The Mainframe Learned to Speak Arm #1875
In this episode, Ray Cochrane breaks down Hot Chips 2026, the engineering conference where IBM, NVIDIA, Intel, AMD, Arm, and Fujitsu all showed how their next processors actually work. The headline disclosure is a mainframe core that runs Arm natively. Ray also covers Apple’s odd M6 Mac mini naming, London’s first autonomous Uber rides, Amazon’s purchase of the company behind DuckDB, GitHub’s HydraFusion, the best of IFA 2026, and new USDA research on farmed salmon.
– Want to start a podcast? It’s easy to get started! Sign-up at Blubrry
– Thinking of buying a Starlink? Use my link to support the show.
Subscribe to the Newsletter.
Email Ray if you want to get in touch!
Like and Follow Geek News Central’s Facebook Page.
Support my Show Sponsor: Best Godaddy Promo Codes
Get 1Password
Full Summary
Cochrane opens the show with a personal update. He apologizes for the late rollout and the missed Monday episode, having spent the week fighting a cold. He’s also heading to Michigan to spend time with family and visit his father’s gravesite. He hopes to record a couple of shows from his dad’s old studio while he is there, including a special episode planned for Tuesday. He also points listeners to a fresh site redesign that trades the old techie look for something cleaner and friendlier.
The featured segment starts from a wrap-up post on Arm’s newsroom. However, Cochrane broadens it to cover the whole conference rather than a single article. Hot Chips has run every August since 1989, and this year’s event was the 38th, held August 23rd through the 25th at Stanford’s Memorial Auditorium.
What Hot Chips Actually Is
Cochrane draws a line between Hot Chips and the big consumer trade shows. CES and Computex exist for product announcements and marketing. Meanwhile, Hot Chips is an IEEE engineering conference where chip architects present block diagrams and die photos for thirty minutes at a stretch.
The audience matters as much as the content. Roughly five hundred people who design chips for a living fill the room, and they would spot a fudged number immediately. No written paper is required, just the talk and the slides. In-person tickets sold out this year, as did Stanford’s dorm housing. Cochrane says he plans to cover the conference annually going forward.
IBM Built a Mainframe Core That Speaks Arm
The disclosure that stopped Cochrane cold came from IBM. Its next processor for IBM Z and LinuxONE runs two completely different instruction sets natively, on every one of its eleven cores. Those are z/Architecture, IBM’s own mainframe language, and AArch64, which is 64-bit Arm.
Crucially, this is not emulation. Nor is it Arm cores glued onto the die beside the mainframe cores. IBM built 2,792 Arm instructions directly into the hardware, which it says is more than double the mainframe instruction count. Each core carries two separate decoders while sharing the caches, branch predictor and register files downstream. It switches between the two in nanoseconds.
IBM even added dedicated hardware to flip byte order, because Arm and the mainframe store numbers in opposite directions. The payoff is Arm SystemReady compliance, meaning off-the-shelf Arm Linux runs on a mainframe unmodified. Patrick Kennedy of ServeTheHome, who was in the room, wrote: “I am sitting here still in awe of what IBM is doing here; this is not Z plus Arm cores, this is Z and Arm in one core.”
Cochrane flags one precision point that is easy to get backward. IBM did not license Arm’s core designs and drop them in. Instead, it took its own mainframe core and taught it AArch64 under an architecture license, which is considerably harder engineering.
The specifications are striking. The chip uses a 2nm process, with eleven cores running above 5.7GHz sustained and no turbo mode at all. Each core gets 36MB of L2 cache, backed by a 3.5GB virtual L4 pool. Furthermore, the reliability target is eight nines, which works out to roughly three tenths of one second of unplanned downtime per year. IBM gave it no name and no ship date, though the press expects “Telum III” around 2028.
Arm, Fujitsu and NVIDIA Show Their Hands
Arm itself had plenty to discuss, starting with an unfortunate name. Its first chip in thirty-five years is called the AGI CPU, which is a product name rather than any claim about artificial general intelligence. For three and a half decades, Arm designed processor blueprints and licensed them out, collecting royalties without competing. That era is now over. The AGI CPU is Arm’s own silicon, co-designed with lead customer Meta, running up to 136 cores on TSMC’s 3nm process at 300 watts. Arm’s CEO says the company has more than $2 billion in customer demand across the next two fiscal years.
