Author page description
August 27
- Huawei Presents UB-Mesh Interconnect for Large AI SuperNodes at Hot Chips 2025
➀ Huawei introduced its UB-Mesh interconnect technology at Hot Chips 2025, targeting large-scale AI SuperNodes with up to a million chips;
➁ The architecture employs a unified protocol and hybrid topology to reduce costs, improve scalability, and ensure reliability for data center-scale deployments;
➂ Key innovations include resilient optical links and hierarchical redundancy to address high error rates and node failures in gigaWatt AI data centers.
- Marvell Shows Dense SRAM Custom HBM and CXL with Arm Compute at Hot Chips 2025
➀ Marvell showcased advanced memory technologies at Hot Chips 2025, including dense 2nm SRAM with 17x higher bandwidth density than standard IP, leveraging TSMC's 2nm process and optimized Vmins for lower power consumption;
➁ The company introduced custom HBM solutions using die-to-die interfaces to reduce on-chip area and power, collaborating with major HBM suppliers to enhance AI accelerator compatibility;
➂ Marvell also demonstrated high-capacity DDR memory expanders with integrated Arm Neoverse v2 CPUs and hardware security, improving latency and bandwidth for large-scale AI workloads.
- NVIDIA Co-Packaged Optics with Silcion Photonics for Switching and Spectrum-XGS Scale-Across
➀ NVIDIA unveiled its co-packaged silicon photonics switches and Spectrum-X Ethernet technology at Hot Chips 2025, targeting gigawatt-scale data center networks for AI workloads;
➁ Spectrum-X reduces network jitter and improves NCCL performance, enabling efficient multi-tenant AI training with 1.9x better scale-out performance compared to traditional Ethernet;
➂ The Spectrum-XGS "scale-across" solution extends AI training across multiple data centers through distance-aware algorithms, with CoreWeave likely being the first deployment partner.
August 26
- NVIDIA GeForce RTX 5090 and the Age of Neural Rendering at Hot Chips 2025
➀ NVIDIA showcased its Blackwell architecture at Hot Chips 2025, emphasizing advancements in neural rendering and machine learning (ML)-based graphics to enhance realism while optimizing compute efficiency;
➁ The RTX 5090 features GDDR7 memory for higher bandwidth, FP4 precision for ML workloads, and an AI Management Processor to dynamically balance graphics and AI tasks;
➂ New techniques like shader execution reordering and frame-generation via ML reduce power consumption, while Universal MIG enables efficient GPU resource partitioning for multi-client scenarios.
- AMD RDNA 4 GPU Architecture at Hot Chips 2025
➀ AMD presented its RDNA 4 GPU architecture at Hot Chips 2025, highlighting enhancements in raytracing, AI/ML capabilities, and media/display engines, with a focus on future gaming workloads;
➁ Raytracing performance doubled through BVH throughput improvements, hardware instance transformation, and oriented bounding boxes, achieving ~2x gains over RDNA 3;
➂ New features include dynamic register allocation for efficient resource use, AV1 B-frame encoding, FP8 support for AI workloads, and modular SoC design for flexible GPU configurations.
- Microsoft Azure Hardware Security to Help Thwart the World’s 3rd Largest GDP
➀ Microsoft unveiled its Azure hardware security architecture at Hot Chips 2025, emphasizing protection for multi-tenant cloud environments through decentralized Hardware Security Modules (HSMs) integrated into every server;
➁ The company introduced custom ASIC-based Azure Integrated HSMs to eliminate centralized TLS handshakes, alongside its Secure Future Initiative and open-source Caliptra 2.0 silicon root of trust;
➂ With cybercrime projected to exceed $10T in 2025, Microsoft highlighted its 34,000 security engineers and global infrastructure (70+ regions, 400+ data centers) as critical defenses against this economic-scale threat.
- Intel Xeon Clearwater Forest with 288 Cores on Intel 18A at Hot Chips 2025
➀ Intel unveiled its next-gen 288-core Xeon Clearwater Forest processor at Hot Chips 2025, built on Intel 18A process and 3D packaging technology, significantly outperforming the Sierra Forest generation with enhanced cache, E-cores, and memory bandwidth.
➁ The processor utilizes an all-E-core design for power-efficient multi-threaded workloads, featuring architectural upgrades including a 9-wide decoder (up from 6-wide), doubled execution ports (26 ports), and a 17% IPC uplift in SPECint 2017 benchmarks.
