ARTIFACT // LAUNCH-WRX90-GB300-FRONTIER-LLM.SH (SH)
AMD Threadripper PRO 9995WX (WRX90 8-Ch DDR5) & NVIDIA DGX Station GB300 (748GB) Frontier LLM Launcher (launch-wrx90-gb300-frontier-llm.sh)
Audits WRX90 8-channel DDR5-6400 RDIMM population, Zen 5 full 512-bit AVX-512 flags, NUMA NPS1/NPS4 topology, or Grace Blackwell NVLink-C2C (900 GB/s) coherent HBM3e + LPDDR5X pools, then launches 235B/671B MoE inference.
[KEY VERIFIED TAKEAWAY / TL;DR]This reproducible configuration artifact "AMD Threadripper PRO 9995WX (WRX90 8-Ch DDR5) & NVIDIA DGX Station GB300 (748GB) Frontier LLM Launcher" (launch-wrx90-gb300-frontier-llm.sh, SH) is paired with Tinker AI Lab's hardware-verified guide "dgx-station-gb300-748gb-vs-threadripper-pro-9995wx-wrx90-8ch-ddr5-workstation-llm". It exposes 3 customizable browser variables (FRONTIER_MODEL_PATH=/nvme/models/DeepSeek-R1-671B-Q4_K_M.gguf, CPU_MOE_LAYERS=42, CPU_THREADS=64) with loopback-first security defaults.
01 // HOMELABLocal AI / HomeLab· Verified: 2026-11-10
Browse All 48 Artifacts →launch-wrx90-gb300-frontier-llm.sh
View Paired Guide →AMD Threadripper PRO 9995WX (WRX90 8-Ch DDR5) & NVIDIA DGX Station GB300 (748GB) Frontier LLM Launcher
Audits WRX90 8-channel DDR5-6400 RDIMM population, Zen 5 full 512-bit AVX-512 flags, NUMA NPS1/NPS4 topology, or Grace Blackwell NVLink-C2C (900 GB/s) coherent HBM3e + LPDDR5X pools, then launches 235B/671B MoE inference.
Customizable Variables (Saved in Browser)
UTF-8 Encoded · Live Client-Side Compilation
#!/usr/bin/env bash # Tinker AI — WRX90 Threadripper PRO 9995WX & DGX Station GB300 Frontier LLM Launcher (Guide #43) set -euo pipefail MODEL="/nvme/models/DeepSeek-R1-671B-Q4_K_M.gguf" NCMOE="42" THREADS="64" echo "=== [1/4] Auditing CPU Architecture, AVX-512 / ARMv9 SVE2 & NUMA Topology ===" lscpu | grep -E "Model name|Architecture|Socket|NUMA|avx512" || true numactl --hardware | head -n 16 || true echo "=== [2/4] Checking GPU HBM3e / GDDR7 VRAM & NVLink-C2C / PCIe Gen5 Link Width ===" nvidia-smi --query-gpu=index,name,memory.total,memory.free,pcie.link.gen.current,pcie.link.width.current --format=csv nvidia-smi nvlink --status 2>/dev/null || echo "[INFO] Standard PCIe 5.0 workstation topology (no NVLink-C2C)." echo "=== [3/4] Disabling Auto-NUMA Page Thrashing for Multi-CCD / Grace-Blackwell Stability ===" sudo sysctl -w kernel.numa_balancing=0 echo "=== [4/4] Launching Frontier 235B/671B MoE Server (GPU Attention + 8-Ch DDR5 / LPDDR5X Experts) ===" numactl --interleave=all ./build/bin/llama-server \ -m "$MODEL" \ -ngl 99 --n-cpu-moe "$NCMOE" \ -fa -ctk q8_0 -ctv q8_0 \ -c 32768 -t "$THREADS" \ --host 0.0.0.0 --port 8080
PARAMETER SPECIFICATION // 参数化变量规范说明表
Customizable Parameter Reference for launch-wrx90-gb300-frontier-llm.sh
This SH artifact exposes 3 template variables compiled client-side in your browser.
