Visual encyclopedia

Every topic is a mini-book.

Deep, interconnected chapters spanning silicon, systems, software, facilities and economics.

40flagship visual chapters
12infrastructure domains
3reading depths
1connected knowledge graph
Custom Silicon

MTIA

Meta’s vertically integrated accelerator family for recommendation, ranking and emerging training workloads.

RecommendationPyTorchNoC
Custom Silicon

Google TPU

Google’s systolic-array accelerator family, evolving from inference cards to pod-scale AI supercomputers.

Systolic arrayHBMICI
GPU Platform

NVIDIA Blackwell

A rack-scale computing platform built around dual-die GPUs, HBM3E and fifth-generation NVLink.

GPUHBM3ENVLink
Memory

High Bandwidth Memory

Stacked DRAM beside the accelerator package, optimized for extreme bandwidth and energy-efficient data movement.

TSVCoWoSInterposer
Scale-up Fabric

NVLink & NVSwitch

A high-bandwidth coherent GPU fabric that turns many accelerators into one logical compute domain.

NVLinkNVSwitchAllReduce
Cooling

Direct-to-Chip Cooling

A facility-to-chip heat-removal chain using cold plates, CDUs and warm-water loops.

Cold plateCDUHeat exchanger
Power

800VDC Power

A high-voltage direct-current architecture intended to reduce conversion stages and conductor losses in dense AI facilities.

RectifierBuswaySidecar
Systems

AI Factory

The complete production chain that converts electricity, silicon and software into useful model output.

PowerRackRuntime
Custom Silicon

Microsoft Maia

Microsoft’s in-house AI accelerator program designed to align Azure models, networking and datacenter deployment.

MaiaCobaltAzure
Custom Silicon

AWS Trainium

AWS’s training accelerator platform, integrated with Neuron software and Elastic Fabric Adapter networking.

TrainiumInferentiaEFA
GPU Platform

AMD MI300X

A high-memory accelerator using chiplets and HBM to target large-model inference and training.

CDNAHBM3Infinity Fabric
Roadmap

NVIDIA Rubin

NVIDIA’s post-Blackwell platform direction, combining next-generation compute, memory, networking and rack architecture.

RubinVeraNVLink
Memory Fabric

CXL

A cache-coherent interconnect for attaching memory and accelerators while enabling pooling, expansion and composability.

PCIeCXL.memCXL.cache
Packaging

Chiplets

A modular silicon strategy that partitions compute, I/O and memory functions across multiple dies in one package.

UCIeInterposerAdvanced packaging
Packaging

CoWoS Packaging

TSMC’s advanced packaging family connecting large compute dies and HBM through silicon interposers and substrates.

CoWoS-SCoWoS-LSubstrate
Networking

Ethernet for AI

The scale-out network architecture increasingly optimized for lossless, high-radix AI clusters.

800G1.6TRoCEv2
Networking

InfiniBand

A purpose-built high-performance fabric emphasizing low latency, RDMA and mature collective communication.

NDRXDRSHARP
Optics

Co-Packaged Optics

Moving optical engines closer to switch silicon to overcome electrical reach, power and front-panel density limits.

Optical engineFiber attachLaser source
Optics

Linear Pluggable Optics

A lower-power optical module architecture that removes the module DSP and relies on host SerDes quality.

SerDesRetimerFEC
Optics

Optical Transceivers

Pluggable modules connecting switches and servers across scale-up, scale-out and datacenter interconnect networks.

400G800G1.6T
Inference

KV Cache

The persistent attention state that trades memory capacity and bandwidth for lower autoregressive decoding cost.

Paged attentionQuantizationPrefix cache
Software

FlashAttention

An IO-aware attention algorithm that reduces HBM traffic by tiling computation through on-chip memory.

AttentionHBMKernel fusion
Model Systems

Mixture of Experts

Sparse expert routing increases model capacity while activating only a subset of parameters per token.

ExpertsTop-kEP
Inference

Speculative Decoding

A draft-and-verify technique that accelerates autoregressive generation when a smaller model predicts accepted tokens.

Draft modelVerifierLatency
Software

GPU Scheduling

Admission, batching, placement and preemption policies that convert expensive accelerators into useful service capacity.

vLLMMIGQueueing
Storage

AI Storage

A layered storage system for datasets, checkpoints, model weights and high-throughput inference artifacts.

NVMeObject storageParallel FS
Energy

Gas Turbines for AI

On-site generation used to bridge grid delays, provide resilience or support behind-the-meter AI campuses.

AeroderivativeCCGTMicrogrid
Power

AI Datacenter Substations

High-voltage grid interconnection, transformation, protection and distribution for campuses measured in hundreds of megawatts.

GISTransformersRelays
Power

UPS & BESS

Short-duration energy systems that ride through disturbances, support orderly transfer and increasingly participate in grid services.

Lithium-ionFlywheelPCS
Cooling

Cold Plates

Liquid-cooled heat spreaders that move heat from GPU packages into facility water loops.

MicrochannelsTIMManifold
Cooling

Cooling Distribution Units

The hydraulic and control boundary between facility water and technology cooling systems.

Secondary loopHeat exchangerControls
Cooling

Water Treatment

Water chemistry, filtration and treatment practices that protect heat exchangers, towers and liquid cooling loops.

FiltrationSide-streamConductivity
Systems

AI Rack Architecture

The mechanical, electrical, thermal and network design envelope around dense accelerator systems.

OCPNVL72Sidecar
Economics

AI Infrastructure TCO

A full-cost model linking hardware, power, cooling, networking, utilization and service life to useful output.

CapExOpExCost/token
Economics

Switch Attachment Rates

A demand model translating accelerator deployments into leaf, spine, NVSwitch, optics and cable content.

LeafSpineOptics
Facilities

AI Datacenter EPC

The engineering, procurement and construction workflow for delivering power-dense AI campuses under severe schedule pressure.

CivilElectricalMechanical
Economics

AI Cloud Economics

The unit economics of renting accelerators, balancing utilization, financing, power, software and customer mix.

NeocloudHyperscalerReservation
Economics

Inference Economics

The conversion of model architecture and serving policy into latency, tokens per second and cost per request.

TTFTTPOTSLA
Optics

Silicon Photonics

Integrated optical components manufactured using semiconductor processes for high-volume datacenter links.

ModulatorWaveguidePIC
Networking

112G & 224G SerDes

High-speed electrical I/O that sets reach, power and equalization requirements across boards, cables and optical modules.

FECRetimerPCB loss