More information
The following documents provide further explanations of the terminologies that are relevant to the corresponding topics:
Terminology
The following terms apply to AI Hypercomputer.
Accelerator
A specialized hardware device, such as a GPU or TPU, designed to perform
specific tasks, such as machine learning workloads, more efficiently than a
general-purpose processor.
Block
A collection of interconnected sub-blocks. Within a block, any GPU is reachable through the network fabric, which ensures ultra-low-latency communication for
your distributed training tasks.
Cluster
A collection of interconnected blocks joined by a high-speed fabric.
The non-blocking east-to-west network fabric lets you scale massive training
jobs across thousands of GPUs without encountering performance or bandwidth bottlenecks.
Clustered GPU
A category of machine families for GPUs, including A4X, A4, A3 Ultra, A3 Mega, and A3 High (8 GPUs). Clustered GPUs let teams deploy and scale thousands of interconnected accelerators as a single, tightly coupled system by using specialized networking fabrics and synchronized maintenance. They are ideal for workloads that require extreme compute, memory, and high-speed networking synchronization. Common use cases include pre-training foundation models with trillions of parameters, frontier model serving, and simulations for drug discovery.
Emergent maintenance
Sometimes referred to as emergency maintenance. An unplanned maintenance event caused by a critical hardware, software, or security issue. For supported accelerator machine types, emergent maintenance provides an advance notification instead of an immediate disruption. This notification gives you time to gracefully drain your workloads and optionally trigger the repair before the system terminates the VMs.
General GPU
A category of GPU machine families, including G2, G4, A2, N1+T4, and A3 Edge. General GPUs let teams deploy and scale NVIDIA accelerators with complete operational autonomy, bypassing the architectural complexity of coordinated supercomputing clusters. They are ideal for workloads that prioritize high availability, self-service provisioning, and independent scaling. Common use cases include real-time inference, RAG, prototyping, and small-to-medium model training.
Network fabric
A network fabric provides high-bandwidth, low-latency connectivity across
all blocks and Google Cloud services in a cluster. Jupiter is Google's data
center network architecture that uses software-defined networking and
optical circuit switches to evolve the network and optimize its
performance.
Node or host
A single physical server machine in the data center. Each host has
associated compute resources, such as accelerators. The number and
configuration of these compute resources depend on the machine family.
Compute Engine instances are provisioned on top of a physical
host.
NVLink domain
An NVLink domain, also referred to as a sub-block, is the core unit of
capacity for A4X Max and A4X machines. An NVLink domain consists of 18 A4X Max
or A4X instances (72 GPUs) that are connected by a multi-node NVLink
system.
Sub-block
A group of hosts and associated connectivity hardware that are on a single
physical rack. In the context of A4X Max and A4X machines, a sub-block is also
referred to as an NVLink domain.
Superblock
A physical grouping of interconnected machines within data centers. By placing
machines inside a superblock, you achieve the extreme network throughput and
sub-millisecond latency required to execute tightly coupled training workloads.