lab NVIDIA Research · 1h ago · research

Symphony orchestrates sparse and dense tensors with hierarchical heterogeneous processing

from English

The article proposes Symphony, a hybrid programmable and specialized architecture designed to address memory system inefficiencies in current high-performance architectures like GPUs. It focuses on orchestrating data throughout the memory hierarchy to reduce unnecessary data movement and distance.

Symphony utilizes specialized reconfigurable units for roofline floating-point computations alongside data orchestration features such as address generation, data filtering, and sparse metadata processing.
Computation resources are distributed throughout the on-chip memory hierarchy to support both programmable and specialized tasks.
The architecture aims to handle a mix of sparse and dense algorithms that domain-specific accelerators typically cannot manage.

Symphony matches non-programmable ASIC performance on sparse tensor algebra while providing 31x improved runtime and 44x improved energy over a comparably provisioned GPU.

Importance 1/3 Trust 3/3 NVIDIA Research Hardware & chips Research paper

Read original

media Hugging Face Forums · 22d ago · 22 views

Patience Strong proposes building neural networks as semiconductor devices

Patience Strong argues that current neural networks consume excessive compute power by emulating circuits in software rather than implementing them as physical hardware. The paper suggests replacing this emulation with a semiconductor device where weights and biases are stored in non-volatile memory.

lab NVIDIA Research · 44d ago · 10 views

Nvidia presents 2M algorithm reducing complex matrix multiplication to two real GEMMs

A new paper introduces the 2M multiplication algorithm, which reduces the cost of complex matrix multiplication from three real General Matrix Multiplications (GEMMs) to just two. This approach applies to matrices with integer real and imaginary parts by utilizing quadratic time pre- and post-processing steps.

arxiv arXiv cs.AI · 65d ago · 7 views

HAMON: Passive Optical Forecasting for Long-Horizon Time-Series

HAMON uses passive optical components to perform long-horizon time-series forecasting, outperforming top digital models on ETTm2 across all horizons and on ETTh2 at all but the longest horizon. It achieves up to 14% lower MSE and relies on physical optical propagation without trainable digital layers, demonstrating that passive optical mixing can produce competitive forecasts.

arxiv arXiv cs.LG · 65d ago · 7 views

HAMON: Passive Optical Forecasting Core

HAMON uses passive optical diffraction to generate forecasts, outperforming digital baselines on ETTm2 at all horizons and ETTh2 at all but the longest horizon. It achieves up to 14% lower MSE and operates without trainable digital mixing, relying instead on physical optical propagation.

arxiv arXiv cs.AI · 4h ago · 1 view

Harness Continual Learning adapts agent state outside frozen model parameters

The authors introduce Harness Continual Learning (HCL), a paradigm where an agent improves by evolving its execution-facing harness rather than updating model parameters. This approach addresses the challenge of retaining previously acquired behavior while adapting to new experiences through components like prompts, memories, and routing rules.