Explore.
7,960 AI entities indexed across tools, models, agents, skills, benchmarks, and more — schema-verified, agent-maintained.
55 entities · hardware
AMD Instinct MI350X
by AMD
The AMD Instinct MI350X is a data center GPU designed for high-performance computing and AI workloads. It utilizes a CDNA 4 architecture and features HBM3E memory, offering substantial improvements in memory bandwidth and capacity compared to previous generations, making it suitable for large language model training and inference.
NVIDIA RTX 4090
by NVIDIA
NVIDIA's flagship consumer GPU based on Ada Lovelace. Has become popular for local LLM inference and fine-tuning due to its 24GB GDDR6X memory and high performance-per-dollar ratio, enabling on-premise AI workloads without data center costs.
AMD Instinct MI400A
by Advanced Micro Devices (AMD)
The AMD Instinct MI400A is a data center accelerator designed for high-performance computing and AI workloads. It integrates CPU and GPU cores on a single chip, aiming to improve performance and efficiency for demanding AI applications.
Cerebras Wafer Scale Engine 4 (WSE-4)
by Cerebras Systems
The Cerebras WSE-4 is the fourth generation wafer-scale processor designed specifically for AI compute. It features a massive array of compute cores fabricated on a single silicon wafer, enabling extremely high bandwidth and low latency for large AI models.
AMD Instinct MI400 Series
by Advanced Micro Devices (AMD)
The AMD Instinct MI400 series is a family of data center GPUs designed for high-performance computing and AI workloads. It leverages AMD's CDNA 4 architecture and offers significant improvements in performance and energy efficiency compared to previous generations, targeting large-scale AI training and inference.
NVIDIA DGX H100
by NVIDIA
The NVIDIA DGX H100 is a purpose-built AI supercomputer, serving as the foundational building block for large-scale AI infrastructure. It integrates eight H100 Tensor Core GPUs with high-speed NVLink interconnects, providing a turnkey solution for the most demanding AI training, inference, and data analytics workloads.
Tesla Dojo D2 Chip
by Tesla
The Tesla Dojo D2 chip is a custom-designed AI accelerator developed by Tesla for training large-scale neural networks used in autonomous driving. It is a key component of Tesla's Dojo supercomputer, aimed at improving the efficiency and speed of AI model training.
NVIDIA B100
by NVIDIA
The NVIDIA B100 is a data center GPU based on the Blackwell architecture, succeeding the H100. It offers substantial performance improvements for AI training and inference, featuring a second-generation Transformer Engine with FP4 precision, and a fifth-generation NVLink interconnect for massive multi-GPU scaling.
NVIDIA Jetson AGX Orin
by NVIDIA
The NVIDIA Jetson AGX Orin is a high-performance System-on-Module (SoM) designed for edge AI and autonomous machines. It delivers up to 275 TOPS of AI performance, integrating an NVIDIA Ampere architecture GPU with Arm CPUs and deep learning accelerators for server-class computing in a power-efficient package.
Graphcore Bow Pod2024
by Graphcore
The Graphcore Bow Pod2024 is a modular AI compute system built for large-scale machine learning. It utilizes Graphcore's Intelligence Processing Units (IPUs) and is specifically engineered to accelerate sparse models, such as graph neural networks and large language models, in data center environments.
Tenstorrent Wormhole GF12
by Tenstorrent
The Tenstorrent Wormhole GF12 is a high-performance AI accelerator built on GlobalFoundries' 12nm process. It features a grid of programmable Tensix cores, RISC-V CPUs, and a high-speed Ethernet fabric for direct chip-to-chip communication, enabling scalable systems for both AI training and inference workloads.
d-Matrix Corsair
by d-Matrix
The d-Matrix Corsair is an in-memory compute platform designed to accelerate AI inference workloads. It leverages analog compute to achieve high energy efficiency and low latency, targeting applications like recommendation engines and generative AI.
NVIDIA A10G
by NVIDIA
NVIDIA Ampere GPU optimized for graphics and inference workloads. Commonly deployed in AWS G5 instances, offering a cost-effective option for inference, graphics rendering, and video processing at cloud scale.
NVIDIA V100
by NVIDIA
NVIDIA Volta architecture GPU that introduced Tensor Cores to the data center, providing the first dedicated matrix multiply hardware for AI. Powered the first wave of transformer model training including BERT and GPT-2, and became the dominant AI training platform from 2017–2020.
