2. Product description

2.1. Bow Pod₂₅₆ reference design

Graphcore’s Bow Pod₆₄ reference design assembles 16 Bow-2000 IPU-Machines together into a logical rack delivering over 22 petaFLOPS of AI compute. The Bow Pod₆₄ can be used individually (64 Bow IPU processors) or as a building block for larger systems such as the Bow Pod₂₅₆ (256 Bow IPUs), going up to 1024 Bow Pod₆₄ racks (64 K Bow IPU processors) delivering nearly 23 exaFLOPS of AI compute.

The Bow Pod₂₅₆ is built from 4 Bow Pod₆₄ racks so contains 64 Bow-2000 IPU-Machines delivering over 89 petaFLOPS of AI compute.

Each Bow Pod₆₄ combines the sixteen Bow-2000 IPU-Machines with network switches and a host server in a rack configuration (switches and host server not provided by Graphcore). The Bow Pod₆₄ system assumes the following default components:

• 1 - 4 approved host servers, see the approved server list for more details. In this datasheet we use the Dell R6525 server with dual-socket AMD Epyc2 CPUs as the default offering. Default number of servers is 1, however up to 4 host servers may be required depending on workload - please speak to Graphcore sales.

• 1 Arista 7060X ToR switch (32x100G + 2 10G)

• 1 Arista 7010T management switch (48p 1G+ 4x1/10G)

• 1 x Arista DCS-7060PX4-32-F (GW-Link switch)

The Bow Pod₂₅₆ is characterized by the following features:

• Disaggregated host architecture allows for different server requirements based on workload

• 89.23 petaFLOPS (FP16.16) of AI compute, 22.31 petaFLOPS @ FP32, and up to 16.6 TBytes of memory

• 2D-torus IPU-Link topology

A high-level view of the cabling for an individual Bow Pod₆₄ is shown in Fig. 2.1.

Fig. 2.1 Bow Pod₆₄ reference design rack

A Bow Pod₂₅₆ system consists of four Bow Pod₆₄ logical racks with optical GW-Link cables used to connect the Bow Pod₆₄ racks together via two redundant switches. Since the GW-Links are optical Ethernet cables the Bow Pod₆₄ logical racks can be installed further apart if necessary, if datacentre layout does not allow for them to be installed adjacent to each other.

Fig. 2.2 shows the Bow Pod₂₅₆ layout and GW-Link cables between the four Bow Pod₆₄ logical racks. A logical rack is a space in one or more physical racks occupied by a single Pod. Since the standard racking of a single Pod may not be possible within one physical rack, we use the term “logical rack” to refer to the set of components making up the single Pod, regardless of where they may be physically installed.

Bow Pod₂₅₆ systems are available as a full implementation through Graphcore’s network of reseller and OEM partners.

Alternatively, customers may directly implement the Bow Pod₂₅₆ system with the help of the Bow Pod₂₅₆ build and test guide, available from the Graphcore document website.

Fig. 2.2 Bow Pod₂₅₆ reference design rack

2.2. Communication for scale-out: 3D IPU-Fabric with GCL

The Bow Pod₂₅₆ reference design builds on the innovative IPU-Fabric, designed to support massive scale out. Fig. 2.3 below shows, on the left, an abstracted view of a Bow-2000 with the IPU-Fabric interconnects comprising IPU-Links™, GW-Links (for jitter-free IPU-to-IPU connectivity), and the Host-Link 100Gbps RDMA connection between the host server and each Bow-2000. The small insert on the right shows how these interconnects are used as part of the Bow Pod₆₄ scale-out: IPU-Links join IPU processors together both within Bow-2000s as well as between them; GW-Links connect between the IPU-Gateway chips in each Bow-2000. The IPU-Link connections in the Bow Pod₆₄ form a 2D torus since the loops are closed top and bottom.

The Graphcore Communication Library (GCL) manages the communication and synchronization between IPUs across any IPU-Fabric, supporting ML at scale.

Fig. 2.3 IPU-Fabric

2.3. Host-Links and GW-Links

Host servers are disaggregated from the Bow-2000 PU-Machines with Host-Links – Graphcore´s low-latency high throughout host-IPU RDMA transport using RoCEv2. The 100 Gbps Ethernet links from the RoCEv2 NICs in the Bow-2000 IPU-Machines connect through switches to the disaggregated host servers. This disaggregated host architecture for the Bow Pod₂₅₆ enables user-defined host:IPU ratios and allows for scalable host server utilisation depending on the machine intelligence workload. The GW-Links are used to connect multiple Bow Pod₆₄ logical racks either directly or through an optional Ethernet switching fabric as shown in Fig. 2.4.

