What stops a model from running locally is almost never the shader count. It is whether the weights and the context fit in memory at all, which is the constraint that decides whether any of the new integrated-GPU machines is useful to someone running models at home. IFA week produced two answers to it, one from NVIDIA and one from AMD, both landing within two months.
Two N1X configurations, and only one of them holds a large model
NVIDIA set October for RTX Spark systems and used the N1X name publicly for the first time. The part comes in two configurations: a 20-core Grace CPU with a 6,144-core Blackwell GPU and 24 to 128GB of unified memory, going into both laptops and mini PCs, and an 18-core CPU with a cut-down 5,120-core GPU and 24 to 32GB, laptops only, per Tom's Hardware on the two N1X tiers.
The 20 percent difference in CUDA cores is the less interesting half. A 24 or 32GB pool covers the 20 to 30 billion parameter models that most local agent setups run at 4-bit quantization, which is a genuinely useful machine. The 128GB ceiling is what puts a 100-billion-parameter mixture-of-experts model in memory at all, and it exists only on the larger die. Anyone buying one of these to run models rather than to own an Arm laptop should treat the memory option as the product and the core count as a footnote.
Two things NVIDIA has not stated are the ones that decide how it performs. The memory bus width is unpublished, so the bandwidth figure is unknown, and on a unified-memory part bandwidth, not capacity, sets the token generation rate once the model fits. NVIDIA's own IFA post puts the GPU at one petaflop without naming the precision behind the number, and a peak figure quoted that way is normally a low-precision one that will not line up against FP8 or FP16 throughput. Pricing and the exact list of systems shipping in October are also still open.
Partner specs give the first hard numbers around the chip. Lenovo's Yoga Pro 9n lists up to 128GB of LPDDR5X-9400 and a TDP of up to 80W on the 15-inch model, while the Yoga 9n 2-in-1 stops at 64GB. Acer is showing a compact desktop. AMD is in the same territory: Gorgon Halo surfaced at the show in Lenovo's ThinkCentre X Ultra, a Ryzen AI Max+ Pro 495 with a Radeon 8065S and 128GB of RAM, due in November. The large unified memory pool behind an integrated GPU is now a category with more than one vendor in it.
PAIR spreads agent sub-tasks across the PCs already on the network
NVIDIA also released Personal AI Router, a free and open-source beta that discovers other PCs on a local network and routes independent parts of an agentic workload to whichever of them has spare GPU cycles. It presents itself as a proxy that LM Studio and Ollama connect to, so the front end does not change; each participating node needs its own PAIR install and one of those runtimes, and discovery runs over mDNS with an IP fallback, as Tom's Hardware describes the setup. Nodes do not have to hold the same models, though more copies of a given model widen the pool that can take a request. Hardware support runs from GeForce RTX 20 Series and newer through RTX PRO workstation cards, DGX Spark, and Apple M4 silicon, with clients on Windows, macOS and Linux.
The limits are structural rather than teething problems. PAIR reserves nothing, so a node whose owner starts a game simply stops contributing, and there is no quality-of-service guarantee to plan around. It helps only where a task genuinely splits into independent sub-agents, which describes long agentic runs and not a single chat stream. Most importantly, it does not enlarge the model anyone can run: every node still executes whatever fits in its own memory, so three 16GB cards remain three 16GB cards and not a 48GB one.
Faster llama.cpp and vLLM, if there is 24GB to spare
The inference gains NVIDIA quoted are up to 1.9x higher throughput in llama.cpp on an RTX 5090 from kernel work, changed speculative decoding and faster prefill, plus 1.2x in vLLM on an RTX PRO 6000 Blackwell Workstation Edition and up to 1.4x across a pair of DGX Sparks. These are vendor-measured with no methodology or model list published, and "up to" is carrying its usual weight, but the work is upstream in llama.cpp and vLLM and reaches users through LM Studio and Ollama rather than a proprietary runtime, which is what makes it worth tracking.
The one-click local agent setups arriving alongside it are gated harder than the speedups are. Both the OpenClaw Windows app and Perplexity's Portable Computer want an RTX GPU with at least 24GB of VRAM, which on the GeForce side means a 3090, 4090 or 5090. On the 8 to 16GB cards most people are running, none of that convenience layer applies and the working path is unchanged: pick a smaller model, quantize it yourself, and watch context length rather than parameter count.