{"video_id":"WTFNWpXiqEM","title":"There’s a new CPU maker.","channel":"Linus Tech Tips","show":"Linus Tech Tips","published_at":"2026-04-04T14:53:29Z","duration_s":438,"segments":[{"start_s":0.0,"end_s":6.68,"text":"If you're watching this right now, you're probably using at least one arm","speaker":null,"is_sponsor":0},{"start_s":3.92,"end_s":10.8,"text":"CPU to do it. Or well, not an arm CPU because arm doesn't actually make CPUs.","speaker":null,"is_sponsor":0},{"start_s":9.52,"end_s":15.8,"text":"Or do they? That's the big news that they sponsored","speaker":null,"is_sponsor":0},{"start_s":13.2,"end_s":22.36,"text":"us down here to their arm everywhere event to announce. Behind me and in my","speaker":null,"is_sponsor":0},{"start_s":18.96,"end_s":25.8,"text":"hand is the arm AGI CPU built for","speaker":null,"is_sponsor":0},{"start_s":22.36,"end_s":28.64,"text":"performance, scale, and as always","speaker":null,"is_sponsor":0},{"start_s":25.8,"end_s":34.56,"text":"efficiency. Up to 136 arm Neoverse V3 cores with 2 megabytes","speaker":null,"is_sponsor":0},{"start_s":31.4,"end_s":37.2,"text":"of level 2 cache each built on TSMC's 3","speaker":null,"is_sponsor":0},{"start_s":34.56,"end_s":41.72,"text":"nanometer process node, and they can run at up to 3.6 GHz, which","speaker":null,"is_sponsor":0},{"start_s":40.2,"end_s":47.6,"text":"right out of the gate raises some questions, doesn't it? Just 3.6 GHz? Can","speaker":null,"is_sponsor":0},{"start_s":45.0,"end_s":52.68,"text":"it like dynamically boost a single core way higher or something?","speaker":null,"is_sponsor":0},{"start_s":49.68,"end_s":55.48,"text":"No. And according to arm, that's","speaker":null,"is_sponsor":0},{"start_s":52.68,"end_s":60.2,"text":"actually a key feature, not a bug. By eschewing SMT multithreading and the","speaker":null,"is_sponsor":0},{"start_s":58.24,"end_s":64.32,"text":"highly variable power consumption that's associated with constantly fluctuating","speaker":null,"is_sponsor":0},{"start_s":62.12,"end_s":69.68,"text":"clock speeds, not to mention designing a 12-channel DDR5 memory controller that","speaker":null,"is_sponsor":0},{"start_s":67.0,"end_s":75.8,"text":"can feed every individual core with a consistent 6 GB/s of bandwidth, arm is","speaker":null,"is_sponsor":0},{"start_s":73.24,"end_s":82.0,"text":"ensuring that every core in the CPU will perform its best at all times and keep","speaker":null,"is_sponsor":0},{"start_s":79.52,"end_s":86.92,"text":"power consumption more consistent, which will allow data centers to design to how","speaker":null,"is_sponsor":0},{"start_s":84.12,"end_s":91.8,"text":"much power their racks will consistently consume rather than having to build in a","speaker":null,"is_sponsor":0},{"start_s":89.24,"end_s":95.92,"text":"buffer for how much they might consume at peak.","speaker":null,"is_sponsor":0},{"start_s":93.24,"end_s":100.16,"text":"And that's huge considering that cooling and especially power are just about the","speaker":null,"is_sponsor":0},{"start_s":98.2,"end_s":105.04,"text":"hottest commodities in a world that is rapidly scaling data center","speaker":null,"is_sponsor":0},{"start_s":101.56,"end_s":107.84,"text":"infrastructure. Each AGI CPU has 96","speaker":null,"is_sponsor":0},{"start_s":105.04,"end_s":113.52,"text":"lanes of PCI Express Gen 6 with support for CXL 3.0 for deploying massive shared","speaker":null,"is_sponsor":0},{"start_s":110.52,"end_s":115.12,"text":"memory pools over PCIe, and arm showed","speaker":null,"is_sponsor":0},{"start_s":113.52,"end_s":120.28,"text":"off node designs with their hardware partners that deployed up to two of","speaker":null,"is_sponsor":0},{"start_s":117.24,"end_s":122.6,"text":"these CPUs on a single