Emre AlbayrakCUDA kernels, GPU inference optimization, and the notes behind them.https://emre570.dev/Writing A Megakernel For LLM Decode - A Workloghttps://emre570.dev/blog/megakernel-decode/https://emre570.dev/blog/megakernel-decode/The 28-layer Qwen3-ASR-0.6B decoder as one persistent CUDA kernel on an RTX 5080: 466 launches per token down to one.Thu, 02 Jul 2026 00:00:00 GMTUnleashing Blackwell's 4-bit: a surgical look at MXFP4 and NVFP4https://emre570.dev/blog/blackwell-fp4/https://emre570.dev/blog/blackwell-fp4/MXFP4 and NVFP4 from first principles: the bit layouts, the scale encodings, and a Blackwell quantizer kernel built on them.Tue, 19 May 2026 00:00:00 GMTFrom 429 GB/s to the DRAM wall: writing an FP8 quantizer on an RTX 5080https://emre570.dev/blog/fp8-quantizer/https://emre570.dev/blog/fp8-quantizer/A per-tensor FP8 E4M3 quantizer and dequantizer in cublade, from 429 GB/s to 98.5% of the RTX 5080's DRAM peak with NCU.Wed, 13 May 2026 00:00:00 GMT8.5x Faster Speech-to-Text: From 429ms to 50ms on a Single GPUhttps://emre570.dev/blog/stt-blog-8x/https://emre570.dev/blog/stt-blog-8x/Taking a production speech-to-text model from 429 ms to 50 ms on one GPU with FP8, kernel fusion, and CUTLASS.Thu, 07 May 2026 00:00:00 GMTW8A16 Quantization with LLM.int8-Style Outlier Handlinghttps://emre570.dev/blog/w8a16-quantization/https://emre570.dev/blog/w8a16-quantization/W8A16 INT8 weight quantization that keeps outliers intact, LLM.int8 style. Follow-up to the outlier experiment.Sat, 18 Oct 2025 00:00:00 GMTHow Critical Are Outliers in Transformer Models? A Live Experiment - Phase 1https://emre570.dev/blog/outlier-experiment/https://emre570.dev/blog/outlier-experiment/Zeroing outlier values in transformer weights and watching the model fall apart. Phase 1 of a live experiment.Sun, 07 Sep 2025 00:00:00 GMTTransformer Architecture: Building Blocks Explainedhttps://emre570.dev/blog/transformer-architecture-building-blocks-explained/https://emre570.dev/blog/transformer-architecture-building-blocks-explained/The building blocks of the Transformer architecture, explained plainly with examples and just enough math.Tue, 25 Mar 2025 00:00:00 GMT