➂ Leveraging 3D stacking (Foveros Direct 3D) and advanced power delivery (BSPDN), Clearwater Forest boasts 1300GB/s DDR5-8000 memory bandwidth per socket and claims a 3.5x rack-level performance-per-watt gain over Sierra, targeting data center power efficiency.
August 25
- NVIDIA Jetson AGX Thor Developer Kit Hands-on Blackwell for Robotics
➀ The NVIDIA Jetson AGX Thor Developer Kit brings Blackwell architecture to robotics with 25x AI performance improvement over previous gen;
➁ Features 4x 25GbE via QSFP28, advanced cooling, and 1TB SSD with dedicated heatsink, but omits PCIe slot from prior models;
➂ Targets robotics developers with $3,499 kit for prototyping before deploying module-based systems.
- Rebellions REBEL-Quad UCIe and 144TB HBM3E Accelerator at Hot Chips 2025
➀ Rebellions unveiled its REBEL-Quad AI accelerator at Hot Chips 2025, featuring four ASICs with 144GB HBM3E memory and UCIe chiplet interconnects.
➁ The PCIe Gen5 accelerator demonstrated live inference of Llama 3.3 70B models at 35.5ms/token using Samsung SF4X and CoWoS-S packaging.
➂ This marks a significant commercial implementation of UCIe technology, showcasing inter-chip communication advancements in AI hardware design.
August 23
- InnoDisk Shows DDR5-12800 MRDIMMs at FMS 2025
➀ InnoDisk showcased DDR5-12800 MRDIMM samples at FMS 2025, targeting future server platforms;
➁ The MRDIMMs offer capacities up to 128GB and are designed to address memory bandwidth bottlenecks in high-core-count servers;
➂ Intel's current Xeon 6 processors support MRDIMMs, with next-gen platforms expected to double throughput via DDR5-12800 and 16-channel configurations.
August 21
- ASUS ESC A8A-E12U 8x AMD Instinct MI325X GPU Server Review
➀ The ASUS ESC A8A-E12U is a 7U server featuring dual AMD EPYC processors and eight AMD Instinct MI325X GPUs with a total of 2TB HBM3e memory, designed for high-performance AI workloads;
➁ The system supports up to ten NVMe SSDs, eleven PCIe slots, and employs a 5+1 redundant 3kW PSU design, alongside a streamlined air-cooling system with ten hot-swappable front fans;
➂ It offers modular tray-based serviceability for GPUs and PCIe components, with front I/O including dual 10Gbase-T ports and a Q-Code display for diagnostics.
August 18
- Samsung 256TB MVP NVMe SSD at FMS 2025 and More
➀ Samsung showcased its 256TB MVP NVMe SSD at FMS 2025, featuring redesigned architecture for higher capacity and improved performance;
➁ HBM4 memory was highlighted as a critical component for next-gen AI accelerators, alongside CXL-based memory modules (CMM-D) to address server memory bandwidth challenges;
➂ The company also demonstrated flexible storage solutions like the PM9G3 SSD series, emphasizing innovation in NAND stacking and form factor adaptability.
August 17
- Phison Pascari D205V 122.88TB NVMe SSD at FMS 2025
➀ Phison unveiled the Pascari D205V PCIe Gen5 NVMe SSD at FMS 2025, featuring a massive 122.88TB capacity and delivering up to 14.7GB/s read and 3.2GB/s write speeds with 0.3 DWPD endurance.
➁ The SSD is offered in both U.2 2.5" and E3.L 1T form factors, with the latter enabling higher storage density for servers and AI clusters requiring multi-petabyte configurations.
➂ As the industry transitions toward PCIe Gen6, the E3.L form factor is positioned to replace U.2, accommodating future high-capacity SSD designs and supporting AI-driven demand for large-scale storage solutions.
August 16
- ASRock Rack AMPONED8-2T BCM AmpereOne Motherboard at FMS 2025
➀ ASRock Rack unveiled the AMPONED8-2T/BCM motherboard at FMS 2025, supporting Arm-based AmpereOne CPUs for servers;
➁ The motherboard features eight DDR5-5200 RDIMM channels, six PCIe Gen5 x16 slots, dual 10Gbase-T ports via Broadcom BCM57416, and a CEB form factor for chassis compatibility;
➂ Targets users seeking PCIe Gen5/DDR5 platforms as an upgrade path from older Ampere Altra systems.
August 14
- GigaPlus GP-S25-1602 Review A Cheap 16-port 2.5GbE and 2-port 10G Switch
➀ The GigaPlus GP-S25-1602 offers 16x2.5GbE ports and 2x10G SFP+ ports as an unmanaged silent switch priced at $160-170;
➀ Its internal design uses four switch chips providing 120Gbps total capacity, differing from typical single-chip configurations;
➂ While functional for basic use, its unique architecture may impact performance in high-demand scenarios.
- GigaPlus GP-S25-1602 Review A Cheap 16-port 2.5GbE and 2-port 10G Switch
➀ The GigaPlus GP-S25-1602 is a budget-friendly 16-port 2.5GbE and 2-port 10G SFP+ switch with low power consumption and silent operation;
➀ Its internal design uses four daisy-chained 30Gbps switch chips connected via 10Gbps links, explaining its 120Gbps switching capacity but revealing performance limitations in specific traffic scenarios;
➂ Performance tests show acceptable throughput in optimal configurations but significant bottlenecks when crossing chip boundaries.
August 13
- Blackmagic Cloud Store Mini Review A M.2 SSD NAS Made Too Simple
➀ The Blackmagic Cloud Store Mini is a compact 8TB NAS featuring four pre-installed M.2 SSDs, designed for portability and simplicity;
➁ It offers multiple connectivity options, including USB-C for direct computer access, dual Ethernet ports (10GbE and 1GbE), and an HDMI output for system monitoring;
➂ Despite its user-friendly approach, the device introduces enhanced security features, addressing feedback from earlier models.
August 12
- NVIDIA RTX Pro 4000 SFF Blackwell Edition and RTX Pro 2000 Blackwell Announced
➀ NVIDIA announced two new workstation GPUs: RTX Pro 4000 SFF Blackwell Edition (dual-slot, low-profile) with 24GB GDDR7, 8960 CUDA cores, and 70W TDP;
➁ The RTX Pro 2000 Blackwell offers 16GB memory, 4352 CUDA cores, and reduced NVENC/NVDEC engines for cost-sensitive applications;
➂ Both target SFF workstations, with the Pro 4000 Blackwell emphasizing compute performance and the Pro 2000 serving display/output needs.
August 11
- Kioxia Shows LC9 for 8PB or More in 2U and 32-Layer Die Stacked NAND at FMS 2025
➀ Kioxia showcased its 245TB LC9 SSD, enabling 8PB+ storage in 2U server chassis, and demonstrated 32-layer stacked BiCS 8 NAND technology at FMS 2025;
➁ A cross-section view of the 32-die stacked BiCS 8 package highlighted advancements in high-density NAND production;
➂ Kioxia's CM9 PCIe Gen5 SSD was shown supporting five NVIDIA H100 GPUs, exceeding typical Gen5 SSD performance for AI workloads.
August 10
- PCIe 8.0 Announced by the PCI-SIG Will Double Throughput Again
➀ PCI-SIG announced PCIe 8.0, targeting a 2028 release, which will double the throughput compared to PCIe 7.0;
➁ A PCIe 8.0 x16 link will offer 1TB/s bandwidth, a significant leap from PCIe 5.0's 128GB/s;
➂ The rapid evolution of PCIe standards is driven by growing demands from AI and high-performance computing applications.
August 8
- Microchip Adaptec SmartRAID 4300 A New Era of NVMe RAID Controller Without Drive Connectivity
➀ Microchip unveiled the Adaptec SmartRAID 4300 at FMS 2025, a PCIe Gen4 x16 NVMe RAID controller without onboard drive connectivity, marking a departure from traditional designs.
➁ The controller performs XOR parity calculations on-device, allowing data to flow directly from the CPU to NVMe SSDs, bypassing traditional bottlenecks and improving performance (e.g., 27M+ IOPS in Linux).
➂ This approach aligns with industry trends adopted by GRAID and Pliops, leveraging PCIe efficiency and requiring colocation with SSDs on the same CPU for optimal latency reduction.
- CWWK X86-P6 NAS Review an Intel N355 M.2 SSD Mini NAS
➀ The CWWK X86-P6 NAS is a compact, low-power device powered by Intel N355 or N150 processors, featuring dual M.2 SSD slots for storage;
➁ It includes dual 2.5GbE ports, USB 3.0, and an external USB fan for cooling, emphasizing quiet operation and portability;
➂ The review highlights its suitability for travel and unexpected performance insights between the N355 and N150 variants.
August 5
- UCIe 3.0 Spec Released with Big Speed Up for Chiplets
➀ The UCIe 3.0 specification doubles data transfer rates to 64GT/s for chiplet-based designs, significantly improving bandwidth density and enabling faster interconnects for 2D/2.5D packaging;
➁ Introduces new features like continuous transmission protocols, enhanced runtime recalibration, and L2 exit handshake to optimize power efficiency and link reliability;
➂ UCIe 3.0 lays the foundation for future chiplet ecosystems, with commercial products expected around 2026-2028 as adoption by major semiconductor companies accelerates.
- Broadcom Jericho4 51.2Tbps AI Router Chip Now Shipping with 3.2Tbps HyperPorts
➀ Broadcom Jericho4 51.2Tbps AI router chip is now shipping, targeting large AI clusters with 3.2Tbps HyperPorts and 3nm process technology;
➁ Features 100Km+ RoCE deployment for long-distance AI cluster communication and 200G PAM4 signaling;
➂ Part of Broadcom's expanded switch/routing portfolio, including Tomahawk 6 and Tomahawk Ultra, for high-tier AI network infrastructure.
- Storage.AI Project by SNIA Looks to Re-frame AI Storage Discussion
➀ SNIA launched the Storage.AI project to optimize AI storage efficiency by integrating storage with high-performance networks like InfiniBand and Ethernet;
➁ The initiative proposes shifting storage to scale-out networks to enhance GPU utilization and reduce pre-processing bottlenecks;
➂ Key industry players including AMD, Dell, and Intel have joined, but NVIDIA's absence raises questions about future collaboration.
August 4
- Dell PowerEdge R6715 Review A Spiffy 1U AMD EPYC Server
➀ The Dell PowerEdge R6715 is a 1U server with direct liquid cooling for AMD EPYC processors, emphasizing efficient thermal management for high-density computing;
➁ It features flexible configurations including multiple drive bays, dual OCP NIC 3.0 slots, and PCIe Gen5 expansion options;
➂ The system is designed for modern data centers with standardized rack liquid cooling integration and redundant Titanium-level power supplies.
August 2
- Deploying AMD Instead of Arm in our Infrastructure 2025 Here is Why
➀ STH explains its choice of AMD EPYC over Arm in 2025 infrastructure, citing Arm's challenges in enterprise adoption, including installed base compatibility, hardware limitations, and insufficient software support;
➁ Power efficiency gains from Arm are now minimal compared to AMD/Intel x86 CPUs, especially with AI GPU servers dominating power consumption (e.g., AMD Instinct MI325X at 1kW/accelerator);
➂ Enterprise Arm adoption faces hardware scarcity and vendor neglect: NVIDIA Grace CPUs are outdated for general workloads, while major vendors like Dell/HPE offer limited Arm options (e.g., HPE ProLiant RL300 Gen11's market failure).
August 1
- Seagate Expansion 28TB External Hard Drive and the HAMR HDD Within Mini-Review
➀ Seagate's 28TB Expansion external hard drive integrates HAMR (Heat-Assisted Magnetic Recording) technology, offering high-capacity storage in a compact design with USB 3.2 and external power support;
➁ Opening the enclosure reveals a Seagate Barracuda 28TB HAMR drive, typically not sold separately and possibly rebadged from enterprise-grade Exos or IronWolf Pro HDDs;
➂ Shucking the external drive provides significant cost savings ($240 less than standalone 28TB HDDs), though sacrificing warranty coverage.
July 30
- Enfabrica Elastic Memory Fabric System aka EMFASYS Launched
➀ Enfabrica launched the Elastic Memory Fabric System (EMFASYS), integrating its 3.2Tbps ACF-S SuperNIC with CXL memory to create a high-bandwidth RDMA solution for AI clusters;
➁ The system consolidates networking (RDMA) and memory expansion (CXL) into a single chip architecture, eliminating multi-chip complexity and enabling memory pooling over 144 CXL lanes;
➂ EMFASYS addresses AI memory scaling limitations by allowing flexible, large-scale memory access beyond traditional HBM/DRAM, with compatibility for devices like Astera Labs CXL modules.
July 29
- Micron 9650 PCIe Gen6 SSD Announced with Micron 6600 ION 122TB and 7650 SSDs
➀ Micron announced three new SSDs: the 9650 PCIe Gen6 (up to 30.72TB) for AI servers with 28GB/s reads and liquid cooling support;
➁ The 6600 ION offers massive 122TB capacity in E3.S form factor, targeting AI data storage;
➂ The mainstream 7650 PCIe Gen5 SSD focuses on low latency and capacities down to 1.92TB, replacing smaller Gen5 drives.