# TensorCore.si > TensorCore.si is an independent educational website that explains NVIDIA Tensor Cores, GPU architecture and AI chips in plain language, with an interactive exploded view of a GPU, live simulations of matrix multiply-accumulate, an AI model hardware calculator and randomized math challenges that award verifiable certificates. Key facts: a Tensor Core computes D = A × B + C on small matrix tiles in one operation. Generations: Volta (2017, FP16), Turing (2018, INT8/INT4), Ampere (2020, TF32/BF16, 2:4 sparsity), Hopper (2022, FP8, Transformer Engine), Blackwell (2024, FP6/FP4 with microscaling), Rubin (2026, HBM4). TensorCore.si is not affiliated with NVIDIA. ## Interactive tools - [Homepage: exploded Tensor Core view and live simulation](https://tensorcore.si/): hover the chip to separate its layers; watch a 4×4 matrix multiply-accumulate run step by step. - [Free AI Hardware Tools and Calculators](https://tensorcore.si/tools/): Hub of all free TensorCore.si calculators, data and reference tools. - [LLM VRAM Calculator](https://tensorcore.si/vram-calculator/): Calculates the GPU memory needed to run an LLM (Llama, Qwen, DeepSeek, Gemma, Mistral, gpt-oss) by quantization (FP16, FP8, GGUF Q8_0 to Q2_K, GPTQ, AWQ) and context length, with GPU fit and tokens-per-second estimates for RTX, Mac and data-center GPUs. - [AI Training Time and Cost Calculator](https://tensorcore.si/training-cost-calculator/): Estimates LLM training compute (6 x parameters x tokens), GPU-hours, wall-clock time, cloud cost and energy on A100, H100, H200, B200, B300, MI300X, MI355X and TPUs. - [AI GPU and Accelerator Specs Database](https://tensorcore.si/gpu-specs/): Per-chip specs database for NVIDIA V100 to B300, AMD MI300X to MI355X and Google TPU v5p to Ironwood: memory, bandwidth, dense FP16/FP8/FP4 and FP64 throughput, power and interconnect, with sources and CSV download. - [Tensor Core Facts and Statistics](https://tensorcore.si/tensor-core-facts/): 25 verified, individually linkable facts and statistics about Tensor Cores, with primary sources. - [AI Number Format Converter](https://tensorcore.si/number-format-converter/): Converts any number to FP32, TF32, BF16, FP16, FP8 E4M3/E5M2 and FP4 E2M1 and shows the rounded value, error and bit pattern. - [Tensor Core Games](https://tensorcore.si/games/): Five browser games that teach Tensor Core matrix math, FP4/FP8/BF16 number formats, the roofline model, GPU history and which core runs which AI task. - [TensorCore.si Challenge](https://tensorcore.si/challenges/): Randomized Tensor Core math problems in three levels with downloadable, verifiable certificates. - [AI Model Explorer](https://tensorcore.si/ai-models/): Calculator for GPU memory, Tensor Core work per token and speed of AI models on H100, H200, B200 and B300. - [Tensor Core & AI Hardware Glossary](https://tensorcore.si/glossary/): Definitions of Tensor Core, GPU and AI hardware terms from A to Z. ## Guides and articles - [2:4 Structured Sparsity Explained: How Sparse Tensor Cores Double AI Throughput](https://tensorcore.si/structured-sparsity-explained/): What 2:4 structured sparsity is, how NVIDIA and AMD tensor cores use it to skip half the math, why spec sheets list "with sparsity" numbers, and how to… - [Mixed Precision Training Explained: FP16, BF16 and FP8 on Tensor Cores](https://tensorcore.si/mixed-precision-training/): How mixed precision training works, why it makes training up to several times faster on tensor cores, the difference between FP16 with loss scaling and… - [How to Run an LLM Locally: Ollama, LM Studio and llama.cpp (Step by Step)](https://tensorcore.si/run-llm-locally/): A beginner-friendly