As many of you know, Cadence (more correctly, “I”) recently performed an industry survey about HLS (High Level Synthesis) to get a fuller view of the productivity experiences and expectations from users and non-users alike.
With nearly 200 responses, roughly half from HLS users and half not, we got a representative picture of what HLS users, potential users, and even skeptics believe about HLS. So let’s dive in.
How familiar are you with high-level synthesis (HLS)?
This was a good cross-section of high-level synthesis users and non-users, which I was very happy to see. The numbers are high enough that they are likely a decent representation of the industry perceptions.
In the analysis of the following questions, I break down the responses by people who have used HLS (the mustard and light blue sections of the above graph) vs non-users (the next three categories). I excluded the responses from people who answered “not at all,” since they self-identified as not even hearing about HLS before this survey.
The next question, only for HLS users, is about what they have designed with HLS.
What types of hardware have you designed with HLS? (select all that apply)
The first takeaway should be that the old opinion that HLS is only used for datapath types of applications is just that… old. Many years ago, that was true, but not today. “Controllers” or “processors” combine to account for 24% of the design types. Of course, some of the other areas, such as “wired networking,” are likely to include a lot of non-datapath processing, as well.
Compared to the survey I did in 2015, “Image Processing,” the combined “networking” categories, and “encryption” have all decreased as an overall percentage. To be clear, this reflects the diversely growing user base of HLS, not an absolute decrease in these categories. (As a matter of fact, wireless was the fastest growing market segment for Stratus™ HLS in 2016.)
The remainder of this year’s survey focuses on productivity, starting with overall productivity compared to an RTL designer. In the following graphs, red bars are the reported users’ experiences, and blue bars are the reported non-users’ expectations.
On average, how productive is an HLS designer compared to an RTL designer?
Most HLS users (red bars) are seeing a good productivity benefit. The spread in productivity didn’t appear to have any correlation with the types of HW being designed. It’s quite possible it’s related to the learning curve, as productivity tends to increase as familiarity with HLS flow increases. Next year, I’ll be sure to ask, “How long have you been using HLS?”
It was interesting to see that 5% of HLS users are exceeding the standard HLS claim of “up to 10x better productivity.” Perhaps we should increase the claim…?
One disappointing result is the shape of the graph for non-users compared to users. As a group, non-users have lower productivity expectations than what is being realized by industry users. In fact, almost a third believe there is no productivity benefit. I guess that gives me and the rest of the HLS community some homework…
The next question asked about how the productivity gained through behavioral IP reuse.
How much productivity is gained through behavioral IP reuse? “Behavioral IP” is defined as high-level IP created for implementation with high-level synthesis (HLS). Behavioral IP can typically be reused or retargeted by changing some controls on the HLS tool.
Again, most HLS users (red bars) are seeing a good productivity benefit from behavioral IP, and 5% are exceeding even the marketing claims.
Unlike the previous question, the shapes of the graphs of the user experiences and non-user expectations were mostly in line, albeit with a few more at the extreme high and low ends of expectations.
The final productivity question is about verification.
How much more productive is verification in the HLS flow?
Once again, most HLS users (red bars) are seeing a good productivity benefit. As before, there didn’t seem to be any correlators with the spread in productivity. Over 6% of users are exceeding the HLS productivity claims. Interestingly, more non-users seem to believe the HLS productivity benefit when it comes to verification.
At this point, you may be going cross-eyed from all the graphs, so let me summarize.
- The types of applications where HLS is being used has broadened significantly.
- HLS users are getting, and sometimes exceeding, the productivity benefits that we EDA vendors claim.
- Non-HLS users accept the productivity benefit when it comes to verification more readily than design.
I’m sure there are other correlations and data that can be gleaned from the results. Maybe I can get my hands on some of that machine learning IP to sift through the raw data….
I’ll close with one final survey result. This one may be immediately applicable to you, and might even save you some money. The Pursley household recently saw three early summer “blockbuster” movies. Previews suggested each could be the movie of the year, so I did a not-so-anonymous survey to see which movie was the best.
As you can see, Guardians of the Galaxy Vol. 2 was the clear winner with 75% of the respondents saying it was the best. It also got a very rare “maybe 10 out of 10 stars” from our resident movie critic. Wonder Woman was also a fantastic movie, getting 25% of the votes and a “8 out of 10 stars” only because it started a little slow.
Pirates of the Caribbean: Dead Men Tell No Tales was a different experience altogether. I think it got negative stars, but I can’t remember because we were almost running to get out of the theater. Yeah, it was that bad… but your mileage may vary.
For more information about the sequel Dead Men Don't Do HLS...sorry, my brain is still mush...I mean Stratus, see the product page.