标签：#gpu

Onur Satici explains how Vortex, an open-source columnar file format under the Linux Foundation, revolutionizes high-throughput data loading. He details how cascading lightweight encodings, layout-based segment pruning, and zero-copy memory pipelines eliminate CPU/NVMe bottlenecks to stream S3 data straight to GPUs at speeds up to 60 Gbps without requiring upfront data reprocessing. By Onur Satici

James Hall discusses the strategic and technical imperative of moving AI workloads from cloud providers to local edge devices. He shares practical approaches using WebGPU, Transformers.js, and DuckDB to achieve near-native performance in JavaScript. Through real-world case studies, he explains how to minimize data privacy risks, optimize browser inference, and build rigorous evaluation suites. By James Hall

📝 Originally published (in Japanese) at forge.workstyle.tech . I Found a Diffusion TTS That Generates Voices from Just a Caption "Calm adult female narrator voice. Slow, warm, and trustworthy delivery, carefully reading long sentences." Pass this description, and it will speak in exactly that voice. You can generate any Japanese voice without preparing even one second of speaker audio. Moreover, the same caption and random seed will always produce the same voice. I thought this could be useful for interactive avatars—eliminating the hassle of preparing voices for each character. In short: it didn’t work for conversations. But it was too valuable to discard, so I repurposed it. Benchmark: 2.5x Slower on the Same GPU Our existing system used a pre-trained TTS (Style-Bert-VITS2-based). We synthesized the same 7.5-second sentence on the same GPU slice to compare. Engine Conditions Generation Time RTF Diffusion TTS 40 steps 2g.20gb 1.9–2.1s 0.25–0.28 Diffusion TTS 24 steps 2g.20gb 1.4–2.0s 0.19–0.27 Diffusion TTS 16 steps 2g.20gb 1.1–1.4s 0.15–0.19 Pre-trained Model 2g.20gb 0.74–0.81s 0.115–0.127 RTF (Real Time Factor) = generation time ÷ audio length. Lower is faster. At 40 steps, it’s 2.5× slower. Even cutting steps to 16 still leaves a 1.5× gap. And below 16 steps, quality collapses (6 steps was immediately rejected on listening). On a smaller GPU slice, the gap widens further. Engine Conditions Generation Time RTF Diffusion TTS 40 steps 1g.10gb 3.4–4.0s 0.46–0.54 Diffusion TTS 16 steps 1g.10gb 2.0s 0.27 In conversation, audio is synthesized and played sentence-by-sentence. The delay until the first sentence plays determines the user experience. A 1–2 second lag per turn is prohibitive. Upgrading GPUs Didn’t Solve It I considered allocating more GPU resources—from 2g.20gb to 7g.80gb (4× the compute). But measurements showed a fixed overhead of ~1.1 seconds . Even cutting steps to 12, generation time bottomed out at 1.1–1.4 seconds. This is due to model loading and tex

📝 Originally published (in Japanese) at forge.workstyle.tech . Building an Unmanned System for 3D Avatar Live Streaming We're developing an unmanned system where 3D avatars automatically handle live streaming. The system boots up a cloud GPU pod at the scheduled start time, the renderer assembles and streams the video, and then the pod is discarded when the segment ends. Since there's no human oversight, three factors directly impact the success of the business and service quality: "how many avatars can run simultaneously," "how the system recovers from failures," and "how quickly it starts up." These questions couldn't be answered through estimates alone. Renting a GPU for a few hours costs only a few hundred yen. In this article, we'll share three stories of how we measured and designed the system, following the structure of "stumbling block → cause → solution." Capacity : How many avatars can run on a single GPU? The answer is 4, at a monthly cost of ¥7,600 per avatar. However, the bottleneck wasn't the GPU. Reliability : Despite using the same image, some hosts crashed every 60 seconds. We implemented a mechanism to automatically switch to a different host. Startup Speed : Reduced the time from pod startup to streaming start from 4 minutes to 95 seconds. These three aspects seem independent but are actually interconnected. Faster startup enabled practical host switching, and understanding capacity allowed us to set prices. Let's dive into each one. 1. How Many Avatars Can Run on a Single GPU? - Measured Result: 4 The first number we desperately needed was "how many avatars can run on a single GPU." Without this, we couldn't determine pricing, and without pricing, we couldn't assess the business viability. Estimates were useless, so we measured it. Here are the results: Item Measured Value GPU RTX 4000 Ada (Community type, $0.28/hour) Simultaneous Streams 4 avatars maintaining 720p30 in real-time (Recorded segment: 89 seconds / 89 seconds) GPU Usage 26% Bottlenec

The new generation of data center systems is increasing efficiency with smarter traffic control instead of just more processor cycles.

