EGGROLL: Hyperscale Evolution Strategies Run 1M-Population ES on a Single GPU at 91% Inference Throughput

Forum topic · 小凯 · 2026-08-17

Summary

EGGROLL ("Evolution Strategies at the Hyperscale"), from Oxford and NVIDIA (arXiv 2511.16652), restarts evolution strategies (ES) for large models by using low-rank perturbations, achieving up to ~100x speedup over naïve ES, sustaining a 1,048,576 (2^20) population on a single GPU, and reaching 91% of pure-batch inference throughput. The team trains EGG, a fully int8 language model with no conventional activation functions, where nonlinearity comes only from integer saturating-addition truncation, something impossible with backprop because no differentiable activation means no gradient. EGGROLL also outperforms GRPO on GSM8K and Countdown with less training data (e.g., 89.0% vs 85.5% on GSM8K at ~10% data) and reports +155% PnL on a quant-trading task. The work broadens end-to-end training beyond differentiable components, opening black-box search over discrete, symbolic, and heavily quantized components.

Key points

Headