About our AI Infrastructure

What is AI Infrastructure?
Server specifications and performance
• Bow-2000 specifications
• M2000 specifications
• Performance tests: Graphcore M2000 vs NVIDIA DGX A100
Tools our AI Infrastructure supports
Deployment time
How to use AI Infrastructure?
How is AI Infrastructure billed?

What is AI Infrastructure?

AI Infrastructure is a cluster of Graphcore servers which are designed for ML tasks. These servers have high specifications and are configured for processing a great amount of data in a short time (you can find the results of performance tests in the Performance tests section).

This infrastructure consists of three entities:

Poplar server — a server that manages all the other servers in the cluster. You have full access to this server via SSH and can work with it directly for managing the AI Infrastructure and running your model.
M2000 or Bow-2000 server (different types are available in different regions) — a Graphcore server which is used for calculations made while training your model. You don’t have access to it, this server receives commands from the Poplar server.
vIPU controller (virtual Intelligence Processing Unit) — a service which configures M2000/Bow-2000 servers of your AI Infrastructure to make them a cluster. It is involved while the cluster is being created and while you’re changing its configuration, e.g. resizing partitions. You have access to vIPU controller via API and can rebuild the cluster if desired.

For datasets storage, you can use Poplar server disk space, external S3 storage, or our S3 storage.

Server specifications and performance

We provide two types of Graphcore servers: M2000 and Bow-2000. M2000 is a second-generation machine and Bow-2000 is a third-generation one.

Bow-2000 specifications

IPU processors	4x Bow IPU processors (IPU frequency 1.85 GHz) 5,888 IPU-Cores™ with independent code execution on 35,328 worker threads
AI compute	1.394 petaFLOPS AI (FP16.16) compute 0.349 petaFLOPS FP32 compute
Memory	Up to ~260 GB memory (3.6 GB In-Processor Memory™ plus up to 256 GB Streaming Memory™) 261 TB/s memory bandwidth
Streaming Memory	2x DDR4-2400 DIMM DRAM Options: 2x 64 GB (default SKU in Bow-2000 Founder’s Edition) or 2x 128 GB (contact sales)
IPU-Gateway	1x IPU-Gateway chip with integrated Arm Cortex quad-core A-series SoC
Internal SSD	RoCEv2 NIC (1 PCIe G4 x16 FH¾L slot) Standard QSFP ports
Mechanical	1U 19 inch chassis (Open Compute compliant) 40 mm (width) x 728 mm (depth) x 1U (height) Weight: 16.395 kg (36.14 lbs)
Lights-out management	OpenBMC AST2520

M2000 specifications

IPU processors	4 Colossus GC200 IPU processors (IPU frequency 1.325GHz) 5,888 IPU-Cores™ with independent code execution on 35,328 worker threads
AI compute	1 petaFLOPS AI compute 0.25 petaFLOPS FP32 compute
IPU-Fabric	8x IPU-Links supporting 2Tbps bi-directional bandwidth 8x OSFP ports Switch-less scalability Up-to 8 M2000s in directly connected stacked systems Up-to 16 M2000s in IPU-POD systems 2x IPU-GW-Links (IPU-Link extension over 100GbE) 2 QSFP28 ports Switch or Switch-less scalability supporting 400Gbp bi-directional bandwidth Up-to 1024 IPU-M2000s connected
IPU-Gateway	1 IPU-Gateway with integrated Arm Cortex quad-core A-series SoC
Streaming Memory	2 DDR4-2400 DIMM DRAM Options: 2x 64GB (default SKU in IPU-M2000 Founder’s Edition) or 2x 128GB or 2x 256GB (contact sales)
Internal SSD	32GB eMMC 1TB M.2 SSD
Mechanical	1U 19inch chassis (Open Compute compliant) 440mm (width) x 728mm (depth) x 1U (height) Weight: 16.395kg (36.14lbs)
Lights-out management	OpenBMC AST2520 2x1GbE RJ45 management ports

Performance tests: Graphcore M2000 vs NVIDIA DGX A100

Here are results of processing of two popular computer vision models on M2000 and NVIDIA DGX A100 (another popular solution for machine learning).

You can find results of a higher number of comparative tests (12 in total) in the article Graphcore Sets New Ai Performance Standards With MK2 IPU System.

Tools our AI Infrastructure supports

Tool class	List of tools	Explanation
Framework	TensorFlow Keras PyTorch Paddle Paddle ONNX Hugging Face	Your model is supposed to use one of these frameworks for correct work
Data platforms	PostgreSQL Hadoop Spark Vertika	You can set up a connection between our cluster and your data platforms of these types to make them work together
Programming languages	JavaScript R Swift Python	Your model is supposed to be written on one of these languages for correct work
Resources for receiving and processing data	Storm Spark Kafka PySpark MS SQL Oracle MongoDB	You can set up a connection between our cluster and your resources of these types to make them work together
Exploration and visualization tools	Seaborn Matplotlib TensorBoard	You can connect our cluster to these tools to visualize your model

Deployment time

Deployment time is about 15 minutes. This is the time between the moment you click “Create cluster” and the moment it is created and ready to work.

How to use AI Infrastructure?

Create a cluster in the G-Core control panel.
Access a Poplar server via SSH.
Upload a clone of the repository with your model to the Poplar server.
Upload datasets to the Poplar server or connect the S3 storage that hosts your datasets.
Run the model.

How is AI Infrastructure billed?

The billing is per minute. You pay for the time spent from a cluster creation to its deletion.

Was this article helpful?