This article provides a step by step deployment guide for serving Google's Gemma 4 on an AWS EC2 G5g instance using pure JAX. The code is here: github.com/xbill9/gemma4-dev What is this project trying to Do? This project aims to serve a modern open model on the cheapest whole CUDA GPU AWS will rent you, and to measure honestly what that costs. Aren't You Using The Wrong GPU? Probably! The T4G is a Turing chip from 2018. It has no bfloat16 and no fp8. But it is cheap, it is available when nothing else is, and it is attached to a Graviton2 host — which makes G5g the rare hardware axis that almost nothing in the ML ecosystem targets: aarch64 and CUDA together . So let's give pure JAX a shot on G5g! AWS EC2 G5g G5g instances pair an AWS Graviton2 (64-bit Arm) processor with NVIDIA T4G Tensor Core GPUs. At g5g.xlarge they are the cheapest EC2 instance carrying a whole NVIDIA GPU , and the only Arm-based GPU family AWS offers. Two GPU instances are cheaper per hour and neither can serve this model (us-east-1, Linux, on-demand, checked against the Pricing API on 2026-08-28): g6f.large at $0.2020 is genuinely NVIDIA and genuinely CUDA — but it is one eighth of a GPU with 3 GB , and the weights alone are 6.155 GB. The first g6f that fits is g6f.4xlarge at $0.9500, which is 1.7x this rig's g5g.2xlarge . g4ad.xlarge at $0.3785 carries an AMD Radeon Pro V520 — no CUDA at any price. Among whole NVIDIA GPUs, G5g is the floor: g5g.xlarge at $0.4200, and the next one up is g4dn.xlarge at $0.5260. More information is available here: https://aws.amazon.com/ec2/instance-types/g5g/ The default in this rig is g5g.2xlarge — 1 GPU, 8 vCPU, 16 GiB RAM. Note- the T4G reports 15,360 MiB of device memory, not the nominal 16 GB. Budget against the measured number. Gemma 4 Gemma is Google's family of open models built from the same research as Gemini. This rig serves google/gemma-4-E2B-it , the instruction-tuned reference release. JAX JAX is Google's array computing library — NumPy semantics, c

Nvidia is worth more than most countries because we cannot figure out how to do cheap...

📝 Originally published (in Japanese) at forge.workstyle.tech . Introduction In recent years, the demand for GPUs has surged with the widespread adoption of AI and machine learning. Particularly, the demand for the latest NVIDIA GPUs is notable, leading to an increase in GPU cloud services offered by cloud providers. Among these developments, containerized data centers are gaining attention. This article explores cost reduction in GPU cloud services and the latest trends in containerized data centers. Main Content Benefits of Containerized Data Centers Containerized data centers offer the following advantages: Rapid Deployment : Compared to traditional building-based data centers, they can be deployed much faster. Suitability for Regional Locations : They can be utilized as locally-focused infrastructure. Expanded Options : Alongside traditional data centers, containerized options are increasingly being offered as an alternative. Efforts to Reduce GPU Power Consumption GPU power consumption significantly impacts data center operational costs. Recent efforts include: Power Consumption Reduction : Proof-of-concept experiments have reported up to 80% reduction in power consumption. Visualization Technology : Combined with locally-focused containerized data centers, next-generation infrastructure solutions are being explored. Key Trends in Japan In Japan, the following trends are observed: Cloud Provider Entry : Multiple cloud providers have launched GPU cloud services. Adoption of Latest GPUs : Services are being deployed with early adoption of NVIDIA's latest GPUs (e.g., H200). Power Issue Solutions : High-efficiency conversion, high-density, and containerization are proposed as solutions to power-related challenges. Conclusion With the growing demand for GPU cloud services, containerized data centers are attracting significant attention. Their benefits, including rapid deployment and reduced power consumption, are diverse. In Japan, services adopting the latest GPUs a

