Tagged articles
memory wall
12 articles · Page 1 of 1
KAIST Laboratory presents a 99-page analysis of next-generation HBM architectures (HBM4–HBM8) for AI transformer workloads, detailing the memory-wall bottleneck, roofline modeling on V100, evolution from HBM1 to HBM3E, cooling progression from air to fluidic TSV, and six key technical directions including 3D near-memory computing and CXL integration.
3D stackingAI AcceleratorsCXL
0 likes · 7 min read

This article analyzes why linked lists often lose to arrays in real-world performance despite theoretical O(1) insertion advantages, exploring CPU cache effects, memory walls, Stroustrup's vector vs list experiments, and scenarios where linked lists still excel in kernels and hybrid structures.
Bjarne StroustrupC++Chandler Carruth
0 likes · 16 min read

This article evaluates High Bandwidth Flash (HBF) as a supplement to HBM for LLM inference through simulation, showing HBF can reduce GPU count and boost throughput for long-context and MoE models, but identifies three critical barriers: read bandwidth must match HBM, flash endurance needs 25× improvement, and write latency must be lowered for practical deployment.
HBFHBMHigh Bandwidth Flash
0 likes · 6 min read

The article analyzes how the growing memory‑wall bottleneck forces GPUs to idle while waiting for data, compares on‑chip SRAM and high‑bandwidth memory (HBM) as remedies, and examines HBM’s technical advantages, supply constraints, and divergent manufacturing routes that may turn it into a new limitation.
AI computeGPUHBM
0 likes · 6 min read


AI WakaMay 12, 2026 · Industry Insights
The article examines how photonic AI accelerators from Q.ANT, now deployed in a German supercomputing center, achieve over 50‑fold throughput gains and dramatically lower power consumption, offering a potential solution to the mounting energy, heat, and memory‑wall challenges of GPU‑driven data centers.
AI hardwareQ.ANTdata center
0 likes · 12 min read

As large‑model inference demand outpaces training, the decode stage hits a memory‑wall that GPUs cannot efficiently cross; SRAM’s on‑chip bandwidth and low‑energy access open a path forward, though capacity and process limits still pose challenges.
AI hardwareGPUSRAM
0 likes · 7 min read

With LLM context windows soaring to millions of tokens, the KV‑cache memory wall threatens scalable inference; Google’s TurboQuant tackles this by compressing KV data up to six‑fold without precision loss and accelerating attention up to eight‑fold, using PolarQuant and 1‑bit QJL techniques, reshaping hardware costs and edge AI possibilities.
AI inferenceKV compressionTurboQuant
0 likes · 25 min read

The rapid rise of AI workloads is reshaping server economics, exposing the memory wall, and spurring a shift toward high‑bandwidth memory, CXL‑based pooling, and advanced DIMM technologies, while manufacturers like SK Hynix, Samsung and Micron race to dominate the emerging market.
AI serversCXLDDR5
0 likes · 9 min read

The article analyzes how AI servers are shifting from pure compute power to storage‑centric designs, detailing the memory‑wall challenge, the rise of HBM and CXL technologies, vendor market shares, upcoming product roadmaps, and the broader supply‑chain opportunities shaping the AI hardware ecosystem.
AI serversCXLHBM
0 likes · 9 min read

The rapid growth of AI models is widening the gap between compute power and memory bandwidth, but the emerging Compute Express Link (CXL) interconnect offers lower latency, memory sharing, and flexible device topologies that can alleviate the memory‑wall bottleneck and reshape future data‑center architectures.
AI computeCXLdata center
0 likes · 10 min read

The article explains how the memory‑wall limitation of traditional von Neumann architectures hampers AI chip performance, describes two in‑memory computing approaches—circuit‑level modifications and new memory devices—highlights recent conference trends, and showcases a Chinese startup’s 8‑bit low‑power in‑memory AI chip that could enable ubiquitous AI on edge devices.
AI chipsin-memory computinglow-power AI
0 likes · 12 min read

This article summarizes a 2017 Gdevops talk that examines the evolution of many‑core processors and non‑volatile memory, explains how memory‑wall effects impact modern DBMS performance, and presents architectural and algorithmic adaptations—including cache‑friendly structures, lock management, and write‑behind logging—to exploit new hardware.
Non-volatile Memorydatabasesmany-core
0 likes · 18 min read
