March 3, 2022, Beijing, China – graphcore ® Today, we officially released a new IPU system product – Bow Series. The bow pod system adopts the world’s first 3D wafer on wafer processor – bow IPU. Bow IPU is the core of the new generation bow pod AI computer system. Compared with the previous generation, it can provide up to 40% performance improvement and 16% power efficiency improvement for real-world AI applications, while the price remains unchanged without changing the existing code. The name of bow comes from a place name in London.

poYBAGIhZ_ aAeBz0AAEFI_ VdNb8627. png

Bow IPU processor

poYBAGIhaAqAL7cEAAK-AOKrZRA686. png

Bow-2000 computing blade

Meanwhile, graphcore will deliver the world’s first super intelligent computer – good in 2024 ™ Computer. Graphcore is developing a new generation of IPU technology for this good ™ Computer provides AI floating-point calculation with more than 10 exa flops; Up to 4pb of storage, with a bandwidth of more than 10PB / S; Support AI models with more than 500 trillion parameters; 3D wafer on wafer logic stack; Get poplar at the same time ® Full SDK support. The computer is expected to cost between $1 million and $150 million (depending on configuration). The reason why this super future intelligent computer is named good is that graphcore is committed to building a “good” computer and pays tribute to Jack good, a pioneer of computer scientists.

Lu Tao, President of graphcore Greater China and global chief revenue officer, said: “With the overall rapid development of AI computing market, customers need computer systems that take into account performance, efficiency and reliability. We release a new generation of IPU precisely to support global technology developers and innovators to make new technological breakthroughs in the field of AI computing and achieve high-quality development. We hope to continue to explore the AI application scenarios of IPU in China, enable efficient AI computing through IPU, and support Chinese innovators to continuously promote the development of machine intelligence The world. “

Excellent performance of bow pod

The flagship product bow pod256 provides AI computing of more than 89 petaflops, while the super large-scale bow pod1024 can provide AI computing of 350 petaflops, supporting machine learning engineers to take a step ahead and make new technological breakthroughs in the field of machine intelligence when the scale of AI model increases exponentially.

Bow pod aims to provide real-world performance on a large scale for various AI applications, from GPT and Bert for natural language processing to efficientnet and RESNET for computer vision, to graph neural networks and many other applications.

In the same peak power range as the MK2 IPU pod system equipped with bow pod system, customers found that the performance of various AI applications improved by up to 40%.

pYYBAGIhaBqAI3g7AAHzhFyooDg103. png

For the most advanced computer vision model efficientnet, the performance of bow pod16 is more than 5 times higher than that of similar NVIDIA DGX A100 system, while the price is only half, so the total cost of ownership is increased by 10 times.

pYYBAGIhaCyAK3mzAAG6w-5dVbU536. png

Compared with the previous generation products, in addition to the 40% performance improvement, the bow pod system can also greatly improve the power efficiency. A series of practical application tests show that the performance per watt of bow pod can be improved by up to 16%.

poYBAGIhaDaAMyvQAAFF3QPP4FQ291. png

Wafer on wafer innovative technology

The bow IPU processor at the core of the bow pod system adopts the world’s first 3D semiconductor technology, which supports the bow pod system to achieve huge performance improvement and higher power efficiency. Wafer on wafer 3D technology is jointly developed by graphcore and TSMC. Wafer on wafer technology has the potential to provide higher bandwidth between silicon wafers, so as to optimize power efficiency and improve the power of colossus architecture at the wafer level.

Wafer on wafer in bow IPU, two wafers are combined to produce a new 3D bare chip. One wafer is used for AI processing, which is architecturally compatible with gc200 IPU processor, has 1472 independent IPU core tiles, can run more than 8800 threads, and has 900mb processor memory, while the second wafer has power supply bare chips.

By adding a deep trench capacitor in the power supply die, which is located next to the processing core and storage, we can supply power more efficiently, so as to realize the AI calculation of 350 teraflops and achieve a 40% performance improvement. Through close cooperation with TSMC, we have fully verified the whole set of technologies, including a number of breakthrough technologies in back silicon via (btsv) and wafer on wafer (wow) hybrid bonding.

As a part of the bow pod system, the latest bow-2000 IPU machine adopts the same powerful system architecture as the second generation ipu-m2000 machine, but now it is equipped with four powerful bow IPU processors, which can provide AI computing of 1.4 petaflops.

Bow-2000 can be fully backward compatible with the existing ipu-pod system, and its high-speed, low delay IPU structure and flexible 1U overall dimensions remain unchanged. Bow-2000 is a basic part of the whole bow pod series. Bow-2000 can be installed on the host servers of leading brands such as Dell, ATOS, supermicro, Inspur and Lenovo to form the bow pod system. The whole bow pod series includes bow pod16 (4 bow-2000 and one host server), bow pod32 (8 bow-2000 and one host server), bow pod64 and larger flagship systems bow pod256 and bow pod1024.

poYBAGIhaEqAOe-fAAHA-O1DSLY277. png

The latest bow pod system is now on the market and is being shipped worldwide.

Leave a Reply

Your email address will not be published. Required fields are marked *