The wait is almost over. Intel’s highly anticipated response to Qualcomm and AMD, the Intel Lunar Lake based laptops are all set to launch on September 3rd at IFA Berlin. Ahead of the official unveiling, we had the opportunity to sit down with Robert Hallock, VP and General Manager, Client AI and Technical Marketing at Intel, to get an exclusive sneak peek at what’s in store. With Lunar Lake, Intel is making a bold bet on the future of computing, promising significant advancements in AI, graphics, and battery life. In our conversation, Robert shared his insights on Intel’s strategy to take the lead in the competitive computing landscape, and addressed the broader industry landscape, including the X86 vs ARM debate and the company’s approach to increasing competition. Here’s the excerpt from our hour-long conversation.
What can we expect from Intel’s Lunar Lake? We saw Meteor Lake back in January this year, which was the first Intel platform to introduce the NPU story. Now, we’re ready with Lunar Lake, and the question is, how big will it be?
I’ll relate the Lunar Lake launch to Meteor Lake as a reference point. On graphics, you can expect a 50% increase in performance, and we’ve reduced SoC power by about 40%, which can vary depending on usage. AI is, of course, the talk of the hour for everyone. Depending on the engine, NPU performance is about 4 times faster, and graphics performance is roughly the same from an AI processing point of view. It’s a monolithic product with two dies – one for the chipset and one for compute. This makes it look like a more conventional Intel CPU. We’ve overhauled cores, graphics, and the AI engine to launch a much more efficient chip. When we launch in September, we’ll have leadership in battery life, graphics, and AI.
On graphics, you can expect a 50% increase in performance, and we’ve reduced SoC power by about 40%
Lunar Lake is doing a few things differently. We’re looking at 8 cores across the lineup, and Intel has dropped hyperthreading, which is big news since we haven’t seen Intel drop hyperthreading in over a decade. What’s the approach here?
It goes back to when CPUs had fewer cores, and getting another core to increase multithreading performance was physically impossible due to transistor size and wafer size limitations. That’s when SMT stepped in to recognize that execution pipelines in CPUs have bubbles or pauses where new information can be inserted. However, duplicating transistors at the front end of the chip costs power and increases costs. A processing thread served by a physical core is always faster than one served by a logical SMT core. With the architectural changes we’ve made in the Skymont E-cores and Lion Cove P-cores, we’ve improved performance per watt by removing SMT. So, it’s a net benefit overall across performance and power to remove it on a product like Lunar Lake. However, SMT will probably return in other places across the roadmap.
Also read: How Intel Lunar Lake chip aims to power new age of AI PCs at scale
With the architectural changes we’ve made in the Skymont E-cores and Lion Cove P-cores, we’ve improved performance per watt by removing SMT
Another big change with Lunar Lake is the introduction of Memory on Package (MOP). What are the trade-offs? It will help with performance, lower latency, and improved efficiency, but it also makes laptops non-upgradeable when it comes to memory. Will that be a big factor?
I agree it’s a nuanced issue for enthusiasts like me who love upgradable laptops. However, you could potentially add hours of battery life to the system by not having upgradeable memory. The power costs of talking to off-chip memory are one of the larger power costs outside of running the core. Moving memory on-package was the right call for maximizing energy efficiency and battery life, which was the overarching goal of the program. However, it might not necessarily be the choice we make for upcoming products, as newer technologies will allow us to make the same power savings elsewhere. So, MOP is a choice for Lunar Lake but not necessarily for our upcoming products.
So, MOP is a choice for Lunar Lake but not necessarily for our upcoming products.
Speaking about power consumption, efficiency, and battery life, when it comes to AI performance, TOPS seems to be the new battleground, and the latest metric, as per Qualcomm, seems to be TOPS per watt measurement. What are your thoughts on TOPS as a benchmark for measuring AI performance?