I want to be upfront about something: this whole project runs on free Kaggle T4 notebooks, an AWS EC2 t3.micro relay that costs almost nothing, and public internet. No A100s. No private datacenter network. No budget. And yet, ShardFlow v2.1 hits 28.10 TPS peak on Qwen2.5-7B across two separate cloud regions over WAN. This is the story of how that happened, and specifically the one fix in v2.1 that I did not see coming. The Problem: Running a 7B Model When You Have No Money A 7B parameter model in FP16 needs roughly 15 GB of VRAM. A single Kaggle T4 has 16 GB. Technically it fits, barely, with nothing left over for a KV cache. The solution is tensor parallelism: split the model across two machines. Node 0 (Iowa) handles layers 0 to 14. Node 1 (Oregon) handles layers 14 to 28, plus the LM head and final verification. They talk to each other through a TCP relay running on an EC2 t3.micro in Ohio. The baseline throughput with this setup and no tricks: 4.92 TPS. Usable, but not fast. Speculative Decoding: The Idea LLM inference is slow because it's sequential. You generate one token, wait, generate another, wait. Each round trip across WAN costs you ~86ms RTT. At 1 token per round trip, you're fighting the network the whole time. Speculative decoding flips this. Instead of sending one token at a time, you run a tiny draft model locally to guess the next K tokens ahead. Then you send all K guesses to the verifier in one shot. If the big model agrees with M of them, you've committed M tokens in a single round trip instead of one. ShardFlow uses Qwen2.5-0.5B as the draft model, running on cuda:1 of Node 0 while the 7B target slice runs on cuda:0. Zero VRAM contention. The drafter proposes 8 candidates, Node 1 verifies them all in parallel, and you get an average of 4.07 tokens per round trip instead of 1. With speculative decoding in eager mode: 14.3 TPS peak. 3x better. The Wall I Hit I thought 14.3 was the ceiling. The network was the obvious bottleneck: two Kaggle instan

The AI buildout shows no signs of slowing. And with hundreds of billions of dollars a year going into data centers and GPUs, compute has become the single biggest cost for anyone building AI products. But for all that spending, there still isn’t a straightforward way to put a price on compute — or for firms to hedge their exposure when the price changes. Silicon Data […]

Choosing the Right GPU for Your Model — A Sizing Method, Not a Guess OK, you're a senior SRE, you've been hearing incessantly about AI models, but aren't quite sure how to determine the correct node size to host your model. - If so ... you're in the right place. Part of a series on running vLLM on AKS. Companion piece: How to avoid flapping . GPU infrastructure setup — coming soon. This piece walks through estimating GPU memory requirements from both a model's parameter count or a concurrent requests requirement. After reading this article you will have enough knowledge to pick a GPU family with confidence. Disclaimer: this process is a rule-of-thumb filter, not a precise calculation — the last step covers how to get exact numbers once the model is actually running. Background: What actually consumes GPU memory AI models live in GPU memory — VRAM — and engines such as vLLM provide novel techniques for managing that memory efficiently [ paper ], but the model isn't the only thing consuming it. Below is a short list of things that consume our precious VRAM: Model weights — the parameters themselves. The big fixed cost: loaded once, never shrinks. KV cache — working memory for in-flight requests. Every token of every active request holds its attention keys/values here. This is the one that determines throughput : more KV cache = more concurrent requests. Everything else — activations (the temporary tensors of a forward pass) plus CUDA/framework overhead. You don't calculate these by hand; vLLM measures activations with a profiling pass at start-up and prints it for our consumption. The sizing question is really: after weights and overhead, how much is left for the KV cache — and is that enough for your traffic? OK, lets get started Step 1 — Choose a model Guidance on which model to choose is outside the bounds of this article. What matters here: once you have a candidate, everything below can be read off its spec sheet — you can then run this method on every model on y

Jordan Nanos discusses how semiconductor constraints, data center expansion, and networking bottlenecks impact AI software architecture. Drawing from SemiAnalysis research, he shares insights on benchmark performance, GPU scaling, and tokenomics from chip fab to model inference. By Jordan Nanos

In mid-2026, FP8 and FP4 have become essential tools for making large-scale AI training and inference more efficient . FP8 uses two main formats-E4M3 for better precision on activations and weights, and E5M2 for wider dynamic range on gradientswhile NVIDIA’s NVFP4 takes things further with 4-bit values and micro-block scaling (shared FP8 scales per 16 elements plus a tensor-level scale). These formats dramatically cut memory use and increase throughput on modern GPUs compared with traditional BF16 or FP16, making it possible to train and serve bigger models on the same hardware. The benefits are clear: roughly 2× memory savings with FP8 and up to 3.5× with NVFP4, higher Tensor Core performance, and better energy efficiency. The trade-offs come from reduced numerical range and precision, which can lead to accuracy loss or instability unless carefully managed with techniques such as delayed scaling, stochastic rounding, Hadamard transforms, and selective quantization that skips sensitive layers. When these methods are applied properly, accuracy often stays within 1–2 % of higher-precision baselines on real workloads. Research has moved quickly from the foundational 2022 FP8 paper to 2025 studies showing stable FP4 pre-training of multi-billion-parameter models. Hardware support is mature for FP8 on Hopper GPUs and reaches its peak on Blackwell with native NVFP4 and MXFP8 acceleration. Among frameworks, PyTorch currently leads with native float8 dtypes, Transformer Engine for production training, and TorchAO for optimized inference. JAX offers solid support through Transformer Engine, TensorFlow/Keras provides simpler quantize-to-FP8 options but relies more on TensorRT for high performance, and libraries such as bitsandbytes remain useful for complementary 4-bit memory savings. Practical adoption is already strong for both training and inference, especially when teams start with proven recipes, monitor scaling factors, and prototype on smaller models. Workarounds for r

600 MiB of headroom My RTX 4070 was running Qwen 35B beautifully after the --cpu-moe trick from a previous run. The tokens/sec were where I wanted them. VRAM sat at 11,714 MiB out of 12,281 — 95% full. That leaves 600 MiB. Not enough for a serious agent. The context window I was giving llama.cpp was -c 4096 . Fine for chat. Not fine when a Claude Code-style agent hands the model 12,000 tokens of tool definitions before it says hello. I wanted -c 32768 . That's an 8× jump. And the memory that grows with context length is the KV cache. Multiply the cache by 8 with 600 MiB free, and llama.cpp dies during warm-up. I know because I tried it first. What actually sits on the GPU After offloading the MoE experts to CPU (the previous chapter's trick), the GPU is holding two things: The attention weights and non-MoE parameters The KV cache — a running record of every token the model has already read The first is fixed. The second grows linearly with context length. Double the context, double the cache. -c 4096 → -c 32768 doesn't just want 8× more tokens processed, it wants 8× more cache resident in VRAM the whole time. There is no room. So the cache itself has to shrink. Two flags llama.cpp takes two flags for KV cache dtype: llama-server -m qwen35.gguf -ngl 99 --cpu-moe -c 32768 \ -ctk q8_0 -ctv q8_0 -ctk is the Key cache, -ctv is the Value cache. Default is f16 (16-bit). q8_0 cuts each in half. Halving both means the KV cache footprint drops by roughly 50%. That freed-up VRAM is exactly what I need to make the context 8× bigger without touching the model weights. The measurement Same prompt, same seed, two runs — one at f16 KV, one at q8_0 KV: KV dtype Max -c I could allocate Tokens/sec (decode) Perplexity delta f16 (default) 4096 ~34.6 baseline q8_0 32768 ~34.1 negligible in my tests The speed loss is inside noise. The context is 8× longer. The quality drop I could not tell apart from run-to-run variance. Community measurements agree: symmetric q8_0 KV lands somewhere unde

I'm training an AlphaZero-style agent (Gumbel MuZero via DeepMind's mctx ) to lay out working factory modules for Factorio: the network places machines, belts and inserters on a grid, and the reward comes from an exact throughput verifier. Everything runs in JAX on a single RTX 5070: 128 environments in one batch, an action space of A = 1729 (3 entity types × 144 cells × 4 rotations + "done"), and a small 474k-parameter conv net in bf16. While benchmarking training configurations I hit this: MCTS simulations per move training throughput XLA compile time 16 143 episodes/s 5 s 32 47 episodes/s 13 s 64 9 episodes/s 60 s Doubling the simulation budget should roughly double the cost — each simulation is one network call plus some tree bookkeeping. Instead, 16→32 costs ×3 and 32→64 costs ×5 . Something in the search was superlinear, and this post is the story of finding it in the compiled HLO and fixing it by rewriting one ~80-line function ( PR #116 ), with bitwise-identical search results. Ruling out the network First, components in isolation (batch 128): one network evaluation takes ~0.8 ms , and the rest of recurrent_fn (environment step + observation + legal-action mask) adds almost nothing on top — the whole function is also ~0.8 ms. So at 64 simulations the network accounts for roughly 50 ms per move. But a full policy step at 64 simulations costs 362 ms . Hundreds of milliseconds were going somewhere else. To localize them I benchmarked three variants of the same policy step: full — production setup; no-net — network replaced by constant logits, real environment; tree-only — no network and no environment: recurrent_fn returns the embedding unchanged. Nothing left but mctx's own tree machinery. sims full no-net tree-only 8 10.7 ms 3.7 ms 3.8 ms 16 20.9 ms 8.3 ms 8.3 ms 32 64.7 ms 34.0 ms 33.7 ms 64 362.1 ms 124.7 ms 125.0 ms The pure tree machinery is superlinear all by itself. Per simulation it costs 0.47 → 0.52 → 1.05 → 1.95 ms as the budget goes 8 → 16 → 32 → 64

A practical framework for test planning, hardware scheduling, artifact traceability, failure classification, and evidence-based quality gates Disclaimer: The views expressed in this article are my own. The architecture, examples, terminology, and code snippets are generalized for educational purposes and do not describe or disclose any employer’s proprietary systems, confidential information, or internal implementation details. A software change can compile successfully, pass unit tests, and still introduce a serious GPU regression. The failure may appear only on one GPU generation. It may depend on a particular driver, firmware revision, operating system, graphics API, or workload. A change may preserve functional correctness while quietly reducing performance. It may also cause an intermittent failure that disappears when the test is rerun. This is why GPU validation cannot be treated as conventional CI/CD with a GPU runner attached to the end of the pipeline. A dependable GPU validation platform must coordinate: Software and firmware artifacts Hardware configurations Test coverage GPU resource scheduling Failure classification Performance baselines Engineering evidence It must do all of this while operating under an important constraint: compatible GPU capacity is limited and expensive. The objective is not simply to run more tests. It is to produce reliable evidence quickly enough to support engineering decisions. Why conventional CI/CD is not enough A conventional application pipeline often resembles: Commit ↓ Build ↓ Unit tests ↓ Integration tests ↓ Deployment A GPU validation pipeline is more multidimensional: Code or configuration change ↓ Build software and firmware artifacts ↓ Determine affected GPU configurations ↓ Reserve compatible hardware ↓ Prepare the driver and runtime environment ↓ Run functional, stability, and performance tests ↓ Collect logs, traces, metrics, and crash artifacts ↓ Classify failures and compare results with baselines ↓ Make a mer

You don't need to know anything about Go to read this. The game is just the fixed yardstick. The story is a hardware benchmark: the same program, the same problem, the same settings — only the machine changed, from a 2017 GPU cluster to a single 2026 graphics card. That makes it a rare clean measurement of one decade of progress. What "solving" means here There are two very different things a computer can do with a board game. It can play it well — that's what AlphaGo did. Or it can solve it: mathematically prove the outcome under perfect play from both sides, leaving no doubt. Solving is the hard one. You explore an enormous tree of "if I play here, they play there…" move sequences until you have an airtight proof. Each node in that tree is one position examined. The target here is a single 7x7 opening called JA . In 2023, a NeurIPS paper ( Game Solving with Online Fine-Tuning , Wu et al.) proved its verdict — the attacker cannot win — using a cluster of twelve GTX 1080Ti GPUs running 384 parallel workers. The solver is guided by a neural network that estimates how hard each branch is, and crucially that network is fine-tuned online — it keeps learning during the solve. I rebuilt that exact solver (same code, same problem, same initial model, same search settings) and ran it on one RTX 5090 . It reached the identical proof . Everything but the hardware was held fixed, so the two runs line up as a generation-vs-generation benchmark — and it doubled as a full shakedown of the new Blackwell workstation. The numbers 1x RTX 5090 (2026) 12x GTX 1080Ti (2017) ratio Worker slots 24 384 1/16 the parallelism Per-slot throughput 284 nodes/s 141 nodes/s 2.01x faster Search work to proof 1.01B nodes 1.73B nodes 0.59x (41% less work) Avg work per sub-job 4,189 nodes 6,136 nodes shallower proofs Live model updates 4,007 208 19.3x more Wall-clock time 41.4 h 8.9 h 4.64x slower Verdict loss (proven) loss identical The single card finished slower in wall-clock time (41 h vs 9 h) — b

