Sponsored Mention AI hardware and you might immediately envision a data centre with racks of specialised servers, with CPUs packing tens of cores that are tuned to handle AI and machine learning, supplemented with dedicated silicon such as GPUs to handle specific tasks. You’d also imagine acres of memory and hectares of storage, with a superfast interconnect holding it all together.
This may all seem a far cry from what you have running in your laptop.
Except that it really isn’t. Intel’s client processors have been following a similar path, with the results being spectacularly evident in Tiger Lake, its 11th Gen Intel®Core architecture, released this year and coming to the Intel vPro® platform for B2B clients in early 2021.
Tiger Lake is built on Intel’s 10nm SuperFin process , offering clock speeds of up to 4.8GHz and up to 12MB of cache, and delivering a 20 per cent plus improvement in CPU performance (check out Intel benchmarks here).It also includes integrated UHD graphics with up to 96 execution units and support for Intel® Optane™ SSDs offering much faster storage, along with integrated Wi-Fi 6 and Thunderbolt™ 4.0 connectivity.
But the platform also builds on the AI integrations found in Ice Lake, Intel’s 10th Gen Core™. These include Intel’s Deep Learning Boost on the CPU itself, while the GPU features Intel® Gaussian Neural Accelerator 2.0. The platform also includes a neural network accelerator that handles noise reduction.
Overall, Intel says Tiger Lake delivers five times better AI performance than its predecessors. But it still might be difficult to grasp exactly what impact the inclusion of this level of AI acceleration has on day to day computer use.
One application of AI on the CPU is frequency tuning, which using machine learning heuristics to figure out the optimal frequency depending on the workload, and to work out, overtime, how to extend the thermal budget. This prevents the scorched thigh syndrome associated with traditional all or nothing turbo boost, but also extends battery life.
Now I see it
Another application which is particularly relevant right now is video conferencing. The three on board AI accelerators combine to do video cleanup and noise filtering, to deliver what Intel’s Client Compute Group UK director Jeff Kilford describes as “an amazing video conferencing experience”.
Similarly, when Tiger Lake comes to Intel’s vPro enterprise client platform, the GPU will take on a central role in accelerated memory scanning.
“I think what is unique to Intel, is our recognition of the fact that the CPU as critical as it is in the platform, it's just one component that can run workloads,” says Kilford. “And what we're doing is now changing the way in which we orchestrate those workloads, in conjunction with the software ecosystem, the operating system vendors, security, companies, and so on.”
This is not just a question of building in algorithms into hardware. For example, in the case of the memory scanning security feature on Intel’s vPro enterprise laptop platform, the system is learning itself.
“It's clever at understanding what's normal for memory behaviour,” says Kilford. “And it can legislate for what is abnormal. And so they are learning at the edge.”
In a sense, Tiger Lake is building in performance for a known unknown. Intel says it wants the platform to be able to support the workloads of 2023, but as Kilford says, it’s hard to say precisely what these workloads will be. What he can say with confidence is that machine learning and AI will be driving many of them.
That’s because of where the software ecosystem is heading, both within the vendor community and in enterprises. New entrants to the workforce have already been steeped in machine learning and AI at university and are itching to apply it in the workplace.
This will result in “a natural transition over time to a new generation, who are prepared for this and find it very normal,” says Kilford, while “Tiger Lake is really where we build a completely radically different AI friendly platform and then release it into the enterprise.”
Intel is working with operating system and application vendors on how they can best exploit these technologies, and Kilford says they are “all in” on the promise of AI.
“We tested it with (predecessor) Ice Lake, and it was really well adopted. And now we're doubling down with Tiger Lake,” says Kilford. “We couldn't possibly imagine the use cases we're going to see over the next two years, because we're going to be totally surprised by the ingenuity of the software community in leveraging this platform. But we're excited to see it.”
Kilford says Tiger Lake represents such a massive leap in machine acceleration in one generation that it really does “switch on a bunch of brainwaves” across the ecosystem.
Is that a chatty workload? Or the sound of unhappy users?
“This is the biggest shift in decades. We're seeing every pretty much every single software vendor or the game companies, the security companies, the operating system vendors, beginning to reorient their stuff for the prevalence of AI at the edge.”
Adobe, for example, has already leveraged the built in AI capabilities in Ice Lake to speed up photo treatments in PhotoShop Elements, while other imaging developers have used them to speed up colourization of black and white images or for upscaling images.
On the vPro platform, when Microsoft’s Windows Defender finds the appropriate version, it will leverage accelerated memory scanning on the GPU. This moves what would otherwise be “a very chatty” workload away from the CPU where it could both degrade the overall user experience and soak up battery power if the machine is unplugged.
But it won’t be just off the shelf software providers that take advantage of the AI integrations. Kilford said IDC research for Intel showed that 29 per cent of software engineers polled had been working on machine learning solutions on the client over the last 12 months, while Universities are increasing their focus on AI at the edge. He said there was also data that showed 20 per cent of workloads that are currently done in the data centre will migrate to the edge. And the interfaces of 2023 are expected to be far more “natural” because of the AI on board.
“In another couple of years, we'll have so much power on the lap or the desk for machine learning,” says Kilford, “It'll be down to the data scientist or the content creator [to ask themselves] do you want to do it at the lab, and then push the rest of your workload to the cloud for the night-time.”
He adds, “Some workloads will never run on a 15-watt device. But a lot of the ones that are in the data centre will migrate out and it will be some true learning as well. We’re just at the start of that.”
There is another, more prosaic challenge that organisations face. How to quantify system performance when it comes to preparing for these new workloads.
Tiger Lake sees workloads federated across the entire system on a chip, simultaneously reducing load on the CPU, while delivering a richer, more powerful performance overall, and extending battery life.
At the same time, how do you account for the fact that the AI engines embedded in the system will be learning, becoming more efficient in response to the particular stresses or use cases that individual users place on them?
Kilford argues that as Tiger Lake and its successors move into the enterprise it will be the experience of what they can achieve that matters, not artificial benchmarks that focus on the CPU while ignoring what is happening - or not happening - in other parts of the system.
With companies buying laptops with an expected lifecycle of three and half years, they should not be setting their baseline at what they think is a reasonable workload today, or what they think is the “the modern threat”.
Rather they need to be thinking about 2022 and beyond: “If all of this starts to be re-architected to leverage machine learning acceleration on the client, will the device I'm buying today run it at all? And if it does run it, will it run it in a way that has a horrible experience?”
The migration of AI to the edge is a “seismic shift” says Kilford. But if you inflict the wrong hardware on your users, the only tremors you’re likely to feel are the sound of your best and brightest heading out the door.
Sponsored by Intel vPro®