I think TOPS are good and bad. They allow for an objective and factual discussion about the theoretical peak performance of the AI engine, just like GPUs have teraflops. However, we don’t necessarily talk about GPU teraflops anymore because people realized that teraflops are not the most accurate metric. History is littered with examples of GPUs with lower teraflops beating those with higher teraflops in real-world performance due to better optimizations, drivers, and software. I believe AI is ultimately headed in the same direction. The software stack for AI is very similar to gaming graphics, with AI models instead of game engines, AI-based applications instead of games, and runtimes and APIs similar to DirectX. So, I believe that over the coming months, the benchmark ecosystem will mature to a point where we can discuss performance pros and cons of AI engines and decide from there. What we like to talk about internally at Intel is performance per TOP. I have some recent measurements on this, where we took the flagship AMD Hawk Point, the Ryzen 7 8840u, vs the Meteor Lake Core Ultra 7 155H. The Ryzen 7 has 25% more TOPS than Intel Core Ultra 7, but on average, across 50 different workloads tested, we saw Meteor Lake coming on top with an average 13% better performance. For me, that’s where the conversation around TOPS starts to break down. On paper, Hawk Point should be better, but because Intel’s software optimization is better, we can produce more performance per TOP in the engine. As far as performance per watt is concerned, of course, it matters and it matters across the board, be it CPU or GPU. However, the advantage we have is having the largest network of software developers, and we talk to them formally. We engage with them to know their long-term plans on what AI software they are working on and the engines they want to use. And, only 30% want to use the NPU, the other 35% want to use graphics, and the rest still want to use the CPU. The reality is, the software developers in the industry are not overly focused on the NPU. Which is why Lunar Lake has these XMX instructions to run AI workloads on the GPU. Our largest content creation accounts will only use graphics as that offers the best performance. So, it’s not that NPU is unimportant, but at the moment, it’s about just 1/3rd of what we need to care about when it comes to AI compute. And that’s where we disagree very much with Qualcomm. By over-indexing on the NPU, you might be under-serving the performance of the workloads that will run on the GPU and CPU engines. And it will ultimately lead to poor outcomes for users. And that’s why we’ve put strong AI on all three accelerators on Lunar Lake.
By over-indexing on the NPU, you might be under-serving the performance of the workloads that will run on the GPU and CPU engines. And it will ultimately lead to poor outcomes for users. And that’s why we’ve put strong AI on all three accelerators on Lunar Lake.
What’s your thought when it comes to ARM-based products entering the Windows Ecosystem, and where do you think X86 has an advantage over ARM-based products?
I think we are about to prove that X86 can certainly have a power advantage over ARM. That is one thing that I am really excited to talk about when we get to the Lunar Lake launch. The real difference maker is the compatibility of X86. Almost every piece of software you have ever touched has X86 at its roots. ARM has this huge hurdle in front of it; it’s got to catch up on 60 years of compatibility. And some of that will probably never move away from the performance cost of emulation. And that’s just the hard truth of trying to overturn 60 years of R&D and coming in at the last minute with something different. Having said that, I do think that Snapdragon is a good offering, the team has done a good job with it. But Lunar Lake is a better solution. Because of the compatibility, the power, and the graphics performance, you get a very well-rounded package that the other guys are still fighting for. And they may never get there, or they will take an extraordinarily long time to get there. You don’t have to wait when you choose the Intel CPU or, for that matter, an AMD CPU. That’s the beauty of X86; it just works well by default.
You don’t have to wait when you choose the Intel CPU or, for that matter, an AMD CPU. That’s the beauty of X86; it just works well by default.
Also read: Intel Lunar Lake AI Performance Preview
We will be there at the launch of Intel Lunar Lake in Berlin, giving you our first impressions. We can’t wait to get our hands on the Lunar Lake laptops from all the major laptop brands. Stay tuned for more coverage on Intel Lunar Lake in the coming week.
Follow Us
Soham Raninga
Soham Raninga is the Chief Editor for Digit.in. A proponent of performance > features. Soham's tryst with tech started way back in Dec 1997, when he almost destroyed his computer, trying to make the Quake II demo run at >30FPS View Full Profile
