DiscoverSilicon To Software | Autonomous Systems | Cybersecurity
Silicon To Software | Autonomous Systems | Cybersecurity
Claim Ownership

Silicon To Software | Autonomous Systems | Cybersecurity

Author: Imran Valiani

Subscribed: 0Played: 0
Share

Description

Autonomous systems and cybersecurity run on hardware most tech writers never see up close. Silicon to Software explores the deep hardware layer behind AI, autonomous systems, and modern technology β€” the circuit boards, chips, and supply chains that make it all work. Hosted by Imran Valiani, an industry insider with 20+ years in PCB electronics manufacturing and technical sales, serving major Bay Area and global tech clients. Each episode is an expert technical teardown: autonomous vehicle vulnerabilities, the future of silicon compute, AI infrastructure, hardware, and cybersecurity.

🌐 Read our latest articles and get in touch at https://silicontosoftware.com

40Β Episodes
Reverse
On the evening of July 10, 2024, a transmission line in Northern Virginia faulted, and about 1,500 MW of load dropped off the grid. None of it was homes. None of it was factories. All of it was data center load. It has happened twice more in the same zone β€” most recently on July 22, 2026, when PJM put the load reduction at 3,800–3,900 MW, the largest of its kind in PJM's history.AI training makes the problem worse. In a normal data center, thousands of unrelated jobs average each other out. AI training doesn't work that way β€” the GPUs compute together and pause together, so their power curves stack instead of canceling. NVIDIA says a single rack can swing from about 30% of its power draw to full power and back in milliseconds.So where does storage actually sit? In three layers: capacitors in the GB300 power shelf (about half a power supply by volume, 65 joules per GPU β€” a shock absorber measured in hundredths of a second, not backup power), lithium-ion battery backup units in the rack, inches from the IT load, and big battery systems at the facility level, where the site connects to the utility.Then the three risks most people pricing this storage aren't seeing: cycle life (smoothing storage cycles all day, and electrolytic capacitors hate heat β€” every 10Β°C cooler roughly doubles their life), fire (the September 2025 fire in the lithium-ion UPS batteries at South Korea's National Government Data Center damaged 96 of 647 government systems and destroyed all 384 UPS battery units), and the board itself (NVIDIA is targeting 2027 to move to 800 VDC β€” 1 MW is about 1,250 A at 800 V versus 18,500 A at 54 V β€” but spacing gets harder, laminate choice matters more, and conductive anodic filament becomes a real reliability risk under years of DC bias).That leaves one question for anyone specifying this storage: if it smooths the workload all day, does it still have enough charge left to ride through a grid fault?For the full written analysis, the sources and the diagrams, head over to silicontosoftware.com.Subscribe and follow Silicon to Software for more breakdowns of the hardware behind AI β€” from power shelves to PCB laminates.Chapters:00:00 The Virginia event: 1,500 MW drops off00:26 It's happened twice more00:47 Where the storage actually sits01:08 The workload: GPUs pause together01:41 Layer one: capacitors in the rack02:24 Layers two and three: BBUs and facility BESS02:40 Risk one: cycle life03:16 Risk two: fire03:52 Risk three: the board04:32 The question to ask
Liquid cooling didn't remove the heat problem from AI racks. It moved it into the board.In late 2024, reports said Blackwell GPUs were overheating in 72-processor servers, and suppliers spent months working through overheating and coolant leaks in GB200 liquid-cooling systems. The root causes were never disclosed β€” but the reports tell you something: a liquid-cooled rack needs a lot more than raw heat transfer to work.In this episode: how the loop actually works β€” a cold plate on each GPU package, coolant through quick disconnects into a rack manifold, then a CDU handing heat to the building's loop. Why today's GB300 racks are still hybrid, about 85% liquid, with optics, storage and some power parts still on air. And why HPE's three rack-power numbers β€” 132 kW nominal, about 155 kW peak, 192 kW provisioning β€” answer three different questions.Then the part most people miss: the cold plate has to press on the package, hard and evenly, so the TIM makes good contact. A 2006 study of one thermal interface material found thermal resistance could differ by up to 45% between high- and low-pressure application. That clamp force is carried through the package into the solder joints, into the board and the backplate. Too much force stresses the BGA and bends the board; too little and you get hot spots. There's a window, and it has to be qualified for each assembly.Plus: Rubin goes to 100% liquid with 45Β°C coolant, PG25 chemistry and quick-disconnect leak risk, liquid-cooled busbars up to 5,000 amps, and Microsoft's lab microfluidics cooling up to three times better than cold plates.Liquid can carry the heat. Whether the rack stays reliable gets decided at the boundaries between suppliers β€” and the board sits right in the middle.Full written analysis with figures and every source: https://www.silicontosoftware.com/direct-to-chip-liquid-cooling-ai-racks/Subscribe and follow Silicon to Software for more breakdowns of where AI hardware actually breaks.Chapters:00:00 Welcome to Silicon to Software00:18 The Blackwell overheating reports00:55 How the loop works01:10 Hybrid racks and three power numbers01:42 Clamp force and the TIM02:59 Rubin goes 100% liquid03:29 The 45Β°C coolant trade-off03:50 PG25, leaks and liquid-cooled busbars04:14 Microfluidics: cooling the silicon itself04:26 What it means for board builders