guide to running large language models on your own computer with Ollama, LM Studio or llama.cpp: hardware you need, which model to… - [How Much VRAM Do I Need to Run an LLM? Formula, Table and Examples](https://tensorcore.si/how-much-vram-for-llm/): A simple formula to work out how much VRAM a large language model needs: weights, KV cache and overhead. Includes a table for 8B to 405B models at FP16,… - [Intelligence Explosion Explained: Recursive Self-Improvement and the Singularity](https://tensorcore.si/intelligence-explosion/): The intelligence explosion is the idea that AI able to improve itself could rapidly become superintelligent. Where the idea came from, how recursive… - [What Is AI Alignment? The Alignment Problem Explained Simply](https://tensorcore.si/what-is-ai-alignment/): AI alignment means making AI systems pursue the goals their designers and users actually intend. What the alignment problem is, examples like reward… - [GPU Memory Bandwidth Explained: Why It Decides LLM Speed](https://tensorcore.si/gpu-memory-bandwidth/): What GPU memory bandwidth is, how to calculate it from bus width and memory speed, why it limits how fast large language models generate tokens, and how… - [GDDR7 vs HBM3E vs HBM4: GPU Memory Types Explained](https://tensorcore.si/gddr7-vs-hbm/): GDDR7 and HBM are the two kinds of memory used on GPUs. How each works, their bandwidth and capacity, why AI data center GPUs use HBM3E and HBM4 while… - [Google Ironwood TPU (TPU v7) Explained: Specs, Pods and How It Compares](https://tensorcore.si/google-ironwood-tpu/): Ironwood is Google's seventh-generation TPU, its first designed mainly for inference. Ironwood's specs, 192 GB of HBM per chip, 9,216-chip pods, how it… - [Cerebras WSE-3 Explained: The Wafer-Scale AI Chip With 900,000 Cores](https://tensorcore.si/cerebras-wse-3/): Cerebras builds an AI chip the size of a whole silicon wafer. How the WSE-3 works, its 4 trillion transistors, 900,000 cores and 44 GB of on-chip memory,… - [AI Demand Sparks a New Chip Shortage: Lead Times Stretch to 32 Weeks](https://tensorcore.si/ai-chip-shortage-2026/): Microchip, Nexperia and STMicroelectronics say AI data center demand is taking over factory, packaging and test capacity, pushing lead times for ordinary… - [Google Unveils TPU 8tx and TPU 8i Max, Expanding Its Eighth-Generation AI Chips](https://tensorcore.si/google-tpu-8tx-tpu-8i-max/): Google has added two specialised chips to its eighth-generation TPU family: TPU 8tx, a confidential-computing training chip co-designed with AMD, and TPU… - [When Will AGI and Superintelligence Arrive? Timelines and the Intelligence Explosion (2026)](https://tensorcore.si/when-will-agi-arrive/): When will AGI and superintelligence arrive? Forecasts from AI labs, researcher surveys and AI 2027's authors, the intelligence explosion, and what could slow it down. - [What Is AGI? Artificial General Intelligence Meaning Explained (2026)](https://tensorcore.si/what-is-agi/): What does AGI mean? Artificial general intelligence explained: how it differs from today's AI, definitions from OpenAI and DeepMind, tests like ARC-AGI, and whether AGI is here. - [Superintelligence by Nick Bostrom: Book Summary and Key Ideas](https://tensorcore.si/superintelligence-nick-bostrom/): Summary of Nick Bostrom's Superintelligence: Paths, Dangers, Strategies: paths to superintelligence, takeoff, the orthogonality thesis, instrumental convergence and the control problem. - [AI 2027 Explained: The Superintelligence Scenario and What Changed (2026)](https://tensorcore.si/ai-2027/): AI 2027 explained: the month-by-month scenario of AI reaching superintelligence by late 