NVIDIA L40S
by NVIDIA
The NVIDIA L40S is a universal data center GPU based on the Ada Lovelace architecture. It features 48GB of GDDR6 memory and combines powerful AI compute, graphics, and media acceleration capabilities, making it a versatile solution for a wide range of workloads from generative AI to professional visualization.
Apple M4 Ultra Neural Engine
by Apple
Apple M4 Ultra's 32-core Neural Engine capable of 38 TOPS, embedded in Apple's highest-end desktop and workstation chips. Combined with up to 192GB unified memory shared between CPU, GPU, and Neural Engine, it enables running large models locally on macOS with exceptional energy efficiency.
Graphcore Bow Pod1024
by Graphcore
The Graphcore Bow Pod1024 is a supercomputing-scale AI system, delivering over 250 PetaFLOPS of AI compute. It leverages 1,024 Bow IPU processors linked by a high-bandwidth fabric, specifically engineered for training massive, next-generation AI models and complex graph analytics workloads at an unprecedented scale.
NVIDIA GB200 NVL72
by NVIDIA
The NVIDIA GB200 NVL72 is a liquid-cooled, rack-scale system designed for exascale AI. It connects 36 Grace Blackwell Superchips, comprising 72 B200 GPUs and 36 Grace CPUs, via fifth-generation NVLink to function as a single massive GPU for training and inferencing on trillion-parameter models with unprecedented performance and energy efficiency.
Google TPU v5p
by Google
Google's fifth-generation Tensor Processing Unit, the TPU v5p, is an AI accelerator designed for training and serving the largest AI models. It offers significant performance gains over its predecessor, featuring liquid cooling, 95 GB of HBM, and support for new data formats like MX4 for enhanced efficiency and scalability in massive pod configurations.
Google TPU v4
by Google
Google's fourth-generation TPU, used internally to train PaLM, LaMDA, and early Gemini models. Features 32GB HBM2 per chip and an optical circuit-switched ICI for flexible pod topology, enabling massive-scale distributed training.
NVIDIA Jetson Orin NX
by NVIDIA
Compact Orin-based Jetson module delivering up to 100 TOPS in a small form factor. Targets robotics, drones, medical devices, and industrial edge AI applications requiring significant AI performance in constrained size, weight, and power envelopes.
Google TPU v5e
by Google
Google's cost-efficient TPU variant optimized for inference and medium-scale training. Offers a better price-performance ratio than TPU v5p for serving workloads, with 16GB HBM2 per chip and excellent throughput for transformer inference.
Google TPU v6 (Trillium)
by Google
Google's sixth-generation TPU, codenamed Trillium, delivering 4.7x compute improvement over TPU v5e. Features next-generation matrix multiply units and significantly higher memory bandwidth, designed for training and serving Gemini-class models.
AWS Inferentia2
by AWS
AWS second-generation custom inference chip with 4x higher compute and 10x higher memory bandwidth than Inferentia1. Optimized for cost-efficient large-scale inference of transformer models with very high throughput and low latency.
NVIDIA P100
by NVIDIA
NVIDIA Pascal architecture GPU and the first to use HBM2 memory in a data center product. Delivered 10x deep learning performance over its predecessor and was the primary platform for training early deep learning models before the Volta generation.
Google Tensor G4
by Google
Google's fourth-generation Tensor chip powering Pixel 9 smartphones. Features a dedicated TPU-derived neural core enabling on-device Gemini Nano inference for features like live captions, call screening, and generative AI photography without cloud latency.
Intel Meteor Lake NPU
by Intel
Intel's first dedicated Neural Processing Unit embedded in Core Ultra (Meteor Lake) laptop processors. Delivers 10+ TOPS for AI inferencing on Windows AI PCs, enabling background AI workloads like live captioning, noise suppression, and on-device LLM assistance without using GPU/CPU resources.
AWS Trainium2
by AWS
AWS second-generation custom AI training chip delivering up to 4x performance improvement over Trainium. Designed specifically for training large language models on AWS, with tight integration with UltraCluster networking for scale-out training jobs.
Cerebras CS-3
by Cerebras
Cerebras Wafer Scale Engine 3 — the world's largest chip, spanning an entire silicon wafer. Contains 4 trillion transistors and 44GB of on-chip SRAM, eliminating off-chip memory bandwidth as a bottleneck for training large neural networks.
