Technology / WORLD

Ember changes data placement when workloads exceed GPU memory

Researchers at the Technical University of Munich developed Ember, a memory manager that tracks which data should remain on a GPU when its memory fills. In the paper’s data-intensive workloads, it reported higher speeds than standard Unified Virtual Memory on three NVIDIA GPUs.

The source does not state the event date; the period shown is the source review period.

Reading sizeSources ↓Share story ↗
Conceptual illustration of technology, processors and connected information.Illustration
Symbolic illustration

Applications that exceed GPU memory must move data between host and device memory. Standard page migration can slow down when it repeatedly moves data that is not useful for the current workload. The adaptive Ember variant maps regions to host memory by default, promotes frequently accessed regions using GPU hardware counters, and reclaims colder regions before device memory fills.

For the static Ember variant at about 1.5× memory oversubscription, the authors report geometric-mean speedups over stock UVM across six data-intensive applications of 9.12× on RTX 5080, 6.83× on A40 and 10.62× on H200. They also tested Qwen2.5 14B and Qwen3 32B generation with GPUs capped at 16 GiB. These are results for the named hardware and workloads, not a general production-performance guarantee. Some baseline UVM runs hit the 1,800-second cap, so speedups in affected comparisons are lower bounds.

Data

Table

GPU memory experiment results

Speedups apply only to the GPUs and workloads specified in the paper.

36Geometric-mean speedup: static Ember / stock UVM
Methodology

Values are geometric-mean speedups of the static Ember variant over stock UVM across six applications, measured on three GPUs at about 1.5× memory oversubscription. Some baseline UVM runs hit the 1,800-second cap, so speedups in affected comparisons are lower bounds.

Updated:

Structured coverage

Subjects and places

Subject proposals

Geographic references

Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.

  • Germany · Place role: Coverage · Place type: Country or area · Location precision: Regional
    Evidence
    • Explicit preserved editorial country/region; no city or national centroid inferredarxiv.org ↗
Taxonomy version: IPTC Media Topics 2026-Q2

Continue reading

Technology

SlackSwap combines data transfers and recomputation in GPU-limited training

Researchers at UC Merced and the University of Texas at Austin introduced SlackSwap, a planner for deep-learning training. It recomputes some activations when the transfer budget between a GPU and host memory is exceeded; reported results come from simulation, not hardware measurements.

arXiv1 min read
Data included