Intel is gearing up the FPGA user base it inherited from Altera for the release of Stratix 10 hardware and companion application acceleration stack.
For those who do not mind more power consumption, the new cards will provide quite a jump in memory capacity and bandwidth and double the logic elements found on the Arria 10.
As Intel did with Arria a year ago, the company is pitching the Stratix 10 as a platform rather than just a device in order to appeal to a wider base of applications that need the memory and bandwidth. These areas include in-line processing and memory-intensive applications with focus on video, data analytics, financial, and machine learning. Intel’s director of their PSG group and head of acceleration platforms marketing, Sabrina Gomez, tells The Next Platform that Arria 10 had deep traction in video transcoding, in part because of the processing potential inside a tighter power budget.
We suspect the power budget for the different FPGA acceleration cards will be the determining factor in who makes the leap to Stratix 10. This means, of course, so too will the needs of the application. As we reported earlier this year, the Arria 10 GX 1150 consumes 70 watts, but according to Intel, their Stratix 10 PAC could eat as many as 225 watts. The former seems very low while the latter seems way higher than an FPGA should be, at least until we look at what’s inside.
The Stratix 10 provides quite a bit of a jump in on-chip and additional off-chip memory. On-chip memory is about 4X of its predecessor, which means much higher throughput for the memory intensive analytics and transcoding applications Intel is targeting. The 32GB of DDR4 memory on the board is also important for bandwidth for most of the applications Intel is packaging software for and with double the logic elements, Gomez says the sizes of the workloads can be extended quite dramatically as well.
With over a year of FPGA server products under its belt, Intel is getting a sense of where traction is best. Arria 10 had some of its biggest use cases in image processing, in part because of the image processing IP to accelerate across various image and resolution types. This was a sweet spot for these users because of the performance to power consumption ratio, something we suspect might not extend to Stratix 10. Luckily, Intel is adding the Stratix to the portfolio, taking the best lessons from packaging software into an FPGA card from Arria while leaving those customers with a supported product.
Financial applications were another important area for the Arria 10 for select algorithmic acceleration and backtesting in particular. Here, power might be an important factor, but this is where the added memory and throughput could be a real boost. This, coupled with the accelerated application IP for financial applications which lends itself well to existing math kernel interfaces, could bring the Stratix 10 into this market over Arria 10.
Real-time analytics is another important market for the Stratix 10. Gomez says they had good traction with Arria 10, especially in retail, but that card only had one form factor and was used in 1U server settings. With Stratix, she says they go into analytics spaces that need much higher bandwidth and in servers that can take a 2U card and accept a more power hungry device. In other words, some users will be limited by what they can do by the servers they’re working with and what can fit.
“The hardware is important here but is not the full story. We have a unique feature on this new card with the on-chip Ethernet interface. There’s a NIC, it’s on-board. The data streams in, the analytics can be accelerated by the FPGA in real time, and the software stack, which is Spark based, can be really useful in areas like retail where this capability is in demand,” says Gomez.
Anytime we get on the phone with Intel we tend to get story about how a device is a right fit for deep learning and AI but with the Stratix 10 PAC, this did not take center stage. Gomez says inference is the target for their FPGAs (versus training) but at a maximum of 225 watts under a normal load, that would be a pretty heavy tool for the job. The nuanced side of that is that FPGAs have the advantage of being versatile, so even if users didn’t put their Stratix 10 systems in place for inference specifically, they can change jobs from running analytics to handle some inference.
This one card/many uses scenario is a valuable one, according to Gomez, who points out that their customers have gone through great lengths to qualify a card for their servers—something that takes time and energy. “But once an FPGA is up and running and validated, they can swap out workloads on the same device and run many applications on the same approved hardware.”
With the Stratix 10 available elsewhere on the market in advance of Intel’s early 2019 release date (with Hewlett Packard Enterprise as the first OEM to deliver), Gomez says the real value is not the hardware so much as the entire packaged piece, which she says provides far more software support for specific workloads as well as tools for newcomers to FPGAs to onboard. The company is developing a workload store where new users can test drive applications and both devices to see what is the best fit for their applications.
As always, Intel (and its competitor, Xilinx) have to keep beefing up the abstraction to bring new users into the FPGA fray. Gomez says they continue pushing for reduced complexity of coding and adding more abstraction through the different interface layers, enhancing the drivers and acceleration IP internally and via partners, and working to bring all of this into the early phase of the release cycle. Interestingly, new developments are starting to be upstreamed to the Linux community so that by next year, the major OS vendors can have integrated support for Intel FPGAs.