Intel E835 Ethernet Adapters Unveiled at Computex 2026

Inference keeps getting carved up, and every cut makes intelligence cheaper. First we split by phase: prefill on one set of chips, decode on another. Then by layer: attention on HBM-rich GPUs, the feed-forward network on SRAM-based silicon. Now by time itself: workloads sliced into execution windows and interleaved across the cluster. Each split recovers wasted utilization. Recovered utilization lowers the cost per token. We think cheaper tokens don't shrink demand, they grow it. That was the real story of MLSys 2026. A quick map of the three cuts. 🟠 Phase. Every request does two jobs. Prefill reads your prompt; decode writes the answer one token at a time. The two stress hardware differently, so each gets its own chips instead of sharing. 🟠 Layer. Attention lets tokens share context, which is memory-hungry. The feed-forward network refines each token on its own, which is compute-hungry. Send attention to HBM-rich GPUs and the FFN to SRAM-based silicon, and each runs where it's fastest. 🟠 Time. The newest cut. "Interleave" just means taking turns: instead of dedicating separate chips, one set runs a slice of one job, then a slice of the next, switching fast enough that nothing sits idle. The pattern under all three: find idle compute and fill it. That's what drives down the cost of intelligence.

The Cube interviewed Mark Papermaster yesterday and it is worth watching. Sounds like AMD gets the need that LiquidMetal AI has been solving - AI Factories need system level optimization to increate yield ($ Profit / GW). "We've become system optimizers, still optimizing every component, but then equally looking at how you optimize how each of the pieces come together... for the data center, we're optimizing at the rack level, hardware and software." https://lnkd.in/g5VSmZHJ #yieldcontrolplane #aifactory

Everyone tunes answering machine detection for connect rate. Almost nobody treats it as a compliance risk. It's both. Here's the mechanic people skip. Stock Asterisk AMD tops out around 80% accuracy, and operators peg its false-positive rate at 15 to 20%. A false positive means a real human picks up and gets handled like a voicemail. They answer, hear a beat of silence, and hang up. Some of them flag your number as spam on the way out. Now read the TCPA. Abandoned calls - a live person who answers and gets dead air, are capped at 3% of answered calls per campaign, per month. Go over, and each one is worth $500 to $1,500. See the collision? A 15 to 20% AMD false-positive rate isn't just costing you conversations. Every one of those misfires is a live human sitting in silence, which is the exact thing the abandoned-call rule was written to punish. Operators obsess over the drop-percentage ceiling in VICIdial, then run AMD settings calibrated back in the Asterisk 1.4 days. The dialer changed. Voicemail greetings changed. The carriers changed. The AMD parameters mostly didn't. Your AMD accuracy and your compliance exposure are the same number. Most centers only track one of them.

Product update from Radiant Intel: this week's work focused on making source-grounded reports easier to trust and use. Reports now carry required intelligence modules as first-class outputs. Evidence, actions, trigger thresholds and disclosed gaps can travel with the report structure instead of getting flattened into general prose. Readiness checks are stricter, too. A report has to clear source evidence, citations, the text readers see and required modules before analysts see it as ready, with clearer progress through longer synthesis and quality review. Library browsing is lighter for large source sets: list views stay focused on compact previews, while deeper text stays in the document detail path. The practical point is discipline between monitored signals and decision-ready analysis: clearer structure, harder evidence gates and less friction when teams move from sources to judgment.

The AI boom continues to drive record growth across the PCB industry, but the real story is upstream. High-end glass fabrics, copper foils and advanced resin systems are increasingly being pulled into AI infrastructure, creating supply and cost pressures across the wider market. As capacity migrates towards higher-value applications, availability—not price—will become the defining challenge for many OEMs and PCB manufacturers. The best time to secure allocation was in Q2 or earlier. The next best time is now. By the time shortages become obvious to everyone, most of the available capacity will already have been allocated elsewhere. Those with qualified alternatives, strong supplier relationships and credible forecasts will be in a far stronger position than those relying on the spot market.

Through our new integration with the NVIDIA DOCA Argus framework, Cortex is bringing defense-in-depth straight to the silicon level. Operating out-of-band directly from the NVIDIA BlueField DPU, we deliver: ✅ Full visibility into AI node memory, file access, and network connections. ✅ 0% host agent overhead, keeping your compute resources dedicated to your business mission. ✅ Structured, hardware-level telemetry parsed directly into the Cortex Data Model (XDM) for automated remediation. Learn more in our latest blog. https://bit.ly/3SxSKiZ

As organizations rapidly scale the AI Factory, traditional security architecture is forcing a false choice between performance and protection—a compromise we can no longer accept. PANW Cortex XSIAM and Nvidia's DOCA ARGUS intergation offers 🔹 For CISOs: Protect your high-performance computing investments and maximize AI ROI without allowing security to become an operational bottleneck. 🔹 For DevOps & Infosec: NVIDIA BlueField DPUs delivers out-of-band, in-silicon visibility to stop threats without draining premium AI compute resources. ✒️ Learn more https://lnkd.in/gjw6fJYj

Through our new integration with the NVIDIA DOCA Argus framework, Cortex is bringing defense-in-depth straight to the silicon level. Operating out-of-band directly from the NVIDIA BlueField DPU, we deliver: ✅ Full visibility into AI node memory, file access, and network connections. ✅ 0% host agent overhead, keeping your compute resources dedicated to your business mission. ✅ Structured, hardware-level telemetry parsed directly into the Cortex Data Model (XDM) for automated remediation. Learn more in our latest blog. https://bit.ly/3SxSKiZ

Through our new integration with the NVIDIA DOCA Argus framework, Cortex is bringing defense-in-depth straight to the silicon level. Operating out-of-band directly from the NVIDIA BlueField DPU, we deliver: ✅ Full visibility into AI node memory, file access, and network connections. ✅ 0% host agent overhead, keeping your compute resources dedicated to your business mission. ✅ Structured, hardware-level telemetry parsed directly into the Cortex Data Model (XDM) for automated remediation. Learn more in our latest blog. https://bit.ly/3SxSKiZ

Through our new integration with the NVIDIA DOCA Argus framework, Cortex is bringing defense-in-depth straight to the silicon level. Operating out-of-band directly from the NVIDIA BlueField DPU, we deliver: ✅ Full visibility into AI node memory, file access, and network connections. ✅ 0% host agent overhead, keeping your compute resources dedicated to your business mission. ✅ Structured, hardware-level telemetry parsed directly into the Cortex Data Model (XDM) for automated remediation. Learn more in our latest blog. https://bit.ly/3SxSKiZ