = [1024, 2048, 4096]
conv2d_shapes = [
(16, 64, 128, 112, 112, 3),
(8, 128, 256, 56, 56, 3)
]
bandwidth_test_size = 1024 MB
Stress Benchmark Preset#
python
warmup_iterations = 30
measured_iterations = 300
matmul_sizes = [2048, 4096, 8192]
conv2d_shapes = [
(16, 64, 128, 224, 224, 3),
(8, 128, 256, 112, 112, 3)
]
bandwidth_test_size = 2048 MB
4.2 Supported Precision Modes#
| Mode |
Description |
VRAM Usage |
Performance |
| FP32 |
Full precision |
100% |
Baseline |
| FP16 |
Half precision |
50% |
2x faster (Tensor Cores) |
| BF16 |
Brain float |
50% |
2x faster (better range) |
4.3 Benchmark Metrics#
- TFLOPS: Tera Floating Point Operations Per Second
- GBPS: Giga Bytes Per Second (memory bandwidth)
- Latency: Average operation time in milliseconds
5. Monitoring Infrastructure#
5.1 Prometheus Scrape Configuration#
Source: /infrastructure/docker/prometheus/prometheus.yml
| Service |
Port |
Scrape Interval |
| API Gateway |
8000 |
10 seconds |
| Avatar Engine |
8001 |
10 seconds |
| Voice Engine |
8002 |
10 seconds |
| Behavior Engine |
8003 |
10 seconds |
| Perception Engine |
8004 |
10 seconds |
| Computer Use |
8005 |
10 seconds |
| Conferencing |
8006 |
10 seconds |
| Orchestrator |
8007 |
10 seconds |
| Knowledge Base |
8008 |
10 seconds |
| Persona Service |
8009 |
10 seconds |
5.2 GPU Metrics Collection#
| Source |
Port |
Interval |
Metrics |
| NVIDIA DCGM |
9400 |
5 seconds |
GPU utilization, memory, temp |
| Triton Server |
8002 |
5 seconds |
Inference latency, throughput |
| Per-Service GPU |
Various |
5 seconds |
Service-specific GPU usage |
5.3 Infrastructure Metrics#
| Service |
Port |
Interval |
| Redis |
6379 |
Default |
| Qdrant |
6333 |
30 seconds |
| MinIO |
9000 |
30 seconds |
Global Settings:
- Default Scrape Interval: 15 seconds
- Evaluation Interval: 15 seconds
Source: /tests/utils/performance.py
Features:
- Context manager timing
- Function/async decorators
- Memory tracking (tracemalloc)
- Benchmark runner with warmup
- Statistical analysis (mean, std, percentiles)
- Markdown report export
6.2 Regression Testing#
Source: /scripts/performance/perf_regression.py
Configuration:
- Warmup iterations: 25
- Measured iterations: 100
- Percentiles: P50, P95
Thresholds:
- PostgreSQL P95: 50 ms
- Redis P95: 10 ms
7.1 Avatar Engine Optimizations#
| Optimization |
Impact |
Implementation |
| Dynamic Resolution |
20-50% FPS gain |
Enable when FPS drops |
| LOD Adjustment |
10-30% VRAM savings |
Distance-based culling |
| TAA Disable |
10-15% FPS gain |
Quality tradeoff |
| Gaussian Culling |
20-40% speedup |
Frustum + occlusion |
| Mixed Precision |
30-50% speedup |
FP16 on Tensor Cores |
7.2 Voice Pipeline Optimizations#
| Optimization |
Impact |
Implementation |
| Streaming STT |
100-200ms reduction |
Real-time transcription |
| LLM Streaming |
200-400ms reduction |
First token latency |
| Parallel Processing |
50-100ms reduction |
Concurrent pipelines |
| Response Caching |
50-90% reduction |
Common queries |
| Model Quantization |
20-40% speedup |
INT8 inference |
7.3 Conferencing Optimizations#
| Optimization |
Impact |
Implementation |
| Frame Buffering |
Smoother playback |
2-3 frame buffer |
| Adaptive Bitrate |
Network resilience |
Quality scaling |
| Jitter Buffer |
Audio quality |
Adaptive sizing |
| Predictive Sync |
Better A/V sync |
Drift compensation |
8. Baseline Measurement Procedures#
