The land grab touched off in the semiconductor world by the rise of artificial intelligence and big data has just produced its second giant corporate deal in a matter of weeks.
With its $35bn all-stock offer for Xilinx, AMD on Tuesday strengthened its claim to being one of the long-term winners in supplying chips for data centres, where much of the computing power required for cloud computing and machine learning is concentrated.
Like Nvidia, which recently agreed to pay at least $33.5bn in cash and stock for chip design company Arm, AMD has just been through a period of strong growth in its existing core market, and is now setting the stage for a broader role in the data centre.
As recently as five years ago, such a turn of events would have seemed improbable. With a stock market value of less than $3bn, AMD had fallen decisively behind its nemesis, Intel, which at the time was broadening its own reach with the purchase of specialised chipmaker Altera.
It is Intel that is now struggling. AMD has hit back with a well-regarded new generation of chips for PCs and servers, and Intel’s shares have fallen 25 per cent since it stunned Wall Street in July with the news of manufacturing delays.
AMD and Xilinx both rely for manufacturing on Taiwan’s TSMC, which now has a clear lead in the most advanced production technology, and combined are now worth almost two-thirds as much as Intel.
In the alphabet soup of chip technologies, the specialised type of chips made by Xilinx and Altera, known as FPGAs, are coming to play a bigger role in data centres. They act as accelerators to speed up the training of neural networks used for deep learning. They can also be reconfigured by customers, making them more adaptable as the demands of particular computing workloads change.
That makes them part of a wave of silicon that is transforming how data centre computing is handled, prompting a scramble for position and a spate of deal making.
56%
AMD’s revenue growth in Q3
“It’s going to take a lot of work and take a lot of years” for all the different technologies to work seamlessly together, said Patrick Moorhead, an analyst at Moor Insights & Strategy. But chip companies without a full range of products to suit different data centre needs are likely to fall behind in the long run, he said.
AMD’s rebound, thanks to its new designs for central processing units, or CPUs, was confirmed with its latest quarterly earnings on Tuesday. Revenue growth of 56 per cent was 14 percentage points ahead of most analysts’ expectations, with sales of data centre chips more than doubling from a year ago.
For its part, Nvidia came from left field to become a force in the data centre, employing its graphics processing units — originally designed to handle video games — to speed up machine learning. The company hopes to consolidate its position by adding CPUs designed by Arm, which has a rival chip architecture to that used by Intel and AMD. Earlier this year, Nvidia also completed the $7bn purchase of Mellanox, whose networking technology is used to shuffle data inside a data centre.
Intel, meanwhile, has been adding to its own armoury, spending $2bn last year to buy Habana, a maker of specialised, single-purpose chips, known as ASICs, which also act as accelerators for machine learning calculations.
AMD has a long way to go to turn its vision into reality. Data centre customers make up only 14 per cent of Xilinx’s sales in the latest quarter, and it is also only a small player in GPUs. News of its Xilinx purchase caused some Wall Street analysts to warn of the dangers of getting distracted at a time when the company should be focusing on trying to build a durable lead over Intel in its core CPU market.
However, AMD can point to advantages as it tries to combine the two companies. There is close technological compatibility in their hardware designs, making it easier to bring their development closer, said Alan Priestley, an analyst at the research firm Gartner. There could also be cultural fit: Victor Peng, Xilinx’s chief executive, once worked at AMD.
In the short term, the main benefit from the deal will come from diversification, said Mr Moorhead, enabling AMD to extend its reach into 5G base stations and the automotive market, where Xilinx gets most of its sales. The drive to diversify has been one of the main forces behind a wave of dealmaking that has transformed the chip industry over the past five years.
But it is the data centre market that Lisa Su, AMD’s chief executive, was focused on as she laid out her plans for the combination with Xilinx. The deal creates the potential to one day mix and match FPGAs and CPUs on a single chip, optimising performance for different workloads.
AMD is not alone in this ambition. Jensen Huang, Nvidia’s chief executive, recently came up with an addition of his own to the alphabet soup of chip acronyms: a DPU, or data processing unit. By combining elements of several other chip types in a single product, this would take integration to a new level.
AMD also pointed to another hoped-for benefit from adding Xilinx. By combining the software tools used to create programs that run on the companies’ chips, it hopes to make life easier for developers. That reflects a wider fight for developer attention among the biggest chip companies.
AMD and Xilinx start with far less of a following among developers than either Nvidia or Intel. But after being all but counted out half a decade ago, AMD has just laid out its latest case for being seen as a central force in the chip industry again.