2027, its race and slowdown endings, who wrote it and how the authors changed their timelines. - [Should Superintelligence Be Banned? The Superintelligence Statement Explained (2026)](https://tensorcore.si/superintelligence-statement/): The Statement on Superintelligence calls for a ban until superintelligent AI is proven safe. What it says, who signed it, what Americans think and the case for and against. - [AI Safety in Australia: The AI Safety Institute, National AI Plan and Superintelligence (2026)](https://tensorcore.si/ai-safety-australia/): How Australia is preparing for advanced AI and superintelligence: the Australian AI Safety Institute, the National AI Plan, AI regulation without an AI Act, and global comparisons. - [Safe Superintelligence (SSI) Explained: Ilya Sutskever's Company and Plan (2026)](https://tensorcore.si/safe-superintelligence-ssi/): Safe Superintelligence Inc. (SSI) is Ilya Sutskever's AI lab. Founders, funding and $32B valuation, Google TPU and NVIDIA Vera Rubin deals, and what safe superintelligence means. - [AGI vs ASI: The Difference Between AGI and Superintelligence (2026)](https://tensorcore.si/agi-vs-asi/): AGI vs ASI explained: AGI matches humans at most tasks, ASI far surpasses them. Definitions of ANI, AGI and ASI, DeepMind's levels of AGI, and how one could lead to the other. - [What Is VRAM? Video Memory Explained for Gaming and AI (2026)](https://tensorcore.si/what-is-vram/): VRAM is the memory on your graphics card. Learn what it does, VRAM vs RAM, GDDR7 vs HBM, how to check yours and how much VRAM you need for gaming and AI. - [What Is an AI Data Center? How AI Factories Work, Explained (2026)](https://tensorcore.si/what-is-an-ai-data-center/): An AI data center trains and runs AI models on thousands of GPUs. See how it differs from a normal data center, what is inside, power and cooling needs, and the biggest sites. - [What Is a GPU Cluster? How AI Training Clusters Work (2026)](https://tensorcore.si/what-is-a-gpu-cluster/): A GPU cluster links many GPU servers with NVLink and InfiniBand to work as one. Learn how clusters are built, how training is split across GPUs and what limits them. - [What Are CUDA Cores? How Many You Need and What They Do (2026)](https://tensorcore.si/what-are-cuda-cores/): CUDA cores are the FP32 units inside NVIDIA GPUs. Learn what a CUDA core is, how SMs and warps work, core counts for RTX 50, RTX 40 and H100, and if they matter for AI. - [LLM Sampling Simulator: Temperature, Top-k and Top-p Explained](https://tensorcore.si/llm-sampling-simulator/): See how temperature, top-k, top-p and min-p change what a language model writes. Sample tokens yourself and watch how often it hallucinates a wrong answer. - [Roofline Model Simulator: Memory-Bound or Compute-Bound?](https://tensorcore.si/roofline-model-simulator/): Interactive roofline model for 11 AI chips including H100, B200, MI355X, TPU Ironwood and Rubin. Drag LLM decode, prefill and matmul workloads to see what limits speed. - [Attention Mechanism Simulator: Self-Attention Step by Step](https://tensorcore.si/attention-mechanism-simulator/): Type a sentence and watch transformer self-attention work: embeddings, queries, keys, values, scores, causal mask, softmax and output, with every number shown. - [LLM Inference Simulator: How GPUs Generate Tokens](https://tensorcore.si/llm-inference-simulator/): Interactive LLM inference simulator: pick a GPU, model, context and users, then watch prefill, decode, KV cache and VRAM. See why inference is memory-bound. - [What Is Superintelligence? How SI Could Be Built and What It Could Do](https://tensorcore.si/what-is-superintelligence/): Superintelligence