Graphcore

Last Call

This investment is about to open. You still can apply but we do not guarantee your participation. If your money arrives too late, it will be returned to your bank account.

Quick Overview

AI-based applications require unprecedented computing power to run. Graphcore builds processors suitable for this task, with applications ranging from medical research and forecasts to robotics and autonomous electronics. The company has issued 67 total patents and 281 total applications and grants. It counts Microsoft, Citadel, Carmot Capital, Oxford University, Simula as clients.

Website

https://www.graphcore.ai/

Products

• Microprocessor for AI and machine learning applications
• Software stack & development tools
• Cloud services for AI and machine learning applications

Total funds raised

$713M

Last round

Dec 2020. Valuation $2.77B

Investors

Sequoia Capital, Foundation Capital, Fidelity International, Schroders, Baillie Gifford

Strategic backers

Business Growth Phase

Product

Graphical processors that are used, say in cell phones, game consoles and PCs, are not designed to accommodate AI-based applications. Graphcore handles this problem by enabling the next breakthrough in machine learning.

The company is developing 3 core types of products:

Hardware
Graphcore has created a new processor, the Intelligence Processing Unit (IPU), with the ability to handle huge amounts of data. The IPU's unique architecture means developers can run current machine learning models orders of magnitude much faster.

 

Software
The company has also developed Poplar, a complete software stack, that is capable of performing synchronous intensive computations. They are required for AI applications that are designed around human thinking and run in parallel processing.

 

Platforms
Graphcloud is a secure cloud service that allows customers to access the power of Graphcore’s Intelligence Processing Unit (IPU) as they scale from experimentation, proof of concept and pilot projects to larger production systems.

Today Graphcore has over 500 employees and offices in Bristol, London, Cambridge, Palo Alto, Oslo, Beijing, Hsinchu, Seoul, New York, Seattle, Austin, Munich, Paris and Tokyo. Microsoft, BMW, Bosch, and Dell are strategic backers of the company.

Technology

According to Pitchbook data, Graphcore has patented technologies in the following fields:

• Coding

• Computer systems based on specific computational models

• Electric digital data processing

• Semiconductor devices

• Transmission of digital information

 

---

The evolution of processors and technology

 

History. Central Processing Unit was first introduced by Intel in 1971 and is designed for scalar processes.

Parallelism. Designed for scalar processes.

Memory Access. Off-chip memory.

 

History. Graphics Processing Unit was first invented by NVIDIA in 1999, and is the most pervasive parallel processor to date. It follows the SIMD (single instruction, multiple data) architecture and is designed for large blocks of dense contiguous data.

Parallelism. SIMD/SIMT architecture. Designed for large blocks of dense contiguous data.

Memory Access. Model and Data spread across off-chip and small cache and shared mem.

 

History. Intelligence Processing Unit was first built by Graphcore in 2016. It is a completely new architecture and is vastly superior to previous types of processors. The 2 main differences between IPUs and CPUs/GPU are in the field of parallelism and memory access.

Parallelism. IPU implements MIMD architecture: multiple instruction, multiple data. This means that all the threads running on the IPUs are independent from each other. This is unlike GPUs that follow the SIMD architecture (single instruction, multiple data). They apply the same operation to large vectors of data, which is much less flexible.

Memory Access. In IPU memory is distributed over the entire processor so that every core on the chip has memory which is local and private to that core. For machine learning that means that the models which need to use many different parameters will run much faster.

Application and Use Cases

There is a strong interest in Graphcore ​from organizations across different areas, such as financial services, healthcare, automotive and internet companies. Graphcore serves over 100 corporate clients, including Microsoft, Carmot Capital, Qwant, University of Bristol, University of Oxford, Simula.

Development

June 2016 – the official foundation of Graphcore with headquarters in Bristol, UK.

July 2018 – Graphcore started to ship its first products to early access customers and generated its first revenue.

November 2019 – IPU products are officially launched, available for purchase from Dell and on Microsoft Azure. Microsoft is among the first customers.

July 2020 – launch of the second generation flagship chip, the GC200, and a new IPU Machine that runs on it, the M2000, with an 8x step-up in performance. The company described it at the time as the first AI computer to achieve a petaflop of processing power “in the size of a pizza box.”

September 2020 – launch of a global network of channel partners.

2021 – expansion in Singapore and Gdansk (to strengthen its R&D capabilities), partnership with Cirrascale Cloud Services to offer the first publicly available Mk2 IPU-POD scale-out cluster.

October 2021 – launch of IPU-POD₁₂₈ and IPU-POD₂₅₆, the latest and the largest products in the ongoing story of scaling Graphcore AI compute systems. These systems are ideal for cloud hyperscalers, national scientific computing labs and enterprise companies with large AI teams in markets like financial services or pharmaceutical.

Market Opportunities

According to Techcrunch, purpose-built AI processors are expected to be significant in the coming years because of increased AI adoption and computing megatrends like cloud technology and 5G.

Major drivers for the market:

• increasingly large and complex datasets driving the need for AI,
• the adoption of AI for improving consumer services and reducing operational costs,
• the growing number of AI applications,
• the improving computing power,
• growing adoption of deep learning and neural networks.

Team

Graphcore was founded by two successful serial entrepreneurs Nigel Toon and Simon Knowles.

Nigel was CEO of two VC-backed silicon companies: Picochip, which was sold to Mindspeed in 2012, and most recently, XMOS, in which Graphcore was incubated for two years before being established as a separate entity. Before that Nigel was co-founder of Icera, a 3G cellular modem chip company that was sold to NVIDIA in 2011 for $435M.

