HBM3E
PulseAugur coverage of HBM3E — every cluster mentioning HBM3E across labs, papers, and developer communities, ranked by signal.
7 day(s) with sentiment data
-
NVIDIA NIXL and LMCache slash LLM latency with Disaggregated Prefill-Decode
A new technique called Disaggregated Prefill-Decode, utilizing NVIDIA's NIXL and LMCache, significantly improves LLM inference performance, particularly for long-context tasks. This method separates the compute-intensiv…
-
NVIDIA revises Rubin GPU specs, reducing memory and NVLink bandwidth
NVIDIA has revised the specifications for its upcoming Rubin GPU, reducing the HBM4 memory bandwidth and NVLink interconnect bandwidth compared to earlier projections. The Rubin GPU will now offer 19.2 TB/second of memo…
-
China's CXMT pushes HBM3E production despite low yields, aiming for AI supply chain independence
China's Changxin Memory Technologies (CXMT) is proceeding with mass production of HBM3E memory despite a 20% yield rate, a level that would typically be financially unsustainable for most chip manufacturers. This aggres…
-
New benchmarks and research push AI coding agents toward complex, long-horizon tasks
Several new research papers and benchmarks are being released to evaluate the capabilities of AI coding agents, particularly in handling long-horizon tasks and complex software engineering workflows. These efforts aim t…
-
China's CXMT to launch HBM3E production amid low yield rates
ChangXin Memory Technologies, China's largest DRAM manufacturer, is preparing to launch limited production of HBM3E chips within two months. However, only 20-30% of these chips are currently meeting specifications, a st…
-
China's CXMT to produce AI memory chips despite low yields
Changxin Memory Technologies (CXMT) in China is preparing to manufacture HBM3E, a high-bandwidth memory chip crucial for AI processors. This initiative proceeds despite significant yield issues, with reports indicating …
-
TSMC eyes 20% CoWoS price hike amid AI chip demand surge · 1 source tracked
TSMC is reportedly considering a 20% price increase for its advanced CoWoS packaging technology due to surging demand for AI chips and extended equipment lead times. This increase is driven by the complex packaging requ…
-
ASI may be an economic loop, not a robot uprising, author suggests
Artificial Superintelligence (ASI) may manifest not as a robot uprising, but as an economic loop where money is converted into tokens to justify hardware purchases, which in turn generate more tokens. This cycle, remini…
-
GPU VRAM Sizing Error for MoE Models Detailed by Logic42
A technical analysis by Logic42 highlights a common miscalculation in sizing GPU clusters for large Mixture-of-Experts (MoE) models like DeepSeek-V4.1-Flash and Moonshot Kimi K2.5. The error involves underestimating VRA…
-
CXMT begins HBM3E risk production, samples with T-Head and Cambricon
ChangXin Memory Technologies (CXMT) has initiated risk production for its HBM3E memory chips. Initial samples are currently undergoing qualification by Alibaba's T-Head and Cambricon. While specific details on yields, p…
-
Huawei unveils AI chip; Meta, Google advance models; AI agents used in ransomware
Huawei has unveiled its new Threadripper Halo chip, designed for AI researchers and featuring substantial memory bandwidth and capacity. In other AI news, Meta is preparing to release its Muse model with open weights, a…
-
AMD launches Threadripper Halo for local AI, Hugging Face deemed too vital for Nvidia acquisition
AMD has launched its Threadripper Halo workstation, designed for researchers needing significant local AI processing power with up to 576 GB of HBM3e memory. In other AI news, Hugging Face is considered too vital to be …
-
AMD launches Threadripper Halo Station for local AI development · 4 sources tracked
AMD has unveiled its Threadripper Halo Station, a high-performance AI workstation designed for local model training and inference. This powerful machine features a 96-core Threadripper Pro 9995WX CPU and dual liquid-coo…
-
Huawei unveils Atlas 950 SuperPoD servers with 1,024 NPUs and 256TB memory · 4 sources tracked
Huawei has showcased its Atlas 950 SuperPoD servers, a system comprising 1,024 NPUs designed for AI workloads. The system features extensive inter-rack connectivity, with each of the 32 LRS units utilizing 4x100G connec…
-
China's ChangXin Memory Technologies starts HBM3E production, challenging US and Korea
ChangXin Memory Technologies has initiated small-scale production of HBM3E memory chips, aiming to challenge the established dominance of South Korea and the United States in this critical technology sector. This move i…
-
China's CXMT begins HBM3E production, aiming to close AI memory gap · 2 sources tracked
China's leading memory manufacturer, ChangXin Memory Technologies (CXMT), has begun risk production of HBM3E chips, a critical component for advanced AI processors. While currently a generation behind industry leaders, …
-
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 …
-
AI hardware advances: Cerebras stacks DRAM, HBM faces growing costs
At Hot Chips 2026, leading memory and chip manufacturers discussed the future of AI hardware, focusing on advancements in wafer-scale processing and High Bandwidth Memory (HBM). Cerebras unveiled its roadmap for stacked…
-
Intel GPU challenges NVIDIA, Apple sues OpenAI, and agentic ransomware emerges
Intel has released a new GPU, the Arc Pro B70, which significantly outperforms NVIDIA's flagship RTX 5090D and RTX 4090D in AI inference tasks at a fraction of the cost. While Intel's hardware shows promise for inferenc…
-
GLM-5.2 deployment on 8x B200 GPUs favors NVFP4 for optimal throughput
A technical analysis reveals that deploying the GLM-5.2 model on 8x NVIDIA B200 GPUs is most efficient using NVFP4 precision across four GPUs, rather than the more intuitive FP8 precision across all eight. This configur…