explained: AI vs AGI vs SI, whether anyone has achieved it, the paths and components needed to build it, what SI could do that AI can't, and the risks. - [How AI Works: A Complete Guide From Beginner to Master](https://tensorcore.si/how-ai-works/): Learn how AI works step by step: training, tokens and transformers, why AI is accurate and why it makes mistakes, what hardware runs it, and master-level topics. - [Groq LPU vs GPU: How the LPU Works and NVIDIA's Groq 3 LPX](https://tensorcore.si/groq-lpu-vs-gpu/): Groq LPU vs GPU: on-chip SRAM and compiler scheduling for fast decode vs HBM capacity, NVIDIA's $20B Groq deal and the Groq 3 LPX with Vera Rubin. - [FlashAttention Explained: How It Makes Transformers Faster on GPUs](https://tensorcore.si/flashattention-explained/): FlashAttention explained: tiling and online softmax keep attention on chip and cut memory traffic. How it works, and versions 1 to 4 on A100, H100 and B200. - [Prefill vs Decode: How LLM Inference Runs on a GPU](https://tensorcore.si/prefill-vs-decode/): Prefill vs decode explained: why prompt processing is compute-bound and token generation is memory-bound, with H100 numbers, batching and split serving. - [How Much Does It Cost to Train an LLM? Compute, GPU-Hours and Real Examples](https://tensorcore.si/llm-training-cost/): How much does it cost to train an LLM? The 6ND rule, GPU-hours and dollars for 7B to 405B models, real Llama 3.1 and DeepSeek-V3 figures, hidden costs. - [B200 vs B300: NVIDIA Blackwell vs Blackwell Ultra Specs Compared](https://tensorcore.si/b200-vs-b300/): NVIDIA B200 vs B300 compared: 180 GB vs 288 GB HBM3e, 9 vs 13.5 PFLOPS dense FP4, same FP8 and FP16, FP64 and INT8 cuts, power, networking and which to choose. - [AMD vs NVIDIA for AI in 2026: ROCm vs CUDA, Instinct vs Blackwell and Rubin](https://tensorcore.si/amd-vs-nvidia-for-ai/): AMD vs NVIDIA for AI in 2026: Instinct MI355X vs B200, ROCm vs CUDA, AMD Matrix Cores vs Tensor Cores, and which GPU to buy for training or local LLMs. - [GPU vs CPU for AI: Why GPUs Run Machine Learning](https://tensorcore.si/gpu-vs-cpu-for-ai/): GPU vs CPU for AI explained: why GPUs run machine learning, parallel cores vs fast cores, Tensor Cores, memory bandwidth, real numbers and when a CPU is enough. - [What Is DLSS? How Tensor Cores Power AI Upscaling and Frame Generation](https://tensorcore.si/what-is-dlss/): What is DLSS? How NVIDIA's AI upscaling and frame generation run on Tensor Cores, every version up to DLSS 4.5, and which RTX cards get each feature. - [Tensor Cores vs RT Cores: What Each Does in an RTX GPU](https://tensorcore.si/tensor-cores-vs-rt-cores/): Tensor Cores vs RT Cores explained: matrix math for AI and DLSS vs ray tracing, how they work with CUDA cores, counts per RTX card and which ones matter for AI. - [What Is CUDA? NVIDIA's GPU Computing Platform Explained](https://tensorcore.si/what-is-cuda/): What is CUDA? NVIDIA's GPU platform explained: kernels, threads and warps, cuBLAS and cuDNN, compute capability, why AI depends on it and alternatives. - [TOPS vs TFLOPS: How to Compare AI Chip Performance Numbers](https://tensorcore.si/tops-vs-tflops/): TOPS vs TFLOPS explained: what each measures, how precision and sparsity inflate AI TOPS, and how to convert RTX, H100, B200 and NPU figures fairly. - [RTX 5090 for AI: Which LLMs It Runs, How Fast, and Is It Worth It?](https://tensorcore.si/rtx-5090-for-ai/): RTX 5090 for AI and local LLMs: 32 GB GDDR7 at 1.79 TB/s, which models fit (Qwen3 32B, Gemma 3 27B, gpt-oss-20b), tokens/s estimates and vs 4090 and DGX Spark. - [NPU vs GPU vs TPU: What Is an NPU and How Is It Different?](https://tensorcore.si/npu-vs-gpu-vs-tpu/): What