Fig. 2.4 Host-Links and GW-Links

2.4. Software

Bow Pod systems are fully supported by Graphcore’s Poplar® software development environment, providing a complete and mature platform for ML development and deployment. Standard ML frameworks including TensorFlow, Keras, ONNX, Halo, PaddlePaddle, HuggingFace, PyTorch and PyTorch Lightning are fully supported along with access to PopLibs through our Poplar C++ API. Note that PopLibs, PopART and TensorFlow are available as open source in the Graphcore GitHub repo https://github.com/graphcore. PopTorch provides a simple wrapper around PyTorch programs to enable the programs to run seamlessly on IPUs. The Poplar SDK also includes the PopVision™ visualisation and analysis tools which provide performance monitoring for IPUs - the graphical analysis enables detailed inspection of all processing activities.

In addition to these Poplar development tools, the Bow Pod₆₄ is enabled with software support for industry standard converged infrastructure management tools including OpenBMC, Redfish, Docker containers, and orchestration with Slurm and Kubernetes.

Fig. 2.5 Bow Pod₆₄ software

Table 2.1 Poplar SDK

Complete end-to-end software stack for developing, deploying and monitoring AI model training jobs as well as inference applications on the Graphcore IPU
ML frameworks	TensorFlow, Keras, PyTorch, Pytorch Lightning, HuggingFace, PaddlePaddle, Halo, and ONNX
Deployment options	Bare metal (Linux), VM (HyperV), containers (Docker)
Host-Links	RDMA based disaggregation between a host and IPU over 100Gbps RoCEv2 NIC, using the IPU over Fabric (IPUoF) protocol
	Host-to-IPU ratios supported: 1:16 up to 1:64
Graphcore Communication Library (GCL)	IPU-optimized communication and collective library integrated with the Poplar SDK stack
	Support all-reduce (sum,max), all-gather, reduce, broadcast
	Scale at near linear performance to 64k IPUs
PopVision	Visualization and analysis tools

To see a full list of supported OS, VM and container options go to the Graphcore support portal https://www.graphcore.ai/support

Table 2.2 Graphcore Virtual IPU SW

IPU-Fabric topology discovery and validation
Provisioning	REST API and SSH/CLI for IPU allocation/de-allocation into isolated domains (vPods)
	Plug-ins for SLURM and Kubernetes (K8)
Resource monitoring	REST API and SSH/CLI for accessing the Bow-2000 monitoring service
	Prometheus node exporter and Grafana (visualization) support

Table 2.3 Lights out management

Baseboard Management Controller (OpenBMC)

Dual-image firmware with local rollback support

Console support, CLI/SSH based

Serial-over-Lan and Redfish REST API

2.5. Technical specifications

Table 2.4 Graphcore Bow Pod₂₅₆ hardware

IPU-Machines	64x Bow-2000 blades
IPUs	256x Bow IPU processors (4 in each Bow-2000)
IPU-Cores™	376,832
Worker threads	2.26 million
AI compute	89.234 petaFLOPS AI (FP16.16) compute
	22.309 petaFLOPS FP32 compute
Memory	Up to 8422.4 GB (includes 230.4 GB In-Processor-Memory (64x 3.6 GB per Bow-2000) and 8192 GB Streaming Memory (64x 64 GB DIMM x2 per Bow-2000)

Table 2.5 Other Bow Pod₂₅₆ hardware

For each Bow Pod64 rack:
Bow Pod64 host server(s)	Default: 1x Dell PowerEdge R6525 server
	Options: 1 – 4 Graphcore approved server/OS options, contact Graphcore sales or your channel partner for details
Bow Pod64 default switches	1x Arista DCS-7060CX-32S-F (100 GbE ToR switch)
	1x Arista DCS-7010T-48-F (1 GbE management switch)
	1x Arista DCS-7060PX4-32-F (GW-Link switch)

Table 2.6 Bow Pod₂₅₆ IPU-Fabric

IPU-Links	128 Tbps (256 x16 Gen4) aggregated bi-directional bandwidth for direct, 2D torus within-rack Bow Pod256 IPU connectivity
	64 Bow-2000s directly connected
	512 standard OSFP ports with DAC cabling
GW-Links	25.6 TBps (128 x 100 Gbps) for between-rack Bow Pod256 connectivity
	Up to 1024 Bow Pod64 (direct) or 256 Bow Pod64 (switched) can be connected
	Standard 100 Gbps QSFP28 ports supporting industry standard transceivers (100G-DR) and DAC cabling

