Adobe is Considering Whether it Wants To Design Its Own Chips (axios.com)

A growing number of technology companies are trying to manufacture their own chips, cutting their reliance on Intel and other chip providers. This week Adobe pondered making a similar move. From a report: At an internal innovation conference on Tuesday, Adobe CTO Abhay Parasnis posed the matter as a question for his colleagues, noting the significant increases in performance from chips designed specifically for specialized tasks, like machine learning. "Do we need to become an ARM licensee?" he said, referring to the company whose underlying chip design is used across a wide range of devices, including computers, servers and phones.

"I don't have the answer, but it is something we are going to have to pay attention to." Later on Tuesday, Parasnis told Axios that there are a range of ways that Adobe could get deeper into silicon. "ARM does afford a model for a software company to package its technology much closer to silicon," he said, adding Adobe could do that without literally making its own chips, including by partnering with an existing chipmaker.

Comments

Adobe?

[jmccue]

hummmm, from the people who brought us flash. I think I will stick with Intel, thank you very much.

Dubyah Tee Eff?

[Crypto Gnome]

Seriously, WHAT FOR?

The entire "lets design the silicon ourselves" push is because YOU'RE ALREADY USING SILICON, just paying someone else for 100% of the work, and the design is generic not customized for your use-case.

If you're a company which has NO HARDWARE PRODUCTS (not even rumors on the horizons) thinking "hey maybe we should license ARM, it worked for Apple" is the WORST KIND OF CORPORATE DRUG INDUCED NIGHTMARE.

Re:Dubyah Tee Eff?

[Sarten-X]

For just about anybody else, I'd agree with you... For Adobe, though, it kinda makes sense.

Adobe's cash cow is the media industry, and one of their biggest performance bottlenecks is video rendering. While not a particularly large market, having a premium hardware product that improves rendering speed is worth quite a lot of money to certain companies. I expect that's what Adobe is looking to capture with this push, with a model that would look very similar to how Bitcoin miners operated: Plug in an ASIC as a coprocessor, and it will handle the application workload.

Re:Dubyah Tee Eff?

[Sarten-X]

We're talking video rendering, which is almost entirely unrelated to the decoding process that's so fast (and already supported by custom silicon like inside that Roku player you mentioned).

As an AC said:

Once an editor is finished editing a film, they have to render it. This process [stitches] all the edits and effects together into one video file.

Now, note that those edits include computing special effects (like chroma key), compositing layers on top of other layers, as well as arranging different clips into one big video, then the whole result must be encoded. Typically, the video codecs are asymmetric, doing a lot more processing during the encoding step so the decoding can be faster and easier (and therefore supporting higher framerates with cheaper decoding hardware).

4K video, in 24-bit color and uncompressed (which is really necessary to do the full compositing operation) is about 25 megabytes per frame. At 60 FPS, that's 1.5 gigabytes per second, or 12 Gbps, to use typical bandwidth units. In comparison, that will just about fully saturate a PCI-e x16 slot and some of the lower DDR4 specs. That's okay, because you won't be storing that data in memory for very long anyway... 64 GB of RAM will only store 42 seconds of uncompressed video. During the encoding process, you'll want to have that old video accessible, because it's useful for making more efficient compression of future frames.

That's a lot of data, all to get a seamless composition, which is really rather important for having modern CGI effects blend invisibly into the recorded footage. Without the full rendering process, the effect layers may get different handling, so they'll appear noticeably different in the final render. In the effort to produce uncompromising results for you, the viewer, studios just take longer for rendering, spending more money on salaries so you get a better result... or they just cut corners and render at a lower resolution.

Having custom devices (and custom silicon) would mean that Adobe (or another vendor) would be able to take advantage of things like dedicated GDDR5X memory for high-bandwidth (256Gbps per chip, and lots of chips to increase capacity) storage, ARM processors for processing (though not necessarily rendering (in the non-video usually-3D sense)) special effects, and ASICs for the compositing and encoding operations, only relying on the host computer for storing the final product. In theory, a shoebox-sized peripheral could replace a data center render farm, enabling near-real-time rendering of edited film. That means directors and production crews can see their results more quickly, allowing them more time to reshoot or otherwise make a better product.

It's certainly a commercial gamble for Adobe... but like I said, they're one of very few companies with a market position that makes custom hardware sensible.

RISC V

[backslashdot]

Why not RISC V, do companies really want to pay ARM forever? It's like, do you want to keep paying for a cloud subscription to software? Hmm.. I guess in Adobe's case they are cool with it.