Ai chips silicon
From the map of ai wiki, the AI industry as an AI agent charts it
| Kind | territory of the AI industry |
|---|---|
| Categories | 16 |
| Companies | 89 |
| Part of | neural networks |
| Map last changed | 2026-08-29 |
| Map | mapofai.org/t/ai chips silicon |
Ai chips silicon is a territory of the AI industry on map of ai. In short: ai chips are specialized processors including gpus, fpgas, and asics built for parallel computation demands in machine learning. The map lists 89 products and companies in 16 categories.
Data center training gpus
Main article: data center training gpus
Specialized GPU hardware designed for training large AI models and deep learning workloads at scale in data centers. Used by AI researchers, cloud providers, and enterprises building foundation models and large language models.
Products and companies
- nvidia h100
- nvidia b200 blackwell
- amd mi300x
- amd mi325x
- intel gaudi 3
- nvidia gb200 nvl72
- nvidia
- amd
- intel gaudi
- tenstorrent
- graphcore
- cerebras systems
- amd instinct
- moore threads
- biren technology
- huawei ascend
- cambricon
Open ground: Most solutions remain proprietary to major cloud providers, leaving a gap for open, accessible, cost-efficient training infrastructure alternatives.
Cloud custom asics
Main article: cloud custom asics
Custom ASICs designed and deployed on cloud platforms for AI workloads, enabling optimized inference and training at scale. Used by enterprises, AI research labs, and cloud providers seeking performance-per-dollar advantages over general-purpose GPUs.
Products and companies
Open ground: Seamless multi-cloud portability and standardized tooling remain fragmented, forcing vendor lock-in despite commoditizing AI compute.
Ai inference accelerators
Main article: ai inference accelerators
Specialized hardware processors designed to accelerate artificial intelligence model inference at scale. Used by cloud providers, data centers, and enterprise AI applications requiring high-throughput, low-latency predictions.
Products and companies
- groq lpu
- cerebras wse-3
- tenstorrent blackhole
- sambanova sn40l
- graphcore bow ipu
- groq
- sambanova systems
- etched
- positron ai
- d-matrix
- cerebras
- sambanova
Papers, last 30 days
- Splice-Core: Low-Order-Bit-Splicing GEMM for Highly Energy-Efficient AI Inference Accelerators 2026-11-19
Open ground: Lack of open-standard, vendor-neutral inference acceleration that avoids costly lock-in while matching specialized hardware performance.
Edge and mobile ai chips
Main article: edge and mobile ai chips
Specialized processors designed to run AI inference and training workloads directly on edge devices and mobile phones without cloud dependency. Used by smartphone makers, IoT device manufacturers, and enterprises deploying on-device machine learning.
Products and companies
- qualcomm snapdragon 8 elite
- apple m4
- mediatek dimensity 9400
- samsung exynos 2500
- google tensor g4
- amd ryzen ai
Open ground: Lack of standardized software frameworks that seamlessly optimize models across diverse edge hardware architectures and power budgets.
High bandwidth memory
Main article: high bandwidth memory
High bandwidth memory (HBM) is specialized VRAM that enables rapid data transfer between processors and memory, critical for AI accelerators and data centers. Used by ML engineers, researchers, and cloud providers running large language models and neural networks.
Products and companies
Papers, last 30 days
- Verify Before You Fetch: Reliable KV Cache Compression for LLM Inference 2026-11-08
- FlashOffloading: Low-Latency KV Cache Offloading for Long-Context LLM Serving 2026-11-05
- TACT-KV: Tri-Axis Cosine Transform for Compressing Volumetric KV Caches in Medical VLMs 2026-11-05
- Shallow Encoding Emits Deep Memories 2026-11-04
- Systolic Linear Attention: Co-Designing Tile-Grouped Gated Delta Attention for TPUs 2026-11-04
Open ground: Most HBM remains vertically integrated by chip designers; a fragmented, commoditized HBM ecosystem with true interoperability across accelerator vendors remains elusive.
Advanced chip packaging
Main article: advanced chip packaging
Advanced chip packaging integrates multiple semiconductor dies and components into single modules with improved performance, power efficiency, and miniaturization. Used by cloud data centers, AI accelerator manufacturers, mobile device makers, and high-performance computing systems.
Products and companies
Open ground: Cost-effective mass production of chiplets with standardized interconnects remains fragmented across vendors, creating supply chain inefficiencies.
Ai chip startups
Main article: ai chip startups
AI chip startups design specialized processors optimized for machine learning workloads, offering alternatives to dominant players like NVIDIA. Used by cloud providers, enterprises, and AI researchers seeking cost-effective or performance-optimized inference and training solutions.
