01
Launched Oct 16 across ASUS, Dell, HP, Lenovo, Microsoft, and MSI: the 20-core Grace ARM CPU (MediaTek co-design) + Blackwell RTX GPU (6,144 CUDA, FP4) with 128GB unified LPDDR5X, ~1 PFLOP of FP4 AI compute, 600 GB/s NVLink-C2C — enough, NVIDIA says, to run 120B-parameter models with 1M context locally. First-wave laptop benchmarks show strong local inference but an Arm/Prism emulation tax on CPU-bound apps and ~20–25% battery throttling under sustained AI load. The standout form factor: ASUS’s ProArt GR1X mini PC, 140W TDP, up to 128GB, fall availability in select regions. NVIDIA is trying to turn the home AI agent into an appliance category — $2,599 for 128GB of unified memory speaks his language (local inference, KV-cache headroom). The x86 box isn’t obsolete; it just got company.
Source: Forkast (launch coverage, specs, first-wave benchmarks).
02
Updated Oct 17: PS5LM (a llama.cpp port, GPU path via the PS5_Vulkan/RADV driver port) turns a jailbroken PS5 into a GGUF server with 11.4GB of usable memory — a Qwen-class 27B in aggressive quantization does ~20 tok/s, Gemma 4 E4B ~61 tok/s, Llama 3.2 3B ~124 tok/s, all at ~86W and 54°C. Ships a TV dashboard (tokens/s, power, thermals, model/KV memory split) plus OpenAI/Anthropic-compatible endpoints on the LAN. Requires firmware 7.00–13.60 and voids the warranty. A jailbroken console now benchmarks at roughly one A770’s worth of local inference for the price of a used PS5 — the ultimate argument that spare gaming silicon is a legitimate inference node.
Source: AlexTech (project rundown, benchmarks, requirements).
03
First leaked die images (via Moore’s Law Is Dead): Nova Lake-HX is 8 P-cores + 16 E-cores + 4 LPE-cores across six tiles on a 360mm² package — the 104mm² compute die on TSMC N2P, SoC and peripheral dies on Intel 18A-P, plus a 4-Xe-core graphics tile on Intel 3-E. Mainstream Nova Lake-H drops to 4+8+4 in a 276mm² package. Intel’s halo laptop chip is a patchwork sewn across two foundries — interesting silicon, but for the 9800X3D box the takeaway is that Intel is still chasing core counts while the local-LLM bottleneck remains VRAM, which Nova Lake doesn’t address.
Source: GameGPU (leak roundup).
⇅ Click a column header to sort — the home inference shootout, per today’s numbers
Sourced claims only (vendor/project-reported). PS5LM figures are the project’s own benchmarks; RTX Spark numbers are launch-day reporting; LFM2.5’s 184 tok/s is on an M5 Max from this week’s launch teardown.
| Machine | Usable local mem (GB) | Representative tok/s | Model / note | Power | Price |
| PS5LM (jailbroken PS5) | 11.4 | ≈20 | Qwen-class 27B, aggressive quant (124 on Llama 3.2 3B, 61 on Gemma 4 E4B) | ~86W | ≈$450 used |
| RTX Spark ProArt GR1X | 128 | n/a | 120B models w/ 1M context locally; ~1 PFLOP FP4 | 140W TDP | from $2,599 |
| 2× M5 Ultra Mac Studio | 1024 | n/a | Muse Spark Open 816B via MLX (Meta’s day-0 recipe) | n/a | 2× Studio |
| M5 Max laptop | 96* | 184 | LFM2.5-24B (2B active); tool-calling pick | n/a | MacBook-class |
* M5 Max unified pool; figure not reconfirmed today — treat as reference, not news.
PS5LM throughput: tokens/s by model size (project’s own benchmarks)