Persist and reuse KV Cache to speedup your LLM.

Shuhai is a benchmarking-memory tool that allows FPGA programmers to demystify all the underlying details of memories, e.g., HBM and DDR4, on a Xilinx FPGA [FCCM 20]

HBM is an application to authorize and manage authorized docker commands using Docker AuthZ plugin

NATSA is the first near-data-processing accelerator for time series analysis based on the Matrix Profile (SCRIMP) algorithm. NATSA exploits modern 3D-stacked High Bandwidth Memory (HBM) to enable efficient and fast matrix profile computation near memory. Described in ICCD 2020 by Fernandez et al. https://people.inf.ethz.ch/omutlu/pub/NATSA_time-…

DreamRAM: A Fine-Grained Configurable Design Space Modeling Tool for Custom 3D Die-Stacked DRAM

A Scalable BFS Accelerator on FPGA-HBM Platform

A Scalable BFS Accelerator on FPGA-HBM Platform

GTCEu 1.12.2 Pack, Large.

GateSeeder is the first near-memory CPU-FPGA co-design for alleviating both the compute-bound and memory-bound bottlenecks in short and long-read mapping. GateSeeder outperforms Minimap2 by up to 40.3×, 4.8×, and 2.3× when mapping real ONT, HiFi, and Illumina reads, respectively.

Repo of high bandwidth memory papers, codes, ...

The complete AI stack, bottom-up from sand to a served token — a source-backed Obsidian vault. 15 domains, 110 topics, beginner to expert: semiconductor physics, fabs, GPUs, datacenters, power, cooling, training, and inference.

Convert NHibernate XML mapping to Code mapping

Simulates data communication between host memory and HBM on an AMD Alveo U55C FPGA board. Uses PCIe with QDMA to transfer data efficiently between the host and HBM.

Deadline-aware KV-cache scheduling for protecting decode-critical request-state under long-context LLM inference pressure.

Open-source AI infrastructure bottleneck radar powered by LLMs, tracking HBM, GPUs, power, networking, cooling, CoWoS, and other supply-chain constraints.

Block-sparse for v5e-1 eliminates the O(N²) HBM attention matrix via online fused softmax, crossing the 240.2 FLOPs/byte ridge point at N≥1024

high bandwidth memory study

Repo for my master thesis