b200

TokenSpeed is a speed-of-light LLM inference engine.

Pure Triton kernels for Qwen3.5-27B inference on NVIDIA B200

A USRP B200 compatible GPSDO board with the u-blox LEA-M8F

Rust + CUDA LLM inference engine for Blackwell (Tuned specifically on RTX PRO 6000, RTX 5090, B200): OpenAI-compatible (+converse and ant) serving, per-model X hardware exactness gates. NVFP4/mixed (fp8 hybrid, 4o6, etc - correctness, performance, hardware specific adapted) main quant support.

Serving and benchmarking 9 frontier MoE checkpoints (490 GiB+) with vLLM on a single 8x B200 node, plus a pre-download HBM fit checker

Run BEVFormer (ECCV 2022) on PyTorch 2.8 + Blackwell B200; SimpleDDP fix for mmcv 1.x on PyTorch 2.x + Tensor Core analysis

A Flexible and High-Performance Inference Serving Engine for Diffusion Language Models

Deploy DeepSeek-V4-Flash-0731 on dual NVIDIA RTX PRO 6000 Blackwell GPUs with vLLM PR #41834 (jasl fork) and DSpark speculative decoding, achieving ~200-227 tok/s in no-overseas-network environments.

Burst-serving Qwen3.8-27B FP8 on an on-demand RunPod B200, joined to a tailnet as b200 — no public endpoint. Measured: what MTP speculative decoding is worth, what it costs, and what is still unmeasured.

Compare H100, H200, B200, MI300X TCO across cloud providers and on-prem — open-source MCP agent demo.

Reproducible vLLM-Omni serving and benchmark reference for NVIDIA Cosmos3-Super on H200 and B200 GPUs

One-B200 DeepSeek V4 Flash deployment, correctness, and reproducible benchmark harness

Prebuilt spconv v2.3.8 wheels for CUDA 12.8 / 13.0 with native Blackwell (RTX 50-series, sm_120) kernels — for default PyPI torch (cu130) or torch +cu128

Spheron — independent third-party profile of a public API surface, by API Evangelist. Spheron Network is a decentralized GPU and cloud compute marketplace that aggregates enterprise-grade NVIDIA GPU capacity from certified Tier 3/4 data centers worldwide and exposes it through a single on-demand, per-minute billed interface.

GPU kernel lab for NVIDIA Blackwell (B200) — run CUDA, CuTe, CUTLASS, Triton, CuTe DSL, and Quack kernels from a single CLI using Modal.