Scalable AI Server
summary
Rack-scale ARM server built from SynQuacer boards plus dedicated CNN accelerators claims GPU-class inference at far lower power: a 32‑board chassis with 16 AI modules reaches ~352 TOPS vs DGX1 numbers while drawing ~701 W vs 3201 W for ~5x efficiency. Fabric is SynQuacer DDT PCIe fabric with one‑hop ~1.5 μs logical latency and CPU‑to‑CPU 30–36 μs (DDT‑R 78 μs) versus millisecond Ethernet. Accelerator is Gyrfalcon 5.6 TOPS at 0.8 W, achieves ~54–55 FPS on CNNs (target 60–75), C++ SDK ready, Python wrapper pending.
AI-written summary · about
chapters (12)
- 0:00Introduction to Scalable AI
- 1:10AI inference data placement
- 2:40Demo of CNN inference process
- 5:57Distributed AI system design
- 8:06AI module hardware overview
- 10:32Direct Data Transactions explained
- 13:10Server room setup and debugging
- 15:22CPU performance and security
- 17:47Storage scaling with DDT
- 19:16AI chip efficiency comparison
- 21:57Performance benchmarks
- 23:02SDK and future release
related talks
33mAI and SDR: Software Meets Hardware Again...
40mFueling the AI Revolution with Gaming
20mLow Power AI Edge Hardware
55mLarge Scale Distributed AI on the JVM
50mAI acceleration at the edge
22mResearch SuperCluster(RSC): Accelerating AI Research at Scale
23mThe Race to the Bottom: Low Latency in the age of the Transformer
04m4 real-world AI deployments at enterprise scale