{"video_id":"fp_kVJo8qZTuS","title":"There’s a new CPU maker (SPONSORED)","channel":"Linus Tech Tips","show":"Linus Tech Tips","published_at":"2026-03-26T17:00:00.051Z","duration_s":437,"segments":[{"start_s":0.0,"end_s":5.2,"text":"If you're watching this right now, you're probably using at least one ARM CPU to do it.","speaker":null,"is_sponsor":0},{"start_s":5.2,"end_s":10.0,"text":"Or, well, not an ARM CPU, because ARM doesn't actually make CPUs, or do they?","speaker":null,"is_sponsor":0},{"start_s":10.8,"end_s":16.96,"text":"That's the big news that they sponsored us down here to their ARM Everywhere event to announce.","speaker":null,"is_sponsor":0},{"start_s":16.96,"end_s":25.76,"text":"Behind me, and in my hand, is the ARM AGI CPU, built for performance, scale, and, as always,","speaker":null,"is_sponsor":0},{"start_s":26.72,"end_s":33.12,"text":"up to 136 ARM Neoverse V3 cores with two megabytes of level 2 cache each,","speaker":null,"is_sponsor":0},{"start_s":33.12,"end_s":41.12,"text":"built on TSMC's 3nm process node, and it can run at up to 3.6GHz, which, right out of the gate.","speaker":null,"is_sponsor":0},{"start_s":41.12,"end_s":47.76,"text":"This is some questions, doesn't it? Just 3.6GHz? Can it, like, dynamically boost a single core","speaker":null,"is_sponsor":0},{"start_s":47.76,"end_s":55.12,"text":"way higher or something? No. And, according to ARM, that's actually a key feature, not a bug.","speaker":null,"is_sponsor":0},{"start_s":55.12,"end_s":60.72,"text":"By eschewing SMT multithreading and the highly variable power consumption that's associated","speaker":null,"is_sponsor":0},{"start_s":60.72,"end_s":66.72,"text":"with constantly fluctuating clock speeds, not to mention designing a 12-channel DDR5 memory controller","speaker":null,"is_sponsor":0},{"start_s":66.72,"end_s":72.8,"text":"that can feed every individual core with a consistent 6GB per second of bandwidth,","speaker":null,"is_sponsor":0},{"start_s":72.8,"end_s":80.48,"text":"ARM is ensuring that every core in this CPU will perform its best at all times and keep power consumption","speaker":null,"is_sponsor":0},{"start_s":80.56,"end_s":86.16,"text":"more consistent, which will allow data centers to design to how much power their racks will","speaker":null,"is_sponsor":0},{"start_s":86.16,"end_s":92.24,"text":"consistently consume rather than having to build in a buffer for how much they might consume at peak.","speaker":null,"is_sponsor":0},{"start_s":93.12,"end_s":98.64,"text":"And that's huge, considering that cooling and especially power are just about the hottest","speaker":null,"is_sponsor":0},{"start_s":98.64,"end_s":104.32,"text":"commodities in a world that is rapidly scaling data center infrastructure. Each AGI CPU has","speaker":null,"is_sponsor":0},{"start_s":104.32,"end_s":111.04,"text":"96 lanes of PCI Express Gen 6 with support for CXL 3.0 for deploying massive shared memory pools","speaker":null,"is_sponsor":0},{"start_s":111.04,"end_s":117.04,"text":"over PCIe and ARM showed off node designs with their hardware partners that deployed up to two","speaker":null,"is_sponsor":0},{"start_s":117.04,"end_s":125.68,"text":"of these CPUs on a single motherboard. Super cool, but not exactly world-changing yet. To see the","speaker":null,"is_sponsor":0},{"start_s":125.68,"end_s":130.16,"text":"vision that led ARM to spend the last few years bringing this to life, you gotta zoom out and look","speaker":null,"is_sponsor":0},{"start_s":130.16,"end_s":138.48,"text":"beyond the individual node to the rack level. This rack contains 32 node 1P servers, so for those","speaker":null,"is_sponsor":0},{"start_s":138.48,"end_s":146.4,"text":"keeping count at home, that's 8,160 CPU cores. Okay, still not that big of a deal. I mean,","speaker":null,"is_sponsor":0},{"start_s":146.4,"end_s":153.36,"text":"dense CPU racks are already a thing. Well, here comes the big reveal. This sick error message","speaker":null,"is_sponsor":0},{"start_s":153.36,"end_s":159.76,"text":"hoodie is now available from LTTstore.com. JK, okay, I mean, it is, but that's not the big reveal.","speaker":null,"is_sponsor":0},{"start_s":159.76,"end_s":166.48,"text":"The big reveal is that everything that I just told you fits in a standard OCP 36 kilowatt air","speaker":null,"is_sponsor":0},{"start_s":166.48,"end_s":175.04,"text":"cooled rack. Each AGI CPU draws just 300 watts, a significant reduction compared to flagship","speaker":null,"is_sponsor":0},{"start_s":175.04,"end_s":182.16,"text":"x86 CPUs. So when you throw liquid cooling at them, the numbers get frankly kind of ridiculous.","speaker":null,"is_sponsor":0},{"start_s":182.16,"end_s":190.64,"text":"In an OCP 200 kilowatt rack, ARM figures, they can pack 42 8 node 1P systems for a grand total","speaker":null,"is_sponsor":0},{"start_s":190.64,"end_s":200.32,"text":"of 45,696 cores and over a petabyte of RAM, all while consuming only about half of that total","speaker":null,"is_sponsor":0},{"start_s":200.32,"end_s":205.84,"text":"available power budget. They are pegging the bottom line performance per watt in the neighborhood of","speaker":null,"is_sponsor":0},{"start_s":205.84,"end_s":212.16,"text":"double compared to x86. And this is largely thanks to carrying less legacy craft, but also","speaker":null,"is_sponsor":0},{"start_s":212.16,"end_s":217.12,"text":"thanks to architectural choices like using fewer chiplets to keep memory latency down,","speaker":null,"is_sponsor":0},{"start_s":217.12,"end_s":221.76,"text":"along with ARM's traditional strength in instructions per clock, and taking just a no","speaker":null,"is_sponsor":0},{"start_s":221.76,"end_s":228.56,"text":"silicon wasted approach to their design. With the cost and scarcity of power, that's a number that","speaker":null,"is_sponsor":0},{"start_s":228.56,"end_s":235.76,"text":"is going to perk up a lot of years. But why though? Everybody knows that CPUs aren't good at AI,","speaker":null,"is_sponsor":0},{"start_s":235.84,"end_s":242.72,"text":"compared to GPUs or application specific neural processors. So what's with the branding?","speaker":null,"is_sponsor":0},{"start_s":243.52,"end_s":249.28,"text":"ARM met that question head on. While GPUs and neural accelerators get all the attention,","speaker":null,"is_sponsor":0},{"start_s":249.28,"end_s":255.12,"text":"CPUs are still chugging along in the background, coordinating tasks, with ARM estimating that","speaker":null,"is_sponsor":0},{"start_s":255.12,"end_s":262.88,"text":"a typical deployment today is going to have about 30 million cores per gigawatt. But here's the thing,","speaker":null,"is_sponsor":0},{"start_s":262.88,"end_s":269.92,"text":"that's with humans handling most of the token requests. AI agents push requests much faster and","speaker":null,"is_sponsor":0},{"start_s":271.12,"end_s":276.48,"text":"don't sleep, meaning that your expensive AI accelerators can end up sitting around because","speaker":null,"is_sponsor":0},{"start_s":276.48,"end_s":283.28,"text":"the CPU coordinators can't keep up with all of those requests. So ARM figures that that 30 million","speaker":null,"is_sponsor":0},{"start_s":283.28,"end_s":289.36,"text":"cores per gigawatt number could go up a lot in the head node next to the accelerator rack,","speaker":null,"is_sponsor":0},{"start_s":289.36,"end_s":296.72,"text":"as high as about four times as many. But here's the thing, when these are doing all the actual AI","speaker":null,"is_sponsor":0},{"start_s":296.72,"end_s":303.76,"text":"work, nobody's going to want to spend more power budget on all of those CPUs. Well, that's where","speaker":null,"is_sponsor":0},{"start_s":303.76,"end_s":309.92,"text":"ARM comes in with their famously power efficient designs. Let's go to Nick from the lab to see","speaker":null,"is_sponsor":0},{"start_s":309.92,"end_s":314.56,"text":"this thing in action. Many of the demos were focused on the ease-supporting software to ARM","speaker":null,"is_sponsor":0},{"start_s":314.56,"end_s":319.2,"text":"and the support they're building for developers, which makes a lot of sense, but isn't very visual,","speaker":null,"is_sponsor":0},{"start_s":319.2,"end_s":325.28,"text":"so let's check out this one instead, where they're encoding a 1080p video from H.264 to H.265","speaker":null,"is_sponsor":0},{"start_s":325.28,"end_s":329.76,"text":"while running computer vision at the same time on the same CPU. Let's go take a look at the man","speaker":null,"is_sponsor":0},{"start_s":329.76,"end_s":336.64,"text":"behind the curtain. That's not a video recording. ARM actually had the stones to do it live,","speaker":null,"is_sponsor":0},{"start_s":336.64,"end_s":344.64,"text":"bringing an actual server running the actual hardware here to the show floor. But awkward","speaker":null,"is_sponsor":0},{"start_s":345.44,"end_s":350.72,"text":"Doesn't all of this put ARM kind