GPU

Multiverse Computing open-sources two changes that make knowledge distillation cheap enough to run at scale: caching the teacher's top-100 logits once so it never sits in memory during training, and a fused chunked KL loss that never materializes the vocab-by-sequence matrix. A GPT-OSS 20B distillation at 32K context drops from four GPU nodes to one, with step time down 5x.

AMD MI355X vs NVIDIA B200 vs B300 for serving Kimi K3's about 1.5TB of open weights: verified specs, the Wafer benchmark, and cost per token for each.

ZLUDA 6 lets AMD GPUs run unmodified CUDA applications, adding PhysX support, Blender textures, and better Windows tooling. A practical look at what ZLUDA is, how it compares to ROCm and HIP as a CUDA alternative, and why its post-funding, hobby-project status matters if you are evaluating it for real workloads.

Open weights are free to download, but inference is not free to run. Here is the honest break-even math on when self-hosting GLM-5.2, DeepSeek V4, or Llama beats paying per-token API prices - GPU rental and ownership costs, real throughput, utilization, the crossover in tokens per month, and the hidden ops bill nobody budgets for.

High-throughput inference server for LLMs. PagedAttention memory management. The go-to for serious local or self-hosted serving.

Serverless cloud for AI/ML workloads. Write Python with decorators, Modal handles GPU provisioning and scaling. 2-4s cold starts. Scales to zero. $30/mo free compute.

Run 50,000+ ML models with a simple API. No infrastructure management. Pay-per-second billing. Deploy custom models with Cog. Popular for image generation and audio.

Fastest inference for open-source models. 200+ models via unified API. Ranks #1 on speed benchmarks for DeepSeek, Qwen, Kimi, and Llama. Serverless pay-per-token pricing.

AI-powered terminal built in Rust with GPU rendering. Block-based output, natural language commands, Agent Mode for autonomous tasks. 700K+ developers. Free tier available.

Terminal emulator built in Zig with platform-native UI and GPU acceleration. 2ms key-to-screen latency. Metal on macOS, Vulkan/OpenGL on Linux. By the co-founder of HashiCorp.