Table 2.7 Bow Pod₂₅₆ connectivity

For each Bow Pod64 rack:
Bow Pod64 server(s) to Bow-2000 connectivity	1 (default) to 4 host servers have connectivity to the 16 Bow-2000s via the Bow Pod64 100 GbE ToR switch (Arista DCS-7060CX-32S-F)
	1x 100 GbE port per Bow-2000 to connect to the ToR switch
	Dual (2x) 100 GbE ports per host server to connect to the ToR switch
Bow Pod64 internal management network connectivity	Aggregated in the Arista DCS-7010T-48-F 1GbE management switch are:
	2x 1 GbE RJ45 management ports from each of the 16 Bow-2000s
	Server management port(s)
	PDU monitoring port

Table 2.8 Bow Pod₂₅₆ thermal characteristics

Air cooled	Built-in N+1 hot-plug fan cooling system in each of the individual components (Bow-2000s, servers and switches)
Rack airflow	All Bow Pod256 components (Bow-2000 IPU-Machines, server(s) and switches) are mounted for airflow direction front of rack (single door, cold aisle side) to back of rack (split door, hot aisle side)
Airflow rate	103 CFM (measured) per Bow-2000 (6592 CFM total in Bow Pod64)

Table 2.9 Bow Pod₂₅₆ rack

For each Bow Pod64 rack:
Rack	42U - 600 mm (W) X 1200 mm (D) x 1991 mm (H)
Weight	450 kg (943 lbs)
PDU	PDU implementation can be customized for target workload and rack power density goals. Please contact Graphcore sales for any help required specifying the PDU implementation
Input power (Vac)	200 - 240 V
Input power (Vdc)	240-310 V for GC-ADA2-30W and GC-ADA2-3EW models
Power cap	1700 W with programmable power cap
Redundancy	1+1 redundancy (with power cap set to 1500W)
Power (nominal)	19 kW

For information on Bow Pod₂₅₆ integration with datacentre infrastructure, please contact Graphcore sales.

2.6. Environmental characteristics

Table 2.10 Environmental characteristics for the Bow Pod₂₅₆

Operating temperature and humidity (inlet air)	10-32° C (50 to 90° F) at 20%-80% RH (*)
Operating altitude	0 to 3,048 m (0-10,000ft) (**)

• (*) Altitude less than 900 m/3000 ft and non-condensing environment

• (**) Max. ambient temperature is de-rated by 1° C per 300 m above 900 m

For power caps higher than 1700W per Bow-2000 please contact Graphcore sales for environmental guidance.

2.7. Standards compliance for Bow-2000 IPU-Machines

Table 2.11 Standards compliance

EMC standards	Emissions: FCC CFR 47, ICES-003, EN55032, EN61000-3-2, EN61000-3-3, VCCI 32-1
	Immunity: EN55035, EN61000-4-2, EN61000-4-3, EN61000-4-4, EN61000-4-5, EN61000-4-6, EN61000-4-8, EN61000-4-11
Safety standards	IEC62368-1 2nd Edition, IEC60950-1, UL62368-1 2nd Edition
Certifications	North America (FCC, UL), Europe (CE), UK (UKCA), Australia (RCM), Taiwan (BSMI), Japan (VCCI)
	South Korea (KC), China (CQC)
	CB-62368, CB-60950
Environmental standards	EU 2011/65/EU RoHS Directive, XVII REACH 1907/2006, 2012/19/EU WEEE Directive

The European Directive 2012/19/EU on Waste Electrical and Electronic Equipment (WEEE) states that these appliances should not be disposed of as part of the routine solid urban waste cycle, but collected separately in order to optimise the recovery and recycling flow of the materials they contain, while also preventing potential damage to human health and the environment arising from the presence of potentially hazardous substances.

The crossed-out bin symbol is printed on all products as a reminder, and must not be disposed of with your other household waste.

Owners of electrical and electronic equipment (EEE) should contact their local government agencies to identify local WEEE collection and treatment systems for the environmental recycling and /or disposal of their end of life computer products. For more information on proper disposal of these devices, refer to the public utility service.

2.8. Ordering information

Bow Pod systems are available to order from Graphcore channel partners – see https://www.graphcore.ai/partners for details of your nearest Graphcore partner.