motherboard.","speaker":null,"is_sponsor":0},{"start_s":120.28,"end_s":125.12,"text":"Super cool, but not exactly world-changing","speaker":null,"is_sponsor":0},{"start_s":124.04,"end_s":128.759,"text":"yet. To see the vision that led Arm to spend","speaker":null,"is_sponsor":0},{"start_s":127.72,"end_s":136.8,"text":"the last few years bringing this to life, you got to zoom out and look beyond the individual node to the rack","speaker":null,"is_sponsor":0},{"start_s":132.8,"end_s":139.04,"text":"level. This rack contains 32-node","speaker":null,"is_sponsor":0},{"start_s":136.8,"end_s":143.68,"text":"1P servers. So, for those keeping count at home, that's 8,160","speaker":null,"is_sponsor":0},{"start_s":141.92,"end_s":148.36,"text":"CPU cores. Okay, still not that big of a deal. I","speaker":null,"is_sponsor":0},{"start_s":145.92,"end_s":152.08,"text":"mean, dense CPU racks are already a thing.","speaker":null,"is_sponsor":0},{"start_s":149.56,"end_s":156.959,"text":"Well, here comes the big reveal. This sick error message hoodie is now","speaker":null,"is_sponsor":0},{"start_s":154.36,"end_s":160.88,"text":"available from lttstore.com. JK, okay. I mean, it is, but that's not","speaker":null,"is_sponsor":1},{"start_s":159.0,"end_s":168.32,"text":"the big reveal. The big reveal is that everything that I just told you fits in","speaker":null,"is_sponsor":1},{"start_s":162.76,"end_s":172.56,"text":"a standard OCP 36 kW air-cooled rack.","speaker":null,"is_sponsor":0},{"start_s":168.32,"end_s":174.44,"text":"Each AGI CPU draws just 300 W,","speaker":null,"is_sponsor":0},{"start_s":172.56,"end_s":180.64,"text":"a significant reduction compared to flagship x86 CPUs. So, when you throw","speaker":null,"is_sponsor":0},{"start_s":177.84,"end_s":186.32,"text":"liquid cooling at them, the numbers get frankly kind of ridiculous. In an OCP","speaker":null,"is_sponsor":0},{"start_s":183.08,"end_s":190.16,"text":"200 kW rack, Arm figures they can pack","speaker":null,"is_sponsor":0},{"start_s":186.32,"end_s":193.92,"text":"42 eight-node 1P systems for a grand","speaker":null,"is_sponsor":0},{"start_s":190.16,"end_s":193.92,"text":"total of 45,696","speaker":null,"is_sponsor":0},{"start_s":194.16,"end_s":202.239,"text":"cores and over a petabyte of RAM, all while consuming only about half of that","speaker":null,"is_sponsor":0},{"start_s":199.92,"end_s":205.519,"text":"total available power budget. They are pegging the bottom line","speaker":null,"is_sponsor":0},{"start_s":203.72,"end_s":210.6,"text":"performance per watt in the neighborhood of double compared to x86. And this is","speaker":null,"is_sponsor":0},{"start_s":208.6,"end_s":215.64,"text":"largely thanks to carrying less legacy cruft, but also thanks to architectural","speaker":null,"is_sponsor":0},{"start_s":213.239,"end_s":219.04,"text":"choices like using fewer chiplets to keep memory latency down, along with","speaker":null,"is_sponsor":0},{"start_s":217.72,"end_s":223.28,"text":"Arm's traditional strength in instructions per clock, and taking just","speaker":null,"is_sponsor":0},{"start_s":221.48,"end_s":229.88,"text":"a no silicon wasted approach to their design. With the cost and scarcity of power,","speaker":null,"is_sponsor":0},{"start_s":227.92,"end_s":235.88,"text":"that's a number that is going to perk up a a of ears. But, why though? Everybody","speaker":null,"is_sponsor":0},{"start_s":233.24,"end_s":241.48,"text":"knows that CPUs aren't good at AI compared to GPUs or application specific","speaker":null,"is_sponsor":0},{"start_s":238.76,"end_s":246.16,"text":"neural processors. So, uh what's with the branding?","speaker":null,"is_sponsor":0},{"start_s":243.6,"end_s":250.72,"text":"Arm met that question head-on. While GPUs and neural accelerators