Fujitsu brought the detail Cochrane called the coolest of the conference. Its MONAKA chip packs 144 Arm-based cores, but the trick is the cache. Rather than sitting alongside the cores and eating die area, the entire last-level cache lives on a separate 5nm die with the 2nm compute die stacked directly on top. It ships in 2027 in 350-watt and 500-watt versions.
NVIDIA had more stage time than anyone, with six sessions. Its new Vera CPU carries 88 cores of NVIDIA’s own Olympus design, which marks a change: the previous Grace CPU used Arm’s off-the-shelf cores. Consequently, Arm’s win here is the instruction set, not the blueprint. The memory disclosure drew the most attention, with a fully loaded system reaching 1.5TB at 1.2TB/s while the whole memory subsystem draws just 30 to 40 watts.
The Caveat on NVIDIA’s Benchmark Slides
Cochrane pushes back on how NVIDIA presented its numbers. On the standard SPEC integer benchmark, Vera scored 925 against AMD’s 128-core EPYC score of 898, about three percent ahead. However, the slide NVIDIA showed normalizes that same result per physical core, which makes a three percent gap look enormous.
NVIDIA defends the choice, arguing that per-core throughput matters when thousands of AI agents run at once. Cochrane grants that it is a fair argument to make. Even so, his verdict is blunt: it is a different number from the headline one, and presenting it that way is not the best look.
Sponsor: GoDaddy
Economy hosting $6.99/month, WordPress hosting $12.99/month, domains $11.99. Website builder trial available. Use codes at geeknewscentral.com/godaddy to support the show.
Apple’s Newest Chip Landed in Its Cheapest Mac
Last week’s episode covered the Mac Studio half of Apple’s August announcement. Tonight Cochrane takes the other half, the new Mac mini, and finds the numbering genuinely strange.
The $899 base Mac mini gets the M6, which is Apple’s first 2nm chip and the newest process the company has shipped. It brings twelve CPU cores, twelve GPU cores, and 170GB/s of memory bandwidth. Apple also introduced a third CPU core class called super cores.
So Apple’s most advanced chip sits in its cheapest desktop, while the M5 Pro, M5 Max and M5 Ultra above it all carry a lower number. It gets stranger. The M6 mini has Thunderbolt 4 while the pricier M5 Pro mini has Thunderbolt 5, and the memory ceiling runs backward too. Apple explains none of it across three announcement pages. Reading between the lines, Cochrane figures the M6 is the entry point of a new generation that shipped ahead of its larger siblings.
London’s Robotaxis Started Carrying Passengers
Arm’s monthly roundup covers everything outside the data center, and Wayve stood out. Transport for London granted the British self-driving company private hire vehicle licenses on August 5th, the same category a minicab needs.
Then on September 3rd the service launched with Uber, marking the first autonomous rides ever offered to UK passengers. A small fleet of Ford Mustang Mach-Es covers anywhere in London except the airports, with a TfL-licensed safety driver still aboard. Over 140,000 Londoners signed up. The compute runs on NVIDIA’s Arm-based automotive platform, which is why Arm claims the win. Cochrane notes the pattern: once you set a standard nobody can move off, these wins keep arriving.
Two more items round it out, both about squeezing AI onto phones. Google’s Pixel 11 shipped in August with the Tensor G6, and Google claims on-device AI runs up to 3.5x faster using 3.5x less energy. Those are Google’s own unbenchmarked figures. Separately, Graphcore’s research team worked with Arm to run an 11-billion-parameter vision model on phone-class processors by squeezing each parameter to 2.7 bits, taking the model from roughly 22GB down to 3.7GB.
Intel Is Pitching AI Infrastructure From a Long Way Back
Intel’s newsroom post previews the AI Infra Summit, running September 15th through 17th in Santa Clara. CEO Lip-Bu Tan takes a fireside chat on Tuesday morning, with three Intel sessions across the show.
Cochrane unpacks two terms first. Physical AI means AI that acts in the real world through sensors and motors rather than living on a screen, and Intel takes it seriously enough to have renamed its PC division the Client Computing and Physical AI Group in May. Disaggregated inference splits the two phases of running a model: reading your prompt is compute-hungry, while writing the answer back is memory-hungry.
