Hot Chips 2026
PulseAugur coverage of Hot Chips 2026 — every cluster mentioning Hot Chips 2026 across labs, papers, and developer communities, ranked by signal.
2 day(s) with sentiment data
Arm AGI server CPU to challenge AMD/Intel in AI inference latency
Arm's new AGI server CPU, with its dual-chiplet design and integrated compute/I/O, is positioned to offer lower latency for AI workloads compared to competing offerings from AMD and Intel. This design choice directly addresses memory bandwidth and latency concerns, suggesting Arm is making a strong play for AI inference markets where these factors are critical.
Emergence of specialized memory tiers (HBM vs. HBF) for AI workloads
The Hot Chips 2026 conference highlights a divergence in memory solutions for AI. While HBM remains the standard for high-throughput inference, High Bandwidth Flash (HBF) is emerging as a viable option for storing very large models due to its capacity and cost, albeit with lower speeds. This suggests a future where AI systems may utilize a tiered memory approach.
IBM's dual-ISA core to enable hybrid AI deployments on mainframes
IBM's new dual-ISA core, capable of running both ARM and z/Architecture natively, could unlock significant opportunities for AI on mainframe systems. This allows existing mainframe infrastructure to leverage the extensive ARM AI software ecosystem without a complete hardware overhaul, potentially accelerating AI adoption in enterprise environments that rely on mainframes.
-
OpenAI discusses AI-driven chip design and agent safety
Tom's Hardware is highlighting its AI Chip Design Week, featuring an interview with OpenAI's hardware VP, Richard Ho. The discussion covers "Jalapeño," an inference ASIC co-developed with Broadcom, and how AI was used i…
-
OpenAI's custom Jalapeño AI chip prioritizes internal use, hints at future rollout
OpenAI has developed a custom AI inference ASIC named Jalapeño, primarily for its internal use to meet growing compute demands. While the chip is designed to be programmable and capable of running various models beyond …
-
Tom's Hardware offers free AI Chip Design Week access
Tom's Hardware is offering free access to its AI Chip Design Week content from September 28 to October 2, requiring only a free membership sign-up. The week's coverage includes an interview with OpenAI's hardware lead, …
-
Samsung plans integrated compute and memory with zHBM roadmap
Samsung has outlined a three-phase strategy to integrate logic and compute capabilities directly into its High Bandwidth Memory (HBM) architecture. The company's zHBM approach aims to place DRAM directly on top of the p…
-
Oxmiq Labs proposes High Bandwidth Flash for AI inference capacity
Oxmiq Labs is proposing High Bandwidth Flash (HBF) as a new capacity tier for AI inference, aiming to offer significantly more storage at a comparable cost to High Bandwidth Memory (HBM). Their presentation at Hot Chips…
-
Intel unveils custom accelerator to challenge Nvidia, AMD at Hot Chips 2026
Servethehome has published a series of articles from Hot Chips 2026 detailing advancements in custom accelerators. Intel presented a new candidate accelerator designed to compete with Nvidia DGX and AMD's 395+/495+ offe…
-
Cloud operators face memory crunch, may spend 68% capex on DRAM/NAND
Cloud operators are facing a significant increase in capital expenditure due to soaring prices for dynamic random-access memory (DRAM) and NAND flash. This trend, highlighted at Hot Chips 2026, suggests that while High …
-
Nvidia touts DSX MaxLPS for maximizing AI compute density within fixed power budgets
Nvidia is promoting its DSX MaxLPS power management technology, designed to optimize compute density within fixed data center power budgets. At Hot Chips 2026, the company highlighted how this system, when paired with i…
-
Fujitsu unveils Monaka AI CPU with stacked cache design
Fujitsu has unveiled its new Monaka server CPU, designed for AI performance and power efficiency in green data centers. The chip features a unique three-die stack: a 2nm core die, a 5nm die for the entire last-level cac…
-
High Bandwidth Flash offers capacity but limited speed for AI workloads
High Bandwidth Flash (HBF), a new memory format presented at Hot Chips 2026, offers significantly more capacity than traditional High Bandwidth Memory (HBM) at a comparable cost. However, HBF's lower bandwidth makes it …
-
d-Matrix unveils 3D DRAM AI accelerator with 100 TB/s bandwidth
d-Matrix has unveiled its Raptor AI accelerator, which features a novel 3D stacking architecture. This design places a TSMC 4nm compute die directly on top of a custom-designed DRAM die, achieving an unprecedented 100 T…
-
Arm details AGI server CPU with dual-chiplet design for AI workloads
Arm has detailed its upcoming AGI server CPU, slated for release in late 2026. This processor features a dual-chiplet design, with each chiplet containing 70 Neoverse V3 cores and utilizing TSMC's N3P technology. A key …
-
Google unveils TPUv8 family with dedicated training and inference chips
Google has announced its eighth generation of Tensor Processing Units (TPUs), the TPUv8 family, which includes separate chips for training (TPU 8t) and inference (TPU 8i). This dual-chip strategy allows Google to optimi…
-
SambaNova details SN50 AI accelerator with focus on bandwidth utilization
SambaNova has unveiled new technical details about its SN50 RDU, a dedicated AI accelerator designed for high-efficiency and low-latency inference. The SN50 features a dataflow architecture with a large on-chip SRAM and…
-
Microsoft details Maia 200 AI accelerator for Azure data centers
Microsoft has unveiled details about its Maia 200 AI accelerator, designed for its Azure data centers. This second-generation chip, built on TSMC's 3nm process, features a 750W TDP and utilizes a Software Defined Local …
-
Cerebras launches CS-4 rack-scale system with WSE-3 Turbo accelerators
Cerebras has unveiled its new CS-4 rack-scale system, featuring the WSE-3 Turbo wafer-scale engine. This system is designed to significantly boost performance and energy efficiency, offering double the tokens and ten ti…
-
Samsung integrates logic into LPDDR5X memory for faster AI inference
Samsung has developed LPDDR5X-PIM, a new memory technology that integrates logic units directly into the memory chips. This Processing-in-Memory (PIM) approach aims to reduce the cost and power consumption associated wi…
-
Nvidia details 88-core Vera CPU for agentic AI workloads
Nvidia detailed its upcoming Vera CPU at Hot Chips 2026, highlighting its 88-core design and custom Olympus core architecture. The CPU features a novel spatial multithreading implementation and a high-bandwidth memory s…
-
OpenAI's Jalapeño chip shows superior inference performance over Nvidia
OpenAI has revealed initial performance data for its custom-designed "Jalapeño" inference chip, showcasing significant improvements in speed and power efficiency. Benchmarks indicate that Jalapeño outperforms competitor…
-
Intel unveils Xeon 7 'Diamond Rapids' with 256 P-cores and 1.28 GB cache
Intel has revealed details about its upcoming Xeon 7 'Diamond Rapids' server CPUs, slated for a 2027 release. These processors will feature up to 256 P-cores and a substantial 1.28 GB of last-level cache. The new archit…