get","speaker":null,"is_sponsor":0},{"start_s":248.32,"end_s":254.92,"text":"all the attention, CPUs are still chugging along in the background","speaker":null,"is_sponsor":0},{"start_s":252.32,"end_s":260.0,"text":"coordinating tasks with Arm estimating that a typical deployment today is going","speaker":null,"is_sponsor":0},{"start_s":257.16,"end_s":263.48,"text":"to have about 30 million cores per gigawatt.","speaker":null,"is_sponsor":0},{"start_s":261.48,"end_s":266.68,"text":"But, here's the thing. That's with humans handling most of the token","speaker":null,"is_sponsor":0},{"start_s":265.52,"end_s":271.12,"text":"requests. AI agents push requests much faster and","speaker":null,"is_sponsor":0},{"start_s":270.2,"end_s":275.68,"text":"um don't sleep. Meaning that your expensive","speaker":null,"is_sponsor":0},{"start_s":273.6,"end_s":279.84,"text":"AI accelerators can end up sitting around because the CPU coordinators","speaker":null,"is_sponsor":0},{"start_s":278.08,"end_s":283.32,"text":"can't keep up with all of those requests.","speaker":null,"is_sponsor":0},{"start_s":281.08,"end_s":288.72,"text":"So, Arm figures that that 30 million cores per gigawatt number could go up a","speaker":null,"is_sponsor":0},{"start_s":285.52,"end_s":291.56,"text":"lot in head node next to the accelerator","speaker":null,"is_sponsor":0},{"start_s":288.72,"end_s":297.36,"text":"rack as high as about four times as many. But, uh here's the thing. When","speaker":null,"is_sponsor":0},{"start_s":294.48,"end_s":302.08,"text":"these are doing all the actual AI work, nobody's going to want to spend more","speaker":null,"is_sponsor":0},{"start_s":298.76,"end_s":303.28,"text":"power budget on all of those CPUs.","speaker":null,"is_sponsor":0},{"start_s":302.08,"end_s":308.24,"text":"Well, that's where Arm comes in with their","speaker":null,"is_sponsor":0},{"start_s":305.32,"end_s":312.08,"text":"famously power efficient designs. Let's go to Nick from the lab to see","speaker":null,"is_sponsor":0},{"start_s":310.04,"end_s":315.44,"text":"this thing in action. Many of the demos were focused on the ease of porting","speaker":null,"is_sponsor":0},{"start_s":313.6,"end_s":319.36,"text":"software to Arm and the support they're building for developers, which makes a","speaker":null,"is_sponsor":0},{"start_s":317.12,"end_s":323.08,"text":"lot of sense, but isn't very visual. So, let's check out this one instead where","speaker":null,"is_sponsor":0},{"start_s":320.96,"end_s":328.64,"text":"they're encoding a 1080p video from H.264 to H.265 while running computer","speaker":null,"is_sponsor":0},{"start_s":326.16,"end_s":331.72,"text":"vision at the same time on the same CPU. Let's go take a look at the man behind","speaker":null,"is_sponsor":0},{"start_s":330.24,"end_s":338.64,"text":"the curtain. That's not a video recording. Arm actually had the stones","speaker":null,"is_sponsor":0},{"start_s":334.68,"end_s":341.4,"text":"to do it live bringing an actual server","speaker":null,"is_sponsor":0},{"start_s":338.64,"end_s":346.24,"text":"running the actual hardware here to the show floor. But, um","speaker":null,"is_sponsor":0},{"start_s":344.24,"end_s":350.92,"text":"awkward question. Doesn't all of this put Arm kind of in direct competition","speaker":null,"is_sponsor":0},{"start_s":348.88,"end_s":354.76,"text":"with their own customers? You know, the ones who license their IP and their","speaker":null,"is_sponsor":0},{"start_s":352.92,"end_s":360.08,"text":"compute subsystems, the guys who got them where they are today? Well, on","speaker":null,"is_sponsor":0},{"start_s":356.92,"end_s":362.8,"text":"paper, yes, um absolutely.","speaker":null,"is_sponsor":0},{"start_s":360.08,"end_s":366.68,"text":"But, from ARM's perspective, this is actually something that many of