Simon is co-founder, CTO and EVP Engineering of Graphcore and is the original architect of the “Colossus” IPU. He has been designing original processors for emergent workloads for over 30 years, focussing on intelligence since 2012. Before Graphcore, Simon co-founded two other successful processor companies – Element14, acquired by Broadcom in 2000, and Icera, acquired by NVIDIA in 2011.

Financials

Graphcore has raised a total of $713M in funding over 7 rounds. Its last primary round Series E in December 2020 valued the company at $2.77B.

According to Pitchbook, Graphcore’s revenue grew by more than 8 times in 2019 and was $10.1M. Nigel Toon, the founder and CEO of Graphcore, told Techcrunch that 2019 was a transformative year for Graphcore as the company moved from development to a full commercial business with volume production products shipping.

The public market of chip stocks is huge. Graphcore's closest public competitors include NVIDIA, Advanced Micro Devices and Intel. These three companies alone are worth $1T (as of Nov 3, 2021). The recent M&A deals confirm the size of this market: NVIDIA is aiming to acquire ARM for $40B, AMD spent $35B to buy Xilinx and Analog Devices bought Maxim Integrated for $21B.

According to the CEO, the company doesn’t need to raise any more money soon, and “the next step would probably be an IPO of the business at some point, when things are much more predictable in the business.” Over the past years “Graphcore has built a ‘phenomenal’ pipeline of new customers, which helped win over the new investors. The company is ‘very hopeful’ that 2021 is going to be a very strong revenue year.”

According to Markets and Markets, the global artificial intelligence chip market size is projected to reach $40B in sales by 2025. If Graphcore captures 5% of this market, the company's revenue to reach $2B by 2025. As of December 2021, NVIDIA is worth 30x its annual revenue, while AMD is worth 12x its annual revenue. Given these valuation estimates, Graphcore could be worth $30-50B by 2025.

Investments Rounds and Share Price

Investors and Their Most Successful Investments

Sequoia Capital – Freshworks, NVIDIA, and WhatsApp

Ontario Teachers' Pension Plan – Ubisoft, Zalando, and Sea

Foundation Capital – Uber, Netflix, and ForgeRock

Baillie Gifford – Flipkart, Spotify, and Ginkgo Bioworks

Microsoft – Facebook, Flipkart, and Cruise

BMW – ChargePoint, Proterra, and Shift

Dell – Guardicore, DocuSign, and MongoDB

Risks Related to Graphcore

Intense competition

Graphcore describes its IPU as the first processor to be designed specifically for AI. However, a number of other companies including NVIDIA, Intel and AMD have made huge investments into this area and have ramped up their pace of development to meet market demands. Tech giants like Tesla, Amazon, Apple and Google are all getting more serious about their own chips making efforts. And there is also a new generation of startups, including Esperanto Technologies, Cerebras, and SambaNova Systems.

Consolidation on both supply and demand sides

The cost of staying competitive in the chip industry has been rising rapidly thus most of the players are trying to consolidate their efforts. NVIDIA, which is currently the biggest U.S. chip company by market capitalization, is trying to acquire ARM, which owns the most widely-used set of standards, for $40B. The deal is subject to regulatory approvals in China, the European Union, and the U.S. If the acquisition is approved, it can lead to reduced innovation in the chip industry. AMD spent $35B to buy Xilinx, Intel sloughed off its memory unit to SK hynix for $9B and Analog Devices bought Maxim for $21B. Graphcore doesn’t have any plans to make acquisitions and wants to grow organically.

Development and training of AI-based technologies requires massive resources, and to bear these costs there is a consolidation even on the customer side. Customer consolidation increases market supply power and makes AI chip companies dependent on a small number of crucial partners.

Chip shortage

Geopolitical tensions in the semiconductor space have increased in recent months, with the ongoing global chip shortage highlighting the world’s reliance on TSMC and Samsung. Graphcore, alongside industry big names such as Qualcomm and NVIDIA, produces its IPU using technology developed by TSMC. And while chip fabs are struggling to produce older, slower chips for conventional needs, there is also a gap between demand and supply of high-end chips like GPUs. These bottlenecks may slow down the adoption of AI across different industries.

Overvaluation of the market

Investors are very bullish about the future potential of the chip market. Share prices of most chipmakers have increased steadily over the last couple of years. Nonetheless, if the growth of the market will be slower than expected, it can affect the whole chip industry, including the companies that produce processors for AI.

Sources

Graphcore Pitchbook Report, Sep 2021

Graphcore Crunchbase Profile, Nov 2021

Artificial Intelligence (Chipsets) Market, Markets and Markets, Nov 2021

Graphcore website, Nov 2021

Graphcore investor presentation, Oct 2021

“Graphcore raises $150M,” Techcrunch, Feb 2020

“AI chipmaker Graphcore raises $222M,” Techcrunch, Dec 2020

“AI Chip Market is Booming: Top 25 Players in AI Chip Market,” Ubuntupit

Offering

Deal structure

Dizraptor Fund 1001 LLC is formed for the purpose of acquiring Common shares of Graphcore. The interests in the fund are offered to accredited investors who after signing a subscription agreement become its members.

View SPV's documents

Limited Liability Company Agreement

Confidential Private Placement Memorandum

Subscription Agreement & Investor Questionnaire

Fees

Fee on management – 5%, paid on top of the investment amount.

Carried interest – 10% of net profit, charged at the closing of investment, after the rest of the commissions are paid.