The context is where this gets interesting. Intel’s revenue rose 25 percent last quarter, its fastest growth since 2011. Nevertheless, in AI accelerators the company barely registers next to NVIDIA. Gaudi has effectively been abandoned, to the point that Intel stopped maintaining its open-source driver, and AMD passed Intel in data center revenue last quarter. Tellingly, when Intel demoed disaggregated inference at Computex, NVIDIA GPUs handled one phase and SambaNova chips the other while Intel supplied the coordinating CPU. Crescent Island, Intel’s actual inference chip, does not sample until later this year, and Intel declined to publish its memory bandwidth.
Amazon Bought the Company Behind DuckDB
On August 26th, Amazon signed a deal to acquire DuckLabs, the Amsterdam company behind DuckDB. Cochrane spends time explaining what DuckDB is, since listeners outside the data world may never have encountered it.
The problem it solves is familiar. Querying a large pile of data files traditionally meant either running a database server or spinning up a data warehouse with a cluster, a bill, and a loading pipeline. Both are heavy machinery for a question you wanted answered in ten seconds. DuckDB instead ships as a library rather than a server. You add it to your program like any other package, point it at your files, and write ordinary SQL directly against them. The data never moves.
The usual comparison is SQLite, which is embedded in nearly every phone and browser on earth. Where SQLite excels at looking up one record, DuckDB rebuilds that embedded idea for chewing through millions of rows. It now sees roughly 62 million monthly downloads on Python’s package index alone, up from about 25 million last October.
AWS says it is buying the company, not the project. DuckDB stays free and open source under the MIT license, held by a Dutch nonprofit foundation, and the founders join AWS while continuing to run technical direction from Amsterdam. Andy Warfield, a VP and distinguished engineer at AWS, described DuckDB as “the glibc of structured data: a lean, unglamorous, ubiquitous dependency that a great deal of software links against and almost nobody has to think about.”
Cochrane sits with what that ownership means. He reaches for an analogy: imagine Daniel Stenberg selling curl. He doubts it would ever happen, but the concept alone is startling given how much infrastructure depends on it. His read is that AWS is betting DuckDB becomes as foundational as curl and SQLite already are.
GitHub Has One AI Model Grade Another One’s Homework
GitHub shipped Project HydraFusion into Copilot as a research preview. Instead of routing your request to a single model, it picks one of three approaches per request. Sometimes one model simply answers. Alternatively, a cheaper model drafts, and a quality gate decides whether to escalate.
The interesting one is Critique. One model writes the code, a separate model from a different family reviews it read-only, and the original gets one pass to revise. Despite the name, nothing is fused here. There is no voting and no merging, just one model at a time with a gate deciding whether to spend more.
GitHub explained the reasoning in an earlier post: “a model reviewing its own work is still bounded by its own training biases: the same training data and techniques, the same blind spots.” Research supports it. A team at NeurIPS in 2024 showed that models recognize their own writing and score it higher than human graders do. GitHub’s earlier number had a Claude Sonnet and GPT critic pairing closing about three-quarters of the gap between Sonnet and the larger Opus model.
This mirrors a workflow Cochrane uses constantly and has described on a previous episode. He runs a cross-check review with a second model from a different company, and it routinely surfaces issues the first model missed. He explains that different training data, different engineers, and different reinforcement approaches build different internal biases about what counts as correct. Looking ahead, he expects more of these “Frankenstein patterns” where models from different training families work together.
The Best of IFA 2026, and What You Can Actually Buy
IFA opened to the public in Berlin for its 102nd year, with about 1,900 brands. Cochrane splits The Verge’s roundup in two, since much of what generates headlines at these shows never ships.
Starting with real products, iRobot’s flagship Roomba Max 875 Combo runs $1,199 and ships in about two weeks. Its SealForce feature drops a hidden skirt from the chassis when it detects carpet, sealing against the fibers so suction concentrates instead of leaking out the sides. That reaches 35,000 pascals, iRobot’s strongest yet. A step-down model at $899 carries the same trick, though Cochrane balks at both prices.
