About AI Tech Engine
We do not copy ad copy or vendor nameplate scores as if they were our results. Reports cover what we have checked on the machines sitting on the desk. tok/s stays blank until we time the same prompt ourselves. The Gigabyte DGX Spark 128GB arrived on 2026-08-31 and is on site. We still have no measured tok/s. The AMD Strix Halo 128GB is still planned.
1. Mission
Local-LLM writeups often treat VRAM and unified memory as the same number. In a real lab, a 16GB CUDA card and 64/96GB Apple UMA bottleneck differently. AI Tech Engine writes that difference against our own machines.
Frontier API pricing and cloud rentals are linked to official product pages only. We do not have affiliate IDs yet, and we do not invent ref codes.
2. Lab hardware (as of August 2026)
| Machine | Memory | Role | Status |
|---|---|---|---|
| MacBook (M1 Max) | 64GB unified memory | Travel, long-context drafts, MLX / llama.cpp | On site |
| Mac Studio (M2 Max) | 96GB unified memory | Always-on local server, 27–32B Q4, 70B Q4 when it fits | On site |
| CUDA workstation (RTX 4080) | 16GB GDDR6X | Daily 7–8B / 14B; 27B Q4 is tight | On site |
| Gigabyte DGX Spark | 128GB unified LPDDR5x | 70B-class, CUDA agent stack | On site (arrived 2026-08-31; no measured tok/s yet) |
| AMD Ryzen AI Max+ 395 (Strix Halo) | 128GB LPDDR5x | 70B-class Q4, local agents, ROCm control | Planned (not arrived) |
This lab does not have an RTX 4090 24GB, an M3 Max 128GB, or an H100 cluster. Older intro copy that listed those specs has been removed.
3. Methodology (no speed table yet)
- What we publish now: whether weights fit in memory, licenses, quantized file sizes, and which path hits the wall first when vision or long context is opened
- What we will measure next: the same GGUF/MLX on four machines (five later), load vs OOM, open context, and wall-clock tool loops. Spark tok/s is still blank
- What we still will not do: quote vendor tok/s as ours, or treat Spark’s 200B nameplate / the unarrived Strix Halo as lab capacity
4. Affiliate transparency
The recommendation cards currently use official product URLs for RunPod, Vast.ai, and Cursor only. No tracking IDs have been issued, and we do not invent fake ref codes. When real affiliate URLs exist they will be swapped in only via PUBLIC_AFFILIATE_RUNPOD_URL, PUBLIC_AFFILIATE_VAST_URL, and PUBLIC_AFFILIATE_CURSOR_URL.
Cloud GPUs & engineering tools
Official sites (affiliate IDs pending)Affiliate tracking IDs have not been issued yet. The links below are official product sites. We do not invent fake ref/click IDs.
Hourly A100 / H100 / RTX 4090 rentals with one-click vLLM templates
Distributed GPU marketplace for local-LLM fine-tuning rentals
AI autocomplete and refactoring IDE