Kimi K3 requires an 8ÃB300 node with 288GB of HBM per GPU due to its 1.4TB weight, exceeding the memory capacity of the 8ÃB200 node. This configuration results in approximately 20% higher hardware costs while supporting 16 concurrent sessions.

An RTX 5080 and RTX 3090 setup achieves over 80 tokens per second on the Qwen 3.6 27B Q8 model. The RTX 3090, with 24GB of memory, significantly enhances performance, allowing for initial speeds of 30 tokens per second, increasing to 50-60 tokens per second with MTP.

WebAssembly modules on Apple Silicon can share linear memory directly with the GPU, eliminating the need for copies, serialization, or intermediate buffers. This allows the CPU and GPU to read and write the same physical bytes, enabling efficient end-to-end computation without serialization overhead.

AMD's ROCm software stack aims to compete with Nvidia's CUDA for data center GPU market share. Success in this endeavor is viewed as a significant challenge due to CUDA's established dominance.

MegaTrain is a memory-centric system that enables the full precision training of large language models with over 100 billion parameters on a single GPU. It utilizes host memory to store parameters and optimizer states, treating GPUs as transient computation units.

Claude Code was given access to 16 GPUs on a Kubernetes cluster and submitted approximately 910 experiments over 8 hours. It determined that scaling model width was more significant than any single hyperparameter and achieved a 2.87% improvement in validation performance, reducing val_bpb from 1.003 to 0.974.

nCPU is a CPU architecture that operates entirely on GPU, utilizing tensors for registers, memory, flags, and the program counter. All arithmetic operations, including addition, multiplication, bitwise operations, and shifts, are performed through trained neural networks, with specific methods like Kogge-Stone carry-lookahead for addition and learned byte-pair lookup tables for multiplication.