Qualcomm is aware of that if it needs builders to construct and optimize AI purposes throughout its portfolio of silicon, the Snapdragon big must make the expertise less complicated and, ideally, higher than what its rivals have been cooking up within the software program stack division.
That is why on Wednesday the fabless chip designer launched what it is calling the Qualcomm AI Stack, which goals to, amongst different issues, let builders take AI fashions they’ve developed for one machine sort, for example smartphones, and simply adapt them for an additional, like PCs. This stack is just for units powered by Qualcomm’s system-on-chips, be they in laptops, cellphones, automobile leisure, or one thing else.
Whereas Qualcomm is finest identified for its cell Arm-based Snapdragon chips that energy many Android telephones, the chip home is hoping to develop into different markets, reminiscent of private computer systems, the Web of Issues, and automotive. This growth means Qualcomm is competing with the likes of Apple, Intel, Nvidia, AMD, and others, on a a lot bigger battlefield.
Given that every one these gamers agree AI is a key future workload for the whole lot from automobiles and IoT units to PCs and smartphones, Qualcomm is hoping to face out with a unified software program stack, which brings collectively all the corporate’s present instruments, frameworks, runtimes and other forms of software program for growing AI purposes on its households of processors.
Qualcomm thinks all this may cowl AI improvement for the whole lot from telephones and PCs to automobiles and datacenters. Click on to enlarge.
In a briefing with journalists, Qualcomm govt Ziad Asghar mentioned the consolidation of the corporate’s AI software program efforts falls consistent with the way it has been taking its core chip expertise and adapting it to totally different markets with customization.
“So an funding turns into an actual optimistic benefit for us as a result of now the funding that we make in a single space is definitely capable of propagate very strongly into all of the totally different areas,” mentioned Asghar, who’s vice chairman of product administration at Qualcomm Applied sciences.
Opening up the toolkit
With Qualcomm AI Stack, the corporate is now hoping to carry this good thing about utilizing one set of applied sciences for a number of market areas to builders and machine makers on the software program facet.
By slicing down on the quantity of labor required to adapt AI purposes from one sort of Qualcomm chip to a different, the biz is promising to avoid wasting firms cash on costly engineering assets.
“We’re giving the power to our OEM base to have the ability to try this to nice extent with out really having to spend so much extra when it comes to [non-recurring engineering],” Asghar mentioned.
What units Qualcomm AI Stack aside from competing software program stacks, in accordance Asghar, is the Qualcomm AI Mannequin Effectivity Toolkit. One in every of its most vital options is the power to compress a power-hungry AI mannequin skilled within the cloud from 32-bit floating level precision to the 8-bit integer format in order that it will probably run effectively on low-power units.
“As you try this, you’ll be able to carry enormous advantages when it comes to energy consumption, as much as 4 occasions energy consumption enchancment in lots of circumstances in eventualities like this,” he mentioned.
Qualcomm can be banking on its Neural Structure Search device, which lets builders optimize AI fashions utilizing varied constraints, reminiscent of larger accuracy, decrease latency, or decrease energy. This may enable builders to, Asghar mentioned, “get much more accuracy or a mannequin that is ready to provide you with decrease latency or a mannequin that is capable of do the identical process at decrease energy.”
Whereas a lot of the underlying software program within the Qualcomm AI Stack is not new, Asghar known as the unification of the weather a “leap ahead” for builders and machine makers and mentioned it would function the inspiration for domain-specific software program improvement kits when future markets take form.
“What we envision is that now with this very strong footing, which is unified throughout all of the totally different enterprise strains, we are able to begin to write extra of those domain-specific SDKs as a brand new vertical emerges the place we see an agglomeration of our prospects,” he mentioned. ®