Back in 2021, Tesla was trying to ramp its 4680 cell β€” and one of the first signs of trouble didn't come from a chemistry lab. It came from the steel rollers. The cathode material was so hard, it was damaging the calender rolls β€” and only when the rolls got bigger.In this episode:- What a dry electrode actually is β€” and why skipping the solvent slurry moves the control problem from the drying ovens to the mixer and the calender- Why the cathode, not the anode, became the bottleneck β€” and the reports of Tesla losing most of its cathodes in test production- The PTFE binder trade-off: it holds the film together, but blocks electrons and ions and doesn't always stick to the foil- The defect that worries a manufacturing guy most: failures that don't show up until months later- The January 2026 update: Tesla says it's making 4680 cells with both electrodes dry-processed in Austin β€” and what still isn't published (yield, scrap rates, cost per kWh)- The real test: copy exactly β€” same yield on every line, every site, every shift. A process isn't solved when it works. It's solved when it's boring.Full analysis: https://www.silicontosoftware.com/tesla-4680-dry-electrode/
Every high-end AI processor sits on a package substrate β€” a tiny, dense wiring board connecting the chip to the motherboard. For decades its core has been organic, a woven glass-fiber laminate. AI packages are getting huge, and at that size a familiar problem gets loud: warpage.In this episode:- Why warpage threatens assembly yield: silicon, the organic core, and copper expand at different rates when heated, so the stack bends β€” and Intel has shown an engineering sample carrying about twice the silicon a lithography tool can print in one shot- What glass is supposed to fix: it is stiff, extremely flat, and its expansion can be tuned closer to silicon- Measured results, with their limits: test vehicles from Amkor and STATS ChipPAC showed roughly a third less warpage with a glass core β€” specific test designs, not a general rule; in one study, packages with no protective edge coating failed and coated ones passed- The simulation that complicates the story: a Unimicron model, reported secondhand and not a physical test, found glass roughly halved strain where the chip meets the substrate but more than doubled strain on the larger solder joints where the substrate meets the board- Why glass cracks: copper vias run through the glass and copper expands far more than glass, with cracks pointed to at via walls, via caps, and panel cut edges- Where the big players stand: Absolics (SKC's subsidiary in Georgia) in an initial reliability evaluation with samples in Taiwan, with a delay acknowledged in SKC's own financing materials; Samsung Electro-Mechanics with two separate targets β€” the glass core in the second half of 2027 and finished substrates from 2028, with a disputed claim about a failed prototype; and Intel, which put glass on its roadmap in 2023 and has shown engineering work but has named no customer or product, with no revised timeline found- The five gates: customer reliability testing is only the first β€” then process-window mapping, holding it lot after lot, sustained yield, and costs that compete with mature organic substrates; no glass-core supplier has published production yield dataThe verdict: a flat substrate is a materials result. A qualified substrate is a manufacturing result. Next time you see a glass announcement, ask which of the five gates it actually clears.Full analysis: https://www.silicontosoftware.com/glass-substrates-for-ai-chips/
A one-megawatt AI rack fed at 54 volts has to pull about 18,500 amps. At 800 volts, the same rack draws about 1,250. NVIDIA estimates a 54-volt megawatt rack would need up to 200 kilograms of copper busbar, and its power shelves alone would eat up to 64 rack units.In this episode:- Why rack power is climbing toward a megawatt: more GPUs packed into one rack so they can talk over short copper NVLink links- The 54-volt comparison: roughly 15 times less current at 800 volts, far less heat in the copper- The comparison most coverage misses: against 480-volt AC already in many data halls, each wire carries almost the same current β€” the gain is two conductors instead of three- The transitional rack: the first 800-volt hardware converts back down to 54 volts inside the compute rack, adding a conversion step rather than removing one; NVIDIA projects up to 5% better efficiency, and as of late October no independent measurements from an operating 800-volt AI data hall were found- Where the hard engineering moves: converters close to the GPU (silicon carbide, gallium nitride, stacked topologies β€” vendor numbers, not field data), and the circuit board, where 800-volt spacing, laminate choice, manufacturing tolerances, and cleanliness become safety issues- Protection and certification: DC arcs are harder to extinguish than AC arcs, grounding is still undecided, solid-state breakers are only mature at lower currents, and key standards were written around lower voltages or ACThe verdict: against a 54-volt bus the physics is decisive, but 800-volt DC does not make the power problem disappear β€” it moves it into converters, boards, connectors, breakers, and certification.Full analysis: https://www.silicontosoftware.com/nvidia-800v-dc-architecture/
loading
CommentsΒ