Better Hardware Could Turn Zeros into AI Heroes
Larger AI models are increasing in size and capability but also in energy consumption and computational demands. Sparse computing, which leverages the abundance of zero values in model parameters, offers a way to reduce energy use and speed up processing. Researchers at Stanford have developed hardware that efficiently handles sparse workloads, significantly improving performance and energy efficiency.
- ▪Large language models are growing in size, with some containing trillions of parameters, increasing their energy and computational demands.
- ▪Sparsity refers to the prevalence of zero or near-zero values in AI models, which can be exploited to skip unnecessary calculations and reduce memory usage.
- ▪Stanford researchers created hardware that, on average, uses one-seventieth the energy of a CPU and performs eight times faster by fully leveraging sparsity.
- ▪Current mainstream hardware like CPUs and GPUs does not efficiently utilize sparsity, requiring redesigns across the hardware and software stack.
- ▪Sparsity can occur naturally, as in social network graphs, or be induced by setting a large percentage of model parameters to zero without sacrificing accuracy.
Opening excerpt (first ~120 words) tap to expand
ComputingAIMagazineFeature Better Hardware Could Turn Zeros into AI Heroes Sparse computing enables leaner, faster AIOlivia HsuKalhan Koul28 Apr 20269 min readVerticalPetra PéterffyPurpleWhen it comes to AI models, size matters.Even though some artificial-intelligence experts warn that scaling up large language models (LLMs) is hitting diminishing performance returns, companies are still coming out with ever larger AI tools. Meta’s latest Llama release had a staggering 2 trillion parameters that define the model.As models grow in size, their capabilities increase. But so do the energy demands and the time it takes to run the models, which increases their carbon footprint.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at IEEE Spectrum — AI.
Discussion
0 commentsMore from IEEE Spectrum — AI