of in direct competition with their own customers, you know,","speaker":null,"is_sponsor":0},{"start_s":350.72,"end_s":355.2,"text":"the ones who license their IP and their compute subsystems, the guys who got them where they are","speaker":null,"is_sponsor":0},{"start_s":355.2,"end_s":363.36,"text":"today? Well, on paper, yes, absolutely. But from ARM's perspective, this is actually something","speaker":null,"is_sponsor":0},{"start_s":363.36,"end_s":369.36,"text":"that many of their customers were asking for. Expanding on that, ARM laid out how their roadmap","speaker":null,"is_sponsor":0},{"start_s":369.36,"end_s":375.04,"text":"and their policies account for how all three of their business models are going to go forward,","speaker":null,"is_sponsor":0},{"start_s":375.04,"end_s":381.12,"text":"and they're positioning this as a choice between IP licensing, compute subsystems licensing,","speaker":null,"is_sponsor":0},{"start_s":381.12,"end_s":387.6,"text":"and physical CPUs, or hey, why not some combination of all three? They'll gladly take your money","speaker":null,"is_sponsor":0},{"start_s":387.6,"end_s":391.04,"text":"any way you want to give it to them. Contact your local sales representative.","speaker":null,"is_sponsor":0},{"start_s":393.52,"end_s":399.28,"text":"Will is here. He can be reached afterwards. And it seems like that's the plan","speaker":null,"is_sponsor":0},{"start_s":399.28,"end_s":404.88,"text":"for the long haul. In a move that I don't think I've ever seen before, ARM stood up on stage and","speaker":null,"is_sponsor":0},{"start_s":404.88,"end_s":411.44,"text":"said the quiet part out loud. This is just a safe first attempt. The best is yet to come with our","speaker":null,"is_sponsor":0},{"start_s":411.44,"end_s":417.2,"text":"second CPU due next year. Like, obviously, given the timelines of silicon development,","speaker":null,"is_sponsor":0},{"start_s":417.2,"end_s":421.76,"text":"but you almost never hear that from a company who probably wants you to buy the hardware they have","speaker":null,"is_sponsor":0},{"start_s":421.76,"end_s":428.24,"text":"today from partners like, you know, for example, Supermicro. Pretty wild. If you guys enjoyed","speaker":null,"is_sponsor":0},{"start_s":428.24,"end_s":433.12,"text":"this video, you might enjoy the one that we did at CES, also in partnership with ARM, highlighting","speaker":null,"is_sponsor":0},{"start_s":433.12,"end_s":437.36,"text":"some of the unexpected places that you can find ARM technology.","speaker":null,"is_sponsor":0}],"full_text":"If you're watching this right now, you're probably using at least one ARM CPU to do it. Or, well, not an ARM CPU, because ARM doesn't actually make CPUs, or do they? That's the big news that they sponsored us down here to their ARM Everywhere event to announce. Behind me, and in my hand, is the ARM AGI CPU, built for performance, scale, and, as always, up to 136 ARM Neoverse V3 cores with two megabytes of level 2 cache each, built on TSMC's 3nm process node, and it can run at up to 3.6GHz, which, right out of the gate. This is some questions, doesn't it? Just 3.6GHz? Can it, like, dynamically boost a single core way higher or something? No. And, according to ARM, that's actually a key feature, not a bug. By eschewing SMT multithreading and the highly variable power consumption that's associated with constantly fluctuating clock speeds, not to mention designing a 12-channel DDR5 memory controller that can feed every individual core with a consistent 6GB per second of bandwidth, ARM is ensuring that every core in this CPU will perform its best at all times and keep power consumption more consistent, which will allow data centers to design to how much power their racks will consistently consume rather than having to build in a buffer for how much they might consume at peak. And that's huge, considering that cooling and especially power are just about the hottest commodities in a world that is rapidly scaling data center infrastructure. Each AGI CPU has 96 lanes of PCI Express Gen 6 with support for CXL 3.0 for deploying massive shared memory pools over PCIe and ARM showed off node designs with their hardware partners that deployed up to two of these CPUs on a single motherboard. Super cool, but not exactly world-changing yet. To see the vision that led ARM to spend the last few years bringing this to life, you gotta zoom out and look beyond the individual node to the rack level. This rack contains 32 node 1P servers, so for those keeping count at home, that's 8,160 CPU cores. Okay, still not that big of a deal. I mean, dense CPU racks are already a thing. Well, here comes the big reveal. This sick error message hoodie is now available from LTTstore.com. JK, okay, I mean, it is, but that's not the big reveal. The big reveal is that everything that I just told you fits in a standard OCP 36 kilowatt air cooled rack. Each AGI CPU draws just 300 watts, a significant reduction compared to flagship x86 CPUs. So when you throw liquid cooling at them, the numbers get frankly kind of ridiculous. In an OCP 200 kilowatt rack, ARM figures, they can pack 42 8 node 1P systems for a grand total of 45,696 cores and over a petabyte of RAM, all while consuming only about half of that total available power budget. They are pegging the bottom line performance per watt in the neighborhood of double compared to x86. And this is largely thanks to carrying less legacy craft, but also thanks to architectural choices like using fewer chiplets to keep memory latency down, along with ARM's traditional strength in instructions per clock, and taking just a no silicon wasted approach to their design. With the cost and scarcity of power, that's a number that is going to perk up a lot of years. But why though? Everybody knows that CPUs aren't good at AI, compared to GPUs or application specific neural processors. So what's with the branding? ARM met that question head on. While GPUs and neural accelerators get all the attention, CPUs are still chugging along in the background, coordinating tasks, with ARM estimating that a typical deployment today is going to have about 30 million cores per gigawatt. But here's the thing, that's with humans handling most of the token requests. AI agents push requests much faster and don't sleep, meaning that your expensive AI accelerators can end up sitting around because the CPU coordinators can't keep up with all of those requests. So ARM figures that that 30 million cores per gigawatt number could go up a lot in the head node next to the accelerator rack, as high as about four times as many. But here's the thing, when these are doing all the actual AI work, nobody's going to want to spend more power budget on all of those CPUs. Well, that's where ARM comes in with their famously power efficient designs. Let's go to Nick from the lab to see this thing in action. Many of the demos were focused on the ease-supporting software to ARM and the support they're building for developers, which makes a lot of sense, but isn't very visual, so let's check out this one instead, where they're encoding a 1080p video from H.264 to H.265 while running computer vision at the same time on the same CPU. Let's go take a look at the man behind the curtain. That's not a video recording. ARM actually had the stones to do it live, bringing an actual server running the actual hardware here to the show floor. But awkward Doesn't all of this put ARM kind of in direct competition with their own customers, you know, the ones who license their IP and their compute subsystems, the guys who got them where they are today? Well, on paper, yes, absolutely. But from ARM's perspective, this is actually something that many of their customers were asking for. Expanding on that, ARM laid out how their roadmap and their policies account for how all three of their business models are going to go forward, and they're positioning this as a choice between IP licensing, compute subsystems licensing, and physical CPUs, or hey, why not some combination of all three? They'll gladly take your money any way you want to give it to them. Contact your local sales representative. Will is here. He can be reached afterwards. And it seems like that's the plan for the long haul. In a move that I don't think I've ever seen before, ARM stood up on stage and said the quiet part out loud. This is just a safe first attempt. The best is yet to come with our second CPU due next year. Like, obviously, given the timelines of silicon development, but you almost never hear that from a company who probably wants you to buy the hardware they have today from partners like, you know, for example, Supermicro. Pretty wild. If you guys enjoyed this video, you might enjoy the one that we did at CES, also in partnership with ARM, highlighting some of the unexpected places that you can find ARM technology."}