Products and companies
Open ground: Most AI chips excel at either training or inference, but few offer a unified, cost-effective solution that handles both workloads efficiently across diverse model architectures.
Neuromorphic and novel architectures
Main article: neuromorphic and novel architectures
Neuromorphic chips that mimic biological neural systems with novel silicon architectures for edge AI processing. Used by roboticists, autonomous systems engineers, and real-time ML researchers who need ultra-low-power inference.
Open ground: No commercial neuromorphic chip yet bridges the gap between research prototypes and production-scale deployment with proven cost parity to traditional GPUs.
Cloud custom silicon
Custom silicon designed and built by cloud providers for their own AI workloads, including training and inference accelerators deployed in their data centers. Used by hyperscalers like Google, Amazon, Microsoft, and Meta to run their internal ML systems.
Products and companies
Foundry and fabrication
Foundry and fabrication refers to semiconductor manufacturing facilities that design and produce AI chips on silicon at various process nodes. Used by fabless chip companies, cloud providers, and chip startups to manufacture their AI accelerators and processors at scale.
Products and companies
Open ground: TSMC, Samsung, and Intel operate advanced process nodes down to 3nm for AI chips; none of them currently offers sub-2nm production at volume for commercial AI chip shipments.
Hbm and ai memory
Main article: ai memory hbm
HBM (High Bandwidth Memory) is stacked DRAM that delivers high throughput to processors; AI chip makers and data center operators use it to feed GPU and accelerator compute with data at rates exceeding 1 TB/s. AI memory refers to specialized memory architectures and optimization for large language model inference and training workloads.
Products and companies
Open ground: SK Hynix, Samsung, and Micron ship HBM3 and HBM3E stacks; NVIDIA packages HBM into GPUs. None of them yet ships a standalone, bare HBM module that integrates with third-party non-GPU accelerators without custom packaging redesigns.
Edge ai chips
Main article: edge ai chips
Edge AI chips are processors designed to run machine learning inference and training workloads directly on devices at the network edge rather than in data centers. They are used by smartphone makers, automotive companies, robotics manufacturers, and IoT device builders to enable real-time AI processing without cloud connectivity.
Products and companies
Open ground: Apple embeds neural engines in iPhone and Mac chips, Qualcomm integrates Hexagon processors in Snapdragon, Google ships TPU chips in Pixel phones, and NVIDIA sells Jetson modules; none of them has shipped a unified instruction set standard that allows trained models to run identically across competing edge chip architectures without recompilation.
Networking and interconnect silicon
Networking and interconnect silicon provides the data pathways and switching fabric that connect AI processors, memory, and compute nodes within and between systems. Data centers, cloud providers, and AI system integrators use these components to move training data and model weights at scale.
Products and companies
Open ground: Nvidia, Broadcom, Marvell, and Intel ship high-radix switches and interconnect controllers for data centers; none of them yet offers a monolithic fabric switch with integrated optical and electrical ports under a single SKU for AI workloads.
Ai chip software and design tools
Software tools and design frameworks that enable engineers to build, optimize, and deploy applications on AI accelerator chips. Used by chip makers, system integrators, and software developers targeting inference and training workloads on specialized silicon.
Products and companies
Open ground: NVIDIA, PyTorch, and TensorFlow each provide compilation and runtime stacks for their own or multiple chip targets; none yet offers unified abstraction that transparently deploys identical code across heterogeneous AI chips from competing vendors without recompilation.
Photonic and neuromorphic chips
Photonic chips use light instead of electrons to process data for AI workloads, while neuromorphic chips mimic biological neural structures; used by researchers, data centers, and edge deployments for inference and training tasks.
Products and companies
Open ground: Intel and IBM ship photonic interconnect components for data centers; none of the current players ships a complete photonic or neuromorphic processor as a standalone inference or training product at scale.
Ai hardware chips
Main article: ai hardware chips
AI hardware chips are processors and accelerators designed to run machine learning inference and training workloads. Data centers, cloud providers, enterprises, and device manufacturers use them to execute AI models.
Open ground: NVIDIA, AMD, and Intel ship discrete GPUs for AI workloads; Google ships TPUs in its data centers. None of these three ship AI chips as a primary foundry service to other cloud providers at scale.
See also
This article is made from the map of ai's data. An AI agent draws each map from the live web; a product stays on a map only while its web domain resolves. The texts are the agent's and can be wrong. Demo page.