8.1 Avatar Rendering Benchmark#
bash
# Run GPU benchmark
cd /home/ubuntu/oshun/serwaa
python scripts/benchmarks/gpu_benchmark.py --preset standard --dtype fp16
# Expected output:
# - MatMul TFLOPS
# - Conv2D TFLOPS
# - Memory Bandwidth GBPS
# - Per-resolution FPS estimates
8.2 End-to-End Latency Measurement#
bash
# Run latency tests
pytest tests/performance/ -v --benchmark-enable
# Metrics collected:
# - P50, P95, P99 latencies
# - Component breakdown
# - Memory usage
8.3 A/V Sync Measurement#
bash
# Run sync tests
pytest tests/voice_engine/sync/ -v
# Metrics collected:
# - Sync offset distribution
# - Drift rate
# - Correction effectiveness
9.1 Availability SLOs#
| Service |
Availability |
Max Downtime/Month |
| Avatar Engine |
99.9% |
43.8 minutes |
| Voice Engine |
99.9% |
43.8 minutes |
| API Gateway |
99.95% |
21.9 minutes |
| Overall Platform |
99.5% |
3.65 hours |
9.2 Latency SLOs#
| Metric |
P50 |
P95 |
P99 |
| API Response |
100 ms |
300 ms |
500 ms |
| Avatar Render |
33 ms |
50 ms |
66 ms |
| Voice Response |
500 ms |
750 ms |
1000 ms |
| A/V Sync |
20 ms |
50 ms |
100 ms |
9.3 Quality SLOs#
| Metric |
Target |
Minimum |
| Avatar SSIM |
0.90 |
0.85 |
| Audio MOS |
4.0 |
3.6 |
| Frame Drop Rate |
< 1% |
< 5% |
| Sync Quality |
Excellent |
Good |
10. Summary Tables#
| Metric |
LOW |
MEDIUM |
HIGH |
ULTRA |
| Resolution |
512x512 |
768x768 |
1024x1024 |
1024x1024 |
| Target FPS |
60 |
45 |
30 |
Max Quality |
| Est. VRAM |
2-3 GB |
3-4 GB |
4-6 GB |
6-8 GB |
| Min GPU |
RTX 3060 |
RTX 3070 |
RTX 3080 |
RTX 4090 |
10.2 Latency Budget Summary#
| Component |
Budget |
Cumulative |
| Audio Capture |
20 ms |
20 ms |
| STT |
100 ms |
120 ms |
| NLU |
40 ms |
160 ms |
| LLM |
200 ms |
360 ms |
| TTS |
100 ms |
460 ms |
| Network |
40 ms |
500 ms |
10.3 Quality Thresholds Summary#
| Metric |
Minimum |
Target |
Excellent |
| PSNR |
25 dB |
30 dB |
35+ dB |
| SSIM |
0.85 |
0.90 |
0.95+ |
| MOS |
3.6 |
4.0 |
4.3+ |
| A/V Sync |
100 ms |
50 ms |
20 ms |
Appendix A: Key Configuration Files#
| File |
Purpose |
/services/avatar-engine/src/avatar_engine/config.py |
Avatar rendering config |
/services/avatar-engine/src/avatar_engine/utils/metrics.py |
Quality metrics |
/services/voice-engine/src/voice_engine/audio/quality_monitoring.py |
Audio monitoring |
/services/voice-engine/src/voice_engine/sync/measurement.py |
Sync measurement |
/services/voice-engine/src/voice_engine/sync/compensation.py |
Sync compensation |
/packages/common/src/common/metrics/performance.py |
Latency tracking |
/scripts/benchmarks/gpu_benchmark.py |
GPU benchmarking |
/scripts/performance/perf_regression.py |
Regression testing |
/tests/utils/performance.py |
Test utilities |
/infrastructure/docker/prometheus/prometheus.yml |
Monitoring config |
Appendix B: Environment Variables#
| Variable |
Default |
Description |
LLM_TIMEOUT_SECONDS |
60 |
LLM request timeout |
RAG_TIMEOUT_SECONDS |
5 |
RAG retrieval timeout |
LONG_TERM_MEMORY_TIMEOUT_SECONDS |
5 |
Memory access timeout |
AVATAR_QUALITY |
HIGH |
Default quality preset |
TARGET_FPS |
30 |
Target frame rate |
CUDA_VISIBLE_DEVICES |
0 |
GPU selection |
| Field |
Value |
| Created |
January 2026 |
| Last Updated |
January 2026 |
| Status |
Complete |
| Owner |
Psyche Domain Migration Team |
| Review Cycle |
Monthly |
| Related Tasks |
24.1.3.1, 24.1.3.2, 24.1.3.3 |