Anker’s Soundcore Sleep 4 Pro earbuds arrive in November at $349.99. The charging case carries its own round touchscreen, so you pick soundscapes, set alarms, and read sleep stats without your phone. Optical sensors read heart rate and variability from the ear canal, which beats the wrist for accuracy, and the case masks a snoring partner. Cochrane remains unconvinced about sleeping with earbuds in.
Philips also has smart rope lights, the Hue Liane 360, on sale now. They glow evenly around the tube rather than showing individual LEDs. They also run $400 for three meters, which works out to about $130 per meter of rope light.
As for concepts nobody can buy, iRobot showed a robot vacuum that carries a smaller robot vacuum on its back in a garage and lowers it to deploy. Lenovo brought a 14-inch laptop whose screen rolls out to 17 inches at the press of a button, which reviewers call the first rollable that feels close to shippable. Tecno showed a phone with essentially no border around the screen, and Acer had a Windows gaming handheld that swivels its screen up over a keyboard.
Two themes ran through the show. Humanoid robots were the loudest thing on the floor, and IFA’s own CEO framed the event as being about robots that work rather than robots that demo. Meanwhile, AI stopped being its own product category and became an ingredient, showing up in refrigerators, treadmills, dishwashers, and motorized TV mounts.
Farmed Salmon Isn’t the Omega-3 Machine It Used to Be
USDA scientists measured farmed salmon and found considerably less of the good fat than the government’s own database claims. EPA and DHA are two fatty acids you get almost entirely from fish. Your body can build them from the plant version, but only in tiny amounts, so the NIH’s position is that eating them is the only practical way to raise your levels.
Those fatty acids are structural pieces of every cell, with DHA concentrating in the brain and retina. That is why the federal dietary guidelines, issued jointly by USDA and Health and Human Services, recommend at least eight ounces of fish a week and steer you toward salmon. That amount is calibrated to deliver about 250 milligrams a day.
Published in Frontiers in Nutrition last month, the study found EPA and DHA in farmed Atlantic salmon came in 54.7 percent lower than USDA’s own reference values, last updated in 2018. A three-ounce serving fell from roughly 1,670 milligrams to about 756. Consequently, two servings a week now fall about 14 percent short of the target. Plant-derived fats meanwhile rose two to three times over.
The likely cause is feed. Salmon are carnivores, and farms once fed them oily little fish. There was never going to be enough of those as the industry scaled, so crops filled the gap: soy, canola, sunflower and linseed. Importantly, the study does not claim to have proven this and calls the feed shift a plausible explanation. Independent corroboration lends it credibility. Researchers at Stirling measured a similar halving in Scottish salmon between 2006 and 2015, and Norway’s marine institute saw it across thousands of samples.
There is a land dimension too. Roughly half the world’s soy grows in South America, where rainforest gets cleared for feed. Matthew Hayek, who studies the environmental cost of protein at NYU, told Inside Climate News that “soy is a major, important protein and oil ingredient in fish farming.” Adding up two decades of soy across all fish farming, he puts the extra forest clearing at around the area of Nicaragua or Bangladesh. That figure covers all fish farming rather than salmon alone, and Hayek notes it is hard to attribute soy use to any single species.
Cochrane closes with two caveats. First, the study measured only eight fish, bought around Maryland, DC and Virginia over six weeks in 2023, and nearly all sourced from Chile. That is not a national survey, and the authors say plainly the sample was not large enough to change government advice. Rather, it flags that a federal database value needs rechecking, which is what the paper set out to do.
Second, on whether you should care, farmed salmon still beats beef, chicken and eggs by a mile, since those carry essentially zero EPA and DHA. What it loses is its crown among fatty fish, dropping to mid-pack behind herring, sardines and mackerel and roughly level with trout. The broader health case is also softer than the 2000s suggested. A review of 86 trials covering 162,000 people found supplements barely moved heart attacks or deaths, so eating fish and swallowing fish oil are not the same claim. No producer has responded to the findings, and USDA, whose own scientists ran the study, declined an interview and did not answer emailed questions.
Cochrane wraps up with housekeeping and a note that he will be back on Labor Day.
The post The Mainframe Learned to Speak Arm #1875 appeared first on Geek News Central.