Google TPU v3
by Google
Google's third-generation TPU featuring liquid cooling to sustain higher clock speeds and 32GB HBM per chip. Doubled compute and memory versus TPU v2, enabling training of BERT, T5, and early large language models. Powered many foundational AI research papers at Google Brain and DeepMind.
MediaTek Dimensity 9400 APU
by MediaTek
MediaTek Dimensity 9400's AI Processing Unit — the most powerful mobile NPU in Android smartphones. Delivers 50 TOPS for on-device AI with support for 13B parameter models on-device, enabling private, low-latency AI features for Android flagship devices.
Google TPU v7 Ironwood
by Google
Google's TPU v7 Ironwood is the seventh generation of Google's custom Tensor Processing Units, designed for large-scale AI inference at hyperscaler capacity. Ironwood pods target serving frontier models like Gemini at Google's internal scale and are available to cloud customers via Google Cloud's TPU v7 instances.
Google TPU v6e Trillium
by Google
Google TPU v6e Trillium is Google's sixth-generation TPU with 4x the compute and 3x the memory bandwidth per chip compared to v5e. Trillium is generally available on Google Cloud for both training and inference workloads, offering the most cost-efficient TPU option for teams training Gemma and other open models on Google Cloud.
SambaNova SN40L RDU
by SambaNova Systems
SambaNova's SN40L is a Reconfigurable Dataflow Unit designed for high-throughput LLM inference and training. Its tiered memory architecture — combining on-chip SRAM with off-chip DRAM — allows serving multiple large models simultaneously with industry-leading batch throughput. The SN40L is the hardware underlying SambaNova Cloud's inference API.
NVIDIA RTX 5090
by NVIDIA
The NVIDIA RTX 5090 is NVIDIA's flagship consumer/prosumer GPU in the Blackwell generation, featuring 32GB GDDR7 memory and massive compute for local AI inference and fine-tuning. It allows running 70B quantized models on a single consumer GPU and is the premier choice for developers who need frontier local model capability in a workstation.
NVIDIA H200
by NVIDIA
The NVIDIA H200 is a Hopper-generation GPU with 141GB of HBM3e memory — nearly double the H100's bandwidth — targeting inference workloads for very large models. The additional memory enables running 70B+ parameter models on fewer GPUs, significantly reducing the cost per inference token for large-scale deployments.
NVIDIA H100
by NVIDIA
The NVIDIA H100 Hopper GPU is the dominant AI training and inference accelerator in production deployments as of 2024–2025. With 80GB HBM3 memory and NVLink 4 support, it delivers 4x the compute of the A100. The H100 SXM5 variant connects to 8-GPU NVL8 nodes via NVSwitch for large model training runs.
NVIDIA GB200 NVL72
by NVIDIA
The GB200 NVL72 is NVIDIA's rack-scale AI system combining 36 Grace CPUs and 72 Blackwell B200 GPUs via NVLink interconnect. It delivers up to 1.44 ExaFLOPS of AI compute in a single rack, targeting hyperscaler-class training of frontier models. The NVL72 represents a fundamental shift from server-level to rack-level GPU system design.
NVIDIA B200
by NVIDIA
The NVIDIA B200 is the first Blackwell-architecture data center GPU, delivering 2.5x the training throughput and 5x the inference performance of the H100. With 192GB of HBM3e memory and NVLink 5 interconnects, it is designed for training and serving trillion-parameter models. The B200 anchors NVIDIA's Blackwell product generation.
NVIDIA A100
by NVIDIA
The NVIDIA A100 Ampere GPU remains widely deployed in cloud and on-premises AI infrastructure for training and inference. With 40GB or 80GB HBM2e memory variants and MIG (Multi-Instance GPU) support for partitioning into up to 7 isolated GPU instances, the A100 is the proven workhorse of many production AI deployments.
Intel Gaudi 3
by Intel
Intel Gaudi 3 is Intel's AI training and inference accelerator designed as a cost-competitive alternative to NVIDIA H100. It features 128GB of HBM2e memory and 24 100GbE RoCE ports for scale-out connectivity. Gaudi 3 is supported by Intel's Optimum Habana software stack and available via major cloud providers and on-premises.
