The bench numbers from Parts 1 through 6 said the card could do this. The four-day telemetry says it does. One Radeon AI PRO R9700, one Qwen3-30B-A3B-AWQ model, ten background agents, four ingestors, three frontends. No retries. No fallbacks. No restarts. Here is the production verdict and what it means for sizing a agentic AI node.
If you only remember four numbers from this nine-part series, remember these. They define what a 32 GB RDNA4 card can do as an always-on agentic AI host.
368 ms mean time-to-first-token over 49,294 samples. This is the number that defines whether the agents feel snappy. It held flat across four days of continuous load — no degradation as VRAM filled with KV cache, no tail spikes that would block the verify loop.
Synthesis dominates volume. Verification keeps the brain honest. User-facing Q&A is a minority of requests but the most latency-sensitive. vLLM continuous batching flattens all three into one rolling forward pass.
The full agentic stack — including the model serving, all ten background agents, all four ingestors, all three frontends, the brain databases, and the perf dashboards — fits on one EPYC socket and one R9700 with room to spare.
| Resource | Used | Available | Utilization |
|---|---|---|---|
| GPU compute (R9700) | ~100% during batches | 1 card | Saturated by design |
| GPU memory | 29.7 GB | 32 GB | 93% |
| GPU power | 92-143 W | 300 W TBP | 31-48% |
| GPU temp (junction) | 77 C | 110 C limit | 70% |
| Host CPU (p50) | 0.3% | 96 cores / 12 CCDs | <1% |
| Host CPU (p95) | 12.9% | 96 cores / 12 CCDs | ~13% |
| Host memory | 24.5 GB peak | 755 GB | 3.2% |
Based on four days of real telemetry. Not a spec sheet. Not a projection.
| Component | Minimum that worked | Why |
|---|---|---|
| GPU | Radeon AI PRO R9700 (32 GB) | Holds Qwen3-30B-A3B-AWQ at 90% mem-util with 32k context |
| GPU memory | 32 GB minimum | Below this, drop to dense 14B (Part 1) or to 8B class |
| CPU | 16-24 cores (2-3 CCDs) | p95 measured at <13% of 96 cores. 2 CCDs covers normal load; a 3rd absorbs speech-preprocessing bursts cleanly. |
| Host memory | 64 GB | Working set is 24 GB peak. 64 GB covers SQLite WAL + page cache + the speech preprocessor buffers comfortably. |
| Storage | 1 TB NVMe | the retrieval store + the synthesis store + meetings.db + agent_history.db = single-digit GB. The rest is for ingested files. |
| OS | Ubuntu 24.04 + ROCm 7.2.1 | The exact stack that produced these numbers |
Each lesson cost engineering time. None of them came from a spec sheet.
Nine parts, three runtimes, six benchmarks, and a four-day production run. The thesis was that a single Radeon AI PRO R9700 could host an entire agentic AI stack for a working AMD Field team. The thesis held.
49,288 requests. 35.24 million prompt tokens. 8.73 million generated. Zero errors. One model. One GPU. One box. The card that started this series as a bench target now runs my actual job.