A practical introduction to WebGPU, WGSL, render pipelines, compute shaders, and the future of high-performance graphics on the web. Your browser can stream 4K video, run a complete code editor, render complex 3D scenes, and host multiplayer games. But for years, web developers accessed the GPU through an API based on an older generation of graphics programming. WebGPU changes that contract. WebGPU is not simply a faster version of WebGL. It is a new approach to graphics and parallel computation on the web—one built around explicit pipelines, modern GPU architecture, compute shaders, predictable resource management, and a shader language designed specifically for the browser. This article expands on the progression presented in the uploaded High Performance Graphics: Introduction to WebGPU material: why WebGPU matters, how it differs from WebGL, how WGSL works, how the rendering pipeline is constructed, and how compute shaders extend the GPU beyond graphics. WebGPU is not WebGL 3.0 This is the first mental model to correct. WebGPU does not build on WebGL. WebGL exposes a browser-friendly version of the OpenGL ES programming model. WebGPU instead uses concepts associated with modern GPU APIs and provides a portable abstraction over the graphics capabilities available on the user’s system. WebGPU and WGSL are W3C standards for accessing GPU acceleration from web applications. The API supports both graphics rendering and general-purpose parallel computation. ( W3C ) WebGL │ └── OpenGL ES-style state machine WebGPU │ ├── Explicit pipelines ├── Explicit resource bindings ├── Command encoding ├── Compute shaders └── Modern GPU execution model The difference is architectural, not cosmetic. In WebGL, you frequently change global rendering state and then issue a draw call. In WebGPU, you describe the pipeline and resources more explicitly, record commands, and submit those commands to the GPU. WebGL mental model Change state ↓ Change more state ↓ Bind resources ↓ Draw WebGPU

The AI hardware landscape has shifted significantly in 2026, with NVIDIA, AMD, and Intel all competing for developers who need GPUs capable of running local large language models and AI inference workloads. Choosing the right GPU for AI workloads requires looking beyond marketing numbers and focusing on the specifications that actually affect real-world performance. Memory capacity, memory bandwidth, and software ecosystem maturity consistently matter more than theoretical compute peaks when running transformer models locally. This comparison covers the most relevant workstation and prosumer GPUs available in mid-2026, including NVIDIA's Blackwell architecture (RTX 50-series), AMD's Radeon AI Pro R9700, and Intel's Arc Pro B70. The goal is to provide a practical reference for developers deciding which hardware best fits their model sizes, software stack, and budget constraints. Which GPU specifications matter for AI workloads Marketing materials from GPU vendors emphasise AI TOPS and tensor performance, but these metrics rarely tell the complete story for local inference. The specifications below are ranked by their actual impact on running large language models. VRAM capacity VRAM is typically the first limiting factor when running LLMs locally. A model cannot execute entirely on the GPU if it does not fit into available memory. Once model weights spill into system RAM, inference performance drops dramatically. Approximate VRAM requirements for common model sizes: Model Size Recommended VRAM 7B 8-12 GB 14B 16 GB 32B 24-32 GB 70B 48-64 GB 120B+ Multiple GPUs For most homelab users, moving from 16 GB to 32 GB of VRAM provides a substantially larger practical benefit than increasing raw compute performance. A 32 GB GPU capable of running an entire model will often outperform a theoretically faster 16 GB GPU forced to offload tensors into system memory. Memory bandwidth Memory bandwidth determines how quickly model weights can be streamed into compute units. Large tran

Concept design for a building is slow and expensive. A homeowner planning an extension, or a contractor trying to win a job, is stuck between two bad options: pay a drafter $500–2,000 for a concept package, or fight SketchUp's learning curve for a week. Meanwhile the actual idea — "a 4-bed duplex with a garage and a palm out front" — fits in one sentence. So I built Forge3D Spaces : you type that sentence, and a few seconds later you're walking through a furnished 3D house in your browser — with measured floor plans, DXF for AutoCAD, and a cost estimate that come out of the same model. No install. Here's how it works under the hood. The pipeline: sentence → structured plan → building The naive approach — "ask an LLM to emit a 3D scene" — falls apart fast. Models are bad at spatial consistency; walls don't meet, rooms overlap, doors float. So the LLM never touches geometry directly. It emits a structured program , and a deterministic solver turns that into a watertight building. The prompt becomes a spec. A strict JSON-schema call (OpenRouter, json_schema response format with every field required) turns "4-bed duplex with a garage" into a room program: room types, target areas, adjacencies, storeys. A slicing-tree solver lays it out. This is the old floorplanning trick from chip design — recursively split a rectangle with horizontal/vertical cuts until every room has its area. A squarify pass keeps rooms from collapsing into corridors. The output is exact rectangles with real dimensions, guaranteed non-overlapping and gap-free. Walls, openings, roof, furniture get generated from the solved plan. Every door and window is placed by rule, not by vibes. Because the plan is a real data structure, the 2D floor plan, the 3D model, the elevations, and the bill of quantities are all views of the same thing . Drag a wall and they all move together. Nothing drifts out of sync, because there's nothing to sync — it's one model. The rendering: WebGPU, and the fallback you actually