Computer news / Archive / July 2026
AMD Unveils Its Broadest AI Hardware Lineup Yet, Anchored by the Helios Rack System
July 26, 2026
AMD launched 6th Gen EPYC CPUs, Instinct MI400 Series GPUs, Helios AI racks, Ryzen AI Embedded X100, and Kria AI robotics platform.
Helios rack (72 MI455X GPUs + 18 EPYC "Venice" CPUs) claims 30% more tokens per dollar than the leading competitor.
OpenAI, Anthropic, Meta, Microsoft and Oracle are adopting Helios; OpenAI targets Q4 2026 deployment.
Roadmap through 2030 revealed: "Zen 7" EPYC (2028), "Zen 8" EPYC (2030), Instinct MI500 (2027) and MI600 (2028).
New MI430X (288 TFLOPS FP64, HPC/sovereign AI) and MI350P (token-economics-focused) GPUs also launched.
Nvidia Deploys Its Own Vera Processor to Accelerate In-House Chip Engineering
July 26, 2026
Uses Vera CPU internally for chip-design work, tuning Cadence Jasper and Synopsys VCS verification tools for it.
Early benchmarks showed up to 1.5x higher performance on selected verification/simulation workloads.
Vera packs 88 custom Nvidia Olympus cores, LPDDR5X memory, and a 2nd-gen Scalable Coherent Fabric.
Faster verification lets engineers test more design alternatives per development cycle.
Successor "Rosa" CPU (built on the Rigel core) is already planned.
Reconfigurable Optical Chip Lets Engineers Dial In Light's Speed for Future Computing Hardware
July 20, 2026
Seoul National University/University of Seoul built a programmable CRIT-based photonic circuit that slows light on demand.
Unlike fixed traditional CRIT devices, two controllable loop couplers let it adjust delay, bandwidth, and signal shape post-fabrication.
Silicon nitride platform simulations showed reliable operation despite losses, phase errors, and thermal crosstalk.
One chip could combine synchronization, delay lines, buffering, and frequency conversion.
Published in Advanced Science; targets energy/complexity reduction in AI servers and data centers.
NVIDIA's New Jetson Thor Modules Bring Blackwell Compute to Smaller, Cheaper Edge Robots
July 17, 2026
NVIDIA launched two new Thor-architecture modules, Jetson/IGX T3000 and Jetson T2000, aimed at mass-market robotics and edge AI deployment.
The T3000 packs 865 FP4 teraflops, an NVIDIA Blackwell GPU, an eight-core Arm Neoverse CPU and 32GB of LPDDR5X memory in roughly half the size and power of the previous T5000 module.
The T2000 targets a broader range of lower-cost edge systems with 400 FP4 teraflops and 16GB of memory.
New "agent skills" automate memory optimization across the Jetson lineup, letting partners like UBTech and Agile Robots cut memory use by up to 15GB.
NVIDIA also introduced Cosmos 3 Edge, a 4-billion-parameter world foundation model compatible with the Thor platform for on-device robot reasoning.
Intel Becomes First Chipmaker to Ship High-Volume Logic Chips Made With ASML's Next-Gen EUV Tools
July 15, 2026
Intel Foundry has entered high-volume manufacturing for a subset of Intel Core Ultra Series 3 ("Panther Lake") processors using ASML's High NA EUV lithography.
Specific Intel 18A process layers are now dual-qualified on High NA EUV at Intel's Oregon site, shipping to customers at yields matching the older NXE platform.
ASML says this is the first time the industry has shipped high-volume logic product using High NA EUV technology.
Intel was earlier the first company to install and pass acceptance testing of ASML's second-generation High NA system, the TWINSCAN EXE:5200B.
Intel and ASML say they will keep incorporating High NA EUV into future process nodes based on customer needs.
Machine Learning helped to Discover Two New Kagome-Lattice Superconductors
July 15, 2026
The SuperC consortium, led by Aalto University's Päivi Törmä, used machine learning to rapidly screen vast numbers of elemental combinations and identify promising superconductor candidates for detailed quantum calculation.
The approach led to the discovery of two new superconductors, YRu3B2 and LuRu3B2, whose properties come from electrons forming flat bands in a kagome lattice structure.
Rice University researchers, led by Emilia Morosan, synthesized the predicted compounds and experimentally confirmed both are genuine superconductors.
Of roughly 7,000 known superconductors discovered mostly by trial and error, only about 20 had previously been theoretically predicted in advance, reflecting how computationally demanding the search normally is.
Published in Physical Review Research, the team says the machine-learning pre-screening method could eventually process billions of candidate materials, accelerating the search for a room-temperature superconductor targeted by 2033.
Scientists Revive an Old Optics Trick to Build Light Patterns for Future Computing
July 15, 2026
NTU Singapore researchers created optical skyrmions — stable, hedgehog-like swirling light patterns — by shining a laser at a small disc, reviving the 200-year-old "Poisson spot" diffraction effect instead of using costly engineered metamaterials.
The simplified method makes optical skyrmions, seen as promising building blocks for data storage, communications, and computing, far more accessible to researchers.
The team found their setup naturally generates four distinct types of skyrmions simultaneously — spin, Stokes, electric field, and magnetic field — within a single light spot.
Led by Asst. Prof. Shen Yijie and published in the journal Optica, the work provides a new way to compare how different light properties form related topological structures.
Researchers say the approach could advance photonics, materials science, and next-generation computing by lowering the technical barrier to studying complex light structures.
NVIDIA Pitches Its Vera CPU as the Answer to AI Agents' Need for Speed
July 15, 2026
NVIDIA describes Vera as a "max single-threaded CPU at scale," built around its custom Olympus core with 50% higher instructions per cycle than the earlier Grace CPU.
Vera offers up to 1.2TB/s of LPDDR5X memory bandwidth and 3.4TB/s of core-to-core bandwidth across its 88 cores.
In loaded agentic workloads, NVIDIA says Vera delivers 1.8x the sustained per-core performance of comparable x86 chips.
Perplexity tested Vera on real coding workflows and measured jobs completing about 1.5x faster than on x86, with concurrent sandboxes starting up to 1.9x faster.
NVIDIA also previewed a successor, the Rosa CPU with a new Rigel core, continuing its CPU roadmap for agentic AI.
Anthropic's Claude Models Go Live on Nvidia's Newest Blackwell Ultra Chips
June 29, 2026
Anthropic's Claude models in Microsoft Foundry, running on Nvidia GB300 Blackwell Ultra GPUs hosted on Microsoft Azure, are now generally available.
The deployment uses Nvidia GB300 NVL72 systems paired with Nvidia Quantum-X800 InfiniBand networking.
Nvidia says it is integrating its developer tools into Anthropic's stack through "NVIDIA verified agent skills" to give Claude-based agents domain-specific capabilities.
Enterprises can run Claude agents on Azure using Nvidia's Secure Agent Workspace Reference Design, described as a blueprint for running agents in a governed environment with infrastructure-level identity and access controls.
The move builds on a strategic three-way partnership between Microsoft, Nvidia, and Anthropic that was first announced in November.
OpenAI and Broadcom Reveal Jalapeño, Their First Jointly Built Inference Chip
June 24, 2026
Jalapeño is described as OpenAI's first custom "Intelligence Processor," architected specifically around OpenAI's approach to LLM inference.
The chip was co-developed from initial design to manufacturing tape-out in just nine months, which the companies call the fastest ASIC development cycle achieved in advanced semiconductors.
The architecture is designed to reduce data movement and balance compute, memory and networking resources to push realized utilization closer to theoretical peak.
Broadcom is contributing silicon implementation and networking technology, including its Tomahawk networking silicon, to bring the platform to production scale.
Jalapeño is the first step in a planned multi-generation compute platform, with initial deployment targeted by the end of 2026 and Celestica providing board, rack and system integration.
Contact us
tamos.projects@gmail.com
© 2021-2026 Computers.lt