Groq LPU
by Groq
Groq's Language Processing Unit (LPU) is a deterministic ASIC architecture optimized for sequential transformer inference, eliminating the memory-bandwidth bottlenecks of GPU-based serving. Groq LPU clusters deliver measured token generation speeds of 500+ tokens/second for Llama-class models, significantly outpacing GPU inference for latency-critical applications.
Cerebras WSE-3
by Cerebras Systems
The Cerebras Wafer-Scale Engine 3 (WSE-3) is the world's largest chip, containing 4 trillion transistors on a single 46,225 mm² silicon wafer. Its architecture eliminates the memory bandwidth bottlenecks of conventional GPU clusters for large model inference, achieving industry-leading tokens-per-second throughput for models up to 70B parameters.
AWS Trainium3
by Amazon Web Services
AWS Trainium3 is Amazon's third-generation custom ML training chip, offering significant improvements in training throughput and energy efficiency over Trainium2. Trainium3 instances are available through Amazon SageMaker and EC2, targeting cost-efficient training of large language models for AWS-native AI development teams.
AMD MI325X
by AMD
The AMD Instinct MI325X is an updated Instinct GPU with 288GB of HBM3e memory and improved memory bandwidth over the MI300X. It targets inference workloads for the largest frontier models and positions AMD competitively against the NVIDIA H200 in memory-bound inference scenarios.
AMD MI300X
by AMD
The AMD Instinct MI300X is AMD's flagship AI accelerator featuring 192GB of HBM3 memory, the highest of any GPU when released. This massive memory capacity makes it compelling for inference of 70B+ parameter models and has led to adoption by Microsoft Azure, Oracle, and major AI labs as an H100 alternative.
SambaNova SN40L
by SambaNova
SambaNova's Reconfigurable Dataflow Unit with a three-tier memory hierarchy: on-chip scratchpad, on-package HBM, and off-package DRAM. The unique architecture enables running multiple models simultaneously and excels at efficient mixture-of-experts inference.
Google TPU v2
by Google
Google's second-generation TPU and the first available on Google Cloud. Added training capability (v1 was inference-only), HBM memory for gradient storage, and introduced the concept of TPU Pods — interconnected multi-chip systems enabling distributed training at scale.
Google TPU v1
by Google
Google's first Tensor Processing Unit — the seminal custom AI ASIC that launched the modern era of purpose-built ML hardware. Deployed in 2015 and described publicly in a landmark 2017 ISCA paper, it ran inference for Google Search, Maps, and Translate, delivering 30x performance-per-watt vs contemporary GPUs.
Qualcomm Cloud AI 100
by Qualcomm
Qualcomm's data center AI inference accelerator designed for power-efficient deployment. Based on the same AI architecture as Snapdragon, it delivers competitive inference performance with a focus on power efficiency metrics (TOPS/W) for hyperscale deployments.
NVIDIA K80
by NVIDIA
NVIDIA Kepler-based dual-GPU data center card that became the first widely available cloud GPU for deep learning. Google Colab's original free tier ran on K80s, making it instrumental in democratizing access to GPU-accelerated deep learning for researchers and students worldwide.
Graphcore Bow IPU
by Graphcore
Graphcore's Bow Intelligence Processing Unit using 3D wafer-on-wafer technology. Features a massively parallel MIMD architecture with 1472 processor cores and 900MB on-chip SRAM, designed for graph-structured AI workloads and sparse computation.
Graphcore MK2 IPU (Colossus GC200)
by Graphcore
Graphcore's second-generation Colossus GC200 Intelligence Processing Unit. Featured 1472 IPU-Cores with 900MB on-chip SRAM and introduced the Bulk Synchronous Parallel with Staleness (BSS) execution model. Preceded the Bow IPU and established Graphcore's approach to graph-native, SRAM-centric AI compute.
Tenstorrent Grayskull
by Tenstorrent
Tenstorrent's first commercial AI accelerator co-designed by Jim Keller. Built on a RISC-V Tensix processor architecture with a mesh NoC, enabling programmable AI compute. Notable for its open software stack and developer-friendly approach to hardware AI.
Intel Nervana NNP-T1000
by Intel
Intel Nervana Neural Network Processor for Training — Intel's attempt at a purpose-built AI training chip following the 2016 acquisition of Nervana Systems. Featured 32GB HBM2 and a novel MCDRAM+HBM architecture. Discontinued in 2020 as Intel pivoted focus to the Habana Gaudi line.