Install
CUDA & Compute
CUDA, creator workloads, AI frameworks, and encoder/decoder blocks.
- 17 Tracked terms
- Last 30 days Feed window
What this topic collects on
An article joins this feed when it matches these terms. Each one is also a search of its own.
Related topics
Latest in CUDA & Compute
What I Learned by Dissecting gemma-4-E2B-it_qualcomm_sm8750.litertlm
2+ hour, 30+ min ago (910+ words) Gemma 4 Challenge: Write about Gemma 4 Submission I recently started exploring on-device AI inference, and honestly, the initial experience was overwhelming. Hundreds of models on HuggingFace, unfamiliar architecture names, quantization formats, chip-specific variants - it felt like drinking from a firehose. When…...
My Fine-Tuned Gemma 4 Loaded Fine, Then Broke on the First Message
2+ hour, 30+ min ago (743+ words) I fine-tuned Gemma 4 E2B. The adapter merged cleanly. The export to.litertlm completed without errors. I pushed the model to my phone, initialized the engine, and everything looked green. Then I tried to create a conversation and got this: No model…...
Nvidia's Vera CPU Packs a 1K-Entry Reorder Buffer in Its Custom Olympus Core
3+ hour, 18+ min ago (865+ words) Nvidia revealed its custom Olympus CPU core at Hot Chips, detailing a Vera processor with a reorder buffer exceeding 1,000 entries, 16-instruction-per-cycle fetch, and 18 issue ports. The design splits the pipeline into front end, mid core, and back end, with the…...
4-bit GGUF Quality for MoE Models: Why Only 3B of 180B Params Fire, and How to Prove Parity
3+ hour, 39+ min ago (944+ words) The practical consequence: the "180B" headline counts total parameters, while the compute and the active memory traffic track the much smaller active set, on the order of 3B. Two numbers describe an MoE and they answer different questions: Quantization touches storage, so…...
New FFT Library Pushes Four-Dimensional Transforms Across Multi-GPU Supercomputers
3+ hour, 47+ min ago (640+ words) A new open-source C++ library called FFTM brings portable distributed 3D and 4D fast Fourier transforms to CUDA and HIP multi-GPU systems, with autotuning and strong scaling results on up to 120 A100 GPUs....
Gemma 4 26B A4B vs Gemma 4 E4B: Benchmarks & Cost
2+ day, 7+ hour ago (388+ words) Updated October 10, 2026. Rank says Gemma 4 26B A4B is ahead. Price, access, and your workload can each overturn that. Public scores include evidence status and uncertainty. This is a same-family comparison, so migration details appear when the source data supports them. Gemma 4 26B A4B vs…...
Gemma 4 26B A4B vs Gemma 4 E2B: Benchmarks & Cost
2+ day, 7+ hour ago (388+ words) Updated October 10, 2026. Rank says Gemma 4 26B A4B is ahead. Price, access, and your workload can each overturn that. Public scores include evidence status and uncertainty. This is a same-family comparison, so migration details appear when the source data supports them. Gemma 4 26B A4B vs…...
Ministral 3 14B vs Raw Phi-4 mini direct logits: Comparison
2+ day, 7+ hour ago (53+ words) Ministral 3 14B vs Raw Phi-4 mini direct logits Estimated · Public rank #199 Raw Phi-4 mini direct logits has no comparable published API token rate. Raw Phi-4 mini direct logits Raw Phi-4 mini direct logits is not ranked on the public lane for…...
Akamai's Khaja Omer: One VLLM Flag Change Doubled Throughput on a Single GPU
15+ hour, 14+ min ago (1279+ words) One flag. That was the difference between an AI endpoint crawling at 2,000 milliseconds of latency and one running at 800 milliseconds with nearly 40% more throughput. In a live workshop environment, Akamai senior software engineer Khaja Omer changed a single VLLM setting…...
GLiNER2.5 multi vs Holo2-4B: Benchmarks & Cost
2+ day, 7+ hour ago (50+ words) BenchLM GLiNER2.5 multi vs Holo2-4B GLiNER2.5 multi has no comparable published API token rate. Holo2-4B has no comparable published API token rate. GLiNER2.5 multi and Holo2-4B are not ranked on the public lane for agentic tasks, so no winner is named for agentic tasks....