High-flying Achronix plans move to ML
Achronix plans to provide increased support for machine learning architectures such as convolutional neural networks and is also planning to offer a new product line called SpeedChip.
Steve Mensor, vice president of marketing, at Achronix disclosed that Achronix was enjoying an “annus mirabilis” in a briefing. When asked how he could be sure of such numbers with the year less than half-way through said: “We have a hard backlog. This is based on what we’ve shipped and on orders. Anything more that comes in the second half will just add to that.”
Achronix quaterly revenue 2Q16 through 2017. Source: Achronix
Achronix was founded in 2004 and took some time to bring its form of FPGA to market eventually gaining support from Intel as a foundry client on the now relatively mature 22nm FinFET process. However, since being able to ship product with I/O tailored to high-speed networking applications, in its Speedster 22i family it has started to achieve some traction, Mensor said.
And in 2016 it went down a path followed by some other FPGA companies by offering to license various cores of its Speedster FPGAs to SoC builders, under the name Speedcore. However, this offering was ported to the popular 16FF+ 16nm FinFET manufacturing process offered by leading foundry TSMC (see Achronix branches out into FPGA IP with TSMC).
Next: Three product types
In 2017 Speedcore will represent about 20 percent of Achronix sales rising to 50 percent by 2020, Mensor said.
$400 million to $500 million annual revenue.
Source: Achronix
Strong sales in networking and data acceleration gives Achronix the opportunity to target other applications or support the current sector in greater depth. “We are going to pivot towards application targeted chips,” said Mensor, adding that support for acceleration of machine learning was a key opportunity.
“Machine learning is running at 66 percent CAGR according to Semico Research. Big data is just growing larger and larger. An automated vehicle going to generate something like 4Tbytes of data per day. We need bigger engines for processing,” said Mensor. “And programmable is good because the architectures are diverse and algorithms will continue to change.
Current engagements for Speedcore include machine learning in data centers, automotive ADAS, 5G wireless infrastructure, network acceleration, SQL and Internet acceleration.
Next: Products, upgrades coming
“Right now Speedster is not optimized for ML. But we are working on our next generation [of Speedster] on a better manufacturing process. There will be more details to come in the second half of 2017,” said Mensor.
Speedcore is also set to jump to from 16FF+ to 7nm FinFET from TSMC. That will be delivered in first half of 2018, Mensor said, but added that it is open to other process nodes, such as 12nm or 16FFC and other foundries, such as Globalfoundries. “We need a lead customer for a port,” said Mensor.
Mensor added that later in 2017 Achronix would launch its Speedchip approach to providing field programmability of hardware.
“This comes in the form of a chiplet for 2.5D integration. It’s something Speedcore customers have asked about because sometimes they want a package integrated solution, rather than a monolithically integrated solution. In some cases adding Speedcore to an SoC increases the die size making it uneconomic. Speedchip is an FPGA but we will customize the interface to customers specific requirements. Speedchip is an ASIC business model,” he said.
Related links and articles:
News articles:
Achronix branches out into FPGA IP with TSMC
22nm FPGA maker takes on Xilinx and Altera
FPGA cores offered for TSMC’s 40ULP process
China’s Gowin expands non-volatile FPGA family
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