is an NPU? NPU vs GPU vs TPU compared: design, TOPS, memory bandwidth and power, with Snapdragon, Intel, AMD and Apple NPUs vs RTX GPUs and Google TPUs. - [AMD MI355X vs NVIDIA B200: Specs, MLPerf Results and Which to Choose](https://tensorcore.si/mi355x-vs-b200/): MI355X vs B200 compared: 288 GB vs 180 GB HBM3e, FP4, FP6, FP8 and FP64 throughput, power, MLPerf 6.0 results, NVLink vs Infinity Fabric and CUDA vs ROCm. - [NVIDIA Rubin Explained: Vera Rubin NVL72 Specs, HBM4 and Rubin vs Blackwell](https://tensorcore.si/nvidia-rubin-explained/): NVIDIA Rubin explained: 288 GB HBM4 at 19.2 TB/s, 35 PFLOPS NVFP4, NVLink 6, the Vera CPU, NVL72 rack totals, the 2026 spec revision and Rubin vs Blackwell. - [NVFP4 Explained: NVIDIA's 4-Bit Format for Blackwell](https://tensorcore.si/nvfp4-explained/): NVFP4 explained: FP4 E2M1 values with an FP8 scale per 16 and an FP32 tensor scale, 4.5 bits per value, NVFP4 vs MXFP4, accuracy and how to use it. - [GB200 NVL72 Explained: 72 GPUs, One NVLink Domain](https://tensorcore.si/gb200-nvl72-explained/): GB200 NVL72: 72 Blackwell GPUs and 36 Grace CPUs in one NVLink domain, 130 TB/s NVLink, 13.4 TB HBM3e, 720 PFLOPS dense FP4. How it works and why it matters. - [B200 vs H100: Blackwell vs Hopper Specs and Speed](https://tensorcore.si/b200-vs-h100/): B200 vs H100: 2.3x FP16/FP8 Tensor Core throughput, 9 PFLOPS FP4, 180 GB at 8 TB/s, 1.8 TB/s NVLink and 1,000 W. Specs, speedups and verdict. - [H100 vs H200: Same GPU, More Memory (Specs Compared)](https://tensorcore.si/h100-vs-h200/): H100 vs H200: same Hopper GPU, but 141 GB of HBM3e at 4.8 TB/s vs 80 GB at 3.35 TB/s. What it means for LLM inference, training and cost. - [H100 vs A100: Specs, Speed and Which to Choose](https://tensorcore.si/h100-vs-a100/): H100 vs A100: 3.2x the dense FP16 TFLOPS, FP8, 1.6x the memory bandwidth, NVLink and power compared, with real-world speedups and buying advice. ## Topics - [AI Chips](https://tensorcore.si/category/ai-chips/): NVIDIA H100, H200, B200, B300 and Rubin, plus AMD Instinct and Google TPU, compared on compute, memory and what they are built for. - [GPU Architecture](https://tensorcore.si/category/gpu-architecture/): Inside the GPU: streaming multiprocessors, warps, CUDA cores, caches, HBM memory and the data paths that feed Tensor Cores. - [Guides](https://tensorcore.si/category/guides/): Plain-language guides to AI hardware concepts: precision formats, quantization, sparsity, FLOPS and how to read a GPU spec sheet. - [News & Analysis](https://tensorcore.si/category/news/): Launches, benchmarks and analysis of the GPUs and AI accelerators that train and run modern AI models. - [Simulations](https://tensorcore.si/category/simulations/): Interactive and animated explainers that show the math inside a Tensor Core step by step. - [Superintelligence](https://tensorcore.si/category/superintelligence/): The future of AI: AGI, superintelligence (SI), how it could be built, the hardware it needs, what it could do and the risks. The .si in TensorCore.si. - [Tensor Cores](https://tensorcore.si/category/tensor-cores/): How NVIDIA Tensor Cores work, from the first FP16 units in Volta to FP4 in Blackwell: matrix multiply-accumulate, precision and throughput explained. ## About and policies - [About TensorCore.si](https://tensorcore.si/about-us/) - [Contact Us](https://tensorcore.si/contact-us/) - [Disclaimer](https://tensorcore.si/disclaimer/) - [Privacy Policy](https://tensorcore.si/privacy-policy/) ## Optional - [Full text of all guides](https://tensorcore.si/llms-full.txt) - [XML sitemap](https://tensorcore.si/sitemap_index.xml) - [RSS feed](https://tensorcore.si/feed/)