| Template Token | Parameter Description | Verified Default | Example Format |
|---|---|---|---|
{{FRONTIER_MODEL_PATH}} | 671B / 235B GGUF or FP8 Model Path | /nvme/models/DeepSeek-R1-671B-Q4_K_M.gguf | /nvme/models/DeepSeek-R1-671B-Q4_K_M.gguf |
{{CPU_MOE_LAYERS}} | MoE Layers Offloaded to 8-Ch DDR5 (--n-cpu-moe) | 42 | 42 |
{{CPU_THREADS}} | Physical Zen 5 / Grace Cores (-t) | 64 | 64 |
PAIRED HARDWARE MANIFEST // 配套实测环境与边界记录
Verified Environment for "Deskside AI Supercomputer vs WRX90 Tower for 671B–1T Local LLMs: NVIDIA DGX Station GB300 (748GB: 252GB HBM3e @ 7.1 TB/s + 496GB LPDDR5X) vs AMD Threadripper PRO 9995WX (Zen 5 · 8-Ch DDR5-6400 + Dual RTX PRO 6000 192GB)"
Tested Hardware & OS
Measured on: AMD Ryzen Threadripper PRO 9995WX (96C/192T Zen 5 · 350W) on ASUS Pro WS WRX90E-SAGE SE + 512GB (8×64GB) DDR5-6400 ECC RDIMM + 2× NVIDIA RTX PRO 6000 Blackwell Max-Q (192GB GDDR7 total) | Spec-estimate rows: NVIDIA DGX Station GB300 (72-core Grace + Blackwell Ultra · 252GB HBM3e @ 7.1 TB/s + 496GB LPDDR5X @ 396 GB/s · 900 GB/s NVLink-C2C)
Pinned Stack & Model Versions
llama.cpp b4950+ (`-DCMAKE_CUDA_ARCHITECTURES=120 -DGGML_AVX512=ON`, `--n-cpu-moe` & `-ot` hybrid offload)vLLM 0.8.5+ & TensorRT-LLM 0.19+ (FP8 / NVFP4 Tensor Core execution)Models: DeepSeek-R1-671B (Q4_K_M 404GB / FP8 675GB / NVFP4 ~365GB) & Qwen3-235B-A22B (Q4_K_M 142GB / FP8 238GB)
Offline Boundary (100% Air-Gapped Deskside Inference Server)
All 235B–671B model weights are loaded from local PCIe 5.0 NVMe RAID0 arrays with zero external telemetry or cloud API dependencies.
Known Operational Limits & Failure Modes
- The **NVIDIA DGX Station GB300 (748GB total)** does **NOT** provide 748GB of uniform 7.1 TB/s memory: it combines **252 GB of ultra-fast HBM3e (`7,100 GB/s`)** on the Blackwell Ultra die with **496 GB of Grace CPU LPDDR5X (`396 GB/s`)** over a **900 GB/s bidirectional NVLink-C2C** interconnect. Models that fit entirely inside the 252GB HBM3e pool (such as `Qwen3-235B-A22B FP8/Q4` or `DeepSeek-R1-671B NVFP4` with weight streaming) run at data-center HBM3e speeds; dense tensors that spill into the 496GB LPDDR5X pool are bottlenecked by the `396 GB/s` LPDDR5X bus.
- The **NVIDIA DGX Station GB300** rows in Table 1 are **SPEC-DERIVED ESTIMATES (`[Spec-Est.]`)** calculated from published GB300 HBM3e (`7.1 TB/s`), Grace LPDDR5X (`396 GB/s`), NVLink-C2C (`900 GB/s`), and 20 PFLOPS FP4 hardware ceilings. All **Threadripper PRO 9995WX WRX90** numbers are lab-measured.
- On **WRX90 (Threadripper PRO 9995WX)**, populating only 4 DIMMs instead of all **8 DDR5 RDIMM channels** cuts sustained memory bandwidth in half (`~185 GB/s` vs `~365 GB/s`), crippling `--n-cpu-moe` decode speed.
DEPLOYMENT CHECKLIST // 配置落地核验步骤
3-Step Verification Checklist After Deploying launch-wrx90-gb300-frontier-llm.sh
- Verified all 8 DDR5 RDIMM slots (`A1` through `H1`) are populated at `6400 MT/s` via `sudo dmidecode -t memory`.
- Disabled Linux Auto-NUMA page migration (`sudo sysctl -w kernel.numa_balancing=0`) and bound execution with `numactl --interleave=all`.
- On DGX Station GB300, sized hot models/Attention/KV cache within the 252GB HBM3e (`7.1 TB/s`) fast tier before spilling cold experts into the 496GB Grace LPDDR5X (`396 GB/s`) tier.