Underfox

Physicist, Telecom Engineering lover, HPC Enthusiast. Prog Rock/Metal fan.

[Underfox @Underfox3]

Researchers have developed a new simulator to predict the throughput of basic blocks of all Intel Core μarchs released in the last decade, demonstrating to be more accurate than the predictions of state-of-the-art tools by more than an order of magnitude.

https://t.co/83UDBQSchX

[Underfox @Underfox3]

These findings represent a masive breakthrough in width-scaling rules using 2D nanoribbon transistors with enhanced performance at narrower channel widths, which is promising for the ultimate scaling of transistors.

[Underfox @Underfox3]

In this paper, researchers have demonstrated atomically thin monolayer and bilayer molybdenum disulfide nanoribbon transistors that break the width-scaling wall down to 15 nm.

https://t.co/l0pKVk1LpD

[Underfox @Underfox3]

The ultra-narrow nanoribbon transistors maintain the highest on/off ratios reported so far (10^6) for similar device dimensions, with improved mobility and threshold-voltage stability, indicating reduced edge scattering and depletion, along with stronger electrostatic control.

[Underfox @Underfox3]

The results show that CRAM-ER presents near-lossless accuracy with 10× better energy efficiency and a 2× improvement in EDP over the A100 GPU. Furthermore, CRAM-ER achieves up to 70× higher energy efficiency than CPUs and GPUs while reaching near-HBM2 throughput.

[Underfox @Underfox3]

In this paper, researchers proposed an error-resilient CRAM architecture for scalable in-memory matrix-vector multiplications, mitigating the impact of device-level errors and demonstrating high area and energy efficiency.

https://t.co/WKAgJYR5vw

[Underfox @Underfox3]

The proposed architecture enables parallel in-situ multiplications and error-resilient additions. Partitioning MACs between CMOS and MRAM at the bit level provides an optimal trade-off between area overhead and processing efficiency.

[Underfox @Underfox3]

In this paper, researchers have demonstrated that an off-the-shelf open-weight LLM, quantized to fit on a single GPU, suffices to drive a worm AI agent that gains privileged access to machines and replicates itself. #DeepLearning

https://t.co/1fQ7AMvbwY

[Underfox @Underfox3]

Researchers have proposed a bounded GPU-aware wait-free queue with explicit theorem-grounded progress guarantees, and a bounded GPU lock-free queue design that uses wave-batched fast paths to maximize throughput.

https://t.co/VzCi2eUvC1

[Underfox @Underfox3]

In this paper is presented a review on the recent progress in quantum-hardware-based simulations of condensed matter, primarily emphasizing gate-based digital quantum computer simulation, with analog experiments discussed as complementary benchmarks.

https://t.co/KfcEHc8UVY

[Underfox @Underfox3]

This unified formulation is enabled by modality-specific encoders, structured token arrangements, and a Mixture-of-Transformers backbone that couples autoregressive reasoning with diffusion-based generation.

[Underfox @Underfox3]

Finally, Nvidia have introduced Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture.

https://t.co/AhC5lE3Xkx

[Underfox @Underfox3]

Moreover, because the worm requires no commercial AI platform, centralized safety controls, such as service refusals or rate limiting, are structurally irrelevant. This results demonstrate that self-sustaining AI-driven cyber-threats are no longer theoretical.

[Underfox @Underfox3]

In this paper, researchers have demonstrated that an off-the-shelf open-weight LLM, quantized to fit on a single GPU, suffices to drive a worm AI agent that gains privileged access to machines and replicates itself. #DeepLearning

https://t.co/1fQ7AMvbwY

[Underfox @Underfox3]

Since the worm is powered by stolen compute, the attacker’s marginal cost per new infection is zero. This creates a destabilizing economic asymmetry between attackers and defenders.

[Underfox @Underfox3]

The results show that HiGS renders up to ∼15.8× faster than the original 3DGS and outperforms every other rasterizer evaluated, while preserving exact front-to-back alpha compositing.

[Underfox @Underfox3]

In this paper, NVIDIA researchers have proposed Hierarchically Tiled Gaussian Splatting (HiGS), a 3D Gaussian Splatting rendering architecture in which spatial partitioning and rasterization operate at different granularities.

https://t.co/nKtdNCLC9e

[Underfox @Underfox3]

This reshapes work decomposition from screen-area proportional to density proportional and eliminates the rasterizer tail effect inherent in single-tile-size pipelines.

[Underfox @Underfox3]

"G-LFQ achieved the highest peak throughput in several settings, while G-WFQ was the most robust across architectures and workload mixes, sustaining performance under contention and degrading more gracefully at high thread counts."

[Underfox @Underfox3]

Researchers have proposed a bounded GPU-aware wait-free queue with explicit theorem-grounded progress guarantees, and a bounded GPU lock-free queue design that uses wave-batched fast paths to maximize throughput.

https://t.co/VzCi2eUvC1

[Underfox @Underfox3]

The results across fixed-duration microbenchmarks on MI210 and MI300A, the bounded ring designs delivered the strongest overall performance and efficiency.