their","speaker":null,"is_sponsor":0},{"start_s":364.08,"end_s":372.2,"text":"customers were asking for. Expanding on that, ARM laid out how their road map","speaker":null,"is_sponsor":0},{"start_s":369.44,"end_s":376.0,"text":"and their policies account for how all three of their business models are going","speaker":null,"is_sponsor":0},{"start_s":374.08,"end_s":381.6,"text":"to go forward. And they're positioning this as a choice between IP licensing,","speaker":null,"is_sponsor":0},{"start_s":379.24,"end_s":386.08,"text":"compute subsystems licensing, and physical CPUs, or hey,","speaker":null,"is_sponsor":0},{"start_s":384.0,"end_s":389.72,"text":"why not some combination of all three? They'll gladly take your money any way","speaker":null,"is_sponsor":0},{"start_s":388.08,"end_s":393.16,"text":"you want to give it to them. Contact your local sales representative.","speaker":null,"is_sponsor":0},{"start_s":393.56,"end_s":399.88,"text":"Will is here. He can be reached afterwards. And it","speaker":null,"is_sponsor":0},{"start_s":397.4,"end_s":404.8,"text":"seems like that's the plan for the long haul. In a move that I don't think I've","speaker":null,"is_sponsor":0},{"start_s":402.16,"end_s":410.4,"text":"ever seen before, ARM stood up on stage and said the quiet part out loud, \"This","speaker":null,"is_sponsor":0},{"start_s":407.56,"end_s":413.88,"text":"is just a safe first attempt. The best is yet to come with our second CPU due","speaker":null,"is_sponsor":0},{"start_s":412.56,"end_s":417.8,"text":"next year.\" Like, obviously, given the timelines of","speaker":null,"is_sponsor":0},{"start_s":416.28,"end_s":421.36,"text":"silicon development, but you almost never hear that from a company who","speaker":null,"is_sponsor":0},{"start_s":419.8,"end_s":426.08,"text":"probably wants you to buy the hardware they have today from partners like, you","speaker":null,"is_sponsor":0},{"start_s":423.48,"end_s":429.08,"text":"know, for example, Supermicro. Pretty wild.","speaker":null,"is_sponsor":0},{"start_s":427.64,"end_s":432.68,"text":"If you guys enjoyed this video, you might enjoy the one that we did at CES,","speaker":null,"is_sponsor":1},{"start_s":431.4,"end_s":439.12,"text":"also in partnership with ARM, highlighting some of the unexpected","speaker":null,"is_sponsor":1},{"start_s":434.68,"end_s":439.12,"text":"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 efficiency. Up to 136 arm Neoverse V3 cores with 2 megabytes of level 2 cache each built on TSMC's 3 nanometer process node, and they can run at up to 3.6 GHz, which right out of the gate raises some questions, doesn't it? Just 3.6 GHz? 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 6 GB/s of bandwidth, arm is ensuring that every core in the 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 got to 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 kW air-cooled rack. Each AGI CPU draws just 300 W, 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 kW rack, Arm figures they can pack 42 eight-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 cruft, 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 a of ears. But, why though? Everybody knows that CPUs aren't good at AI compared to GPUs or application specific neural processors. So, uh 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 um 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 head node next to the accelerator rack as high as about four times as many. But, uh 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 of porting 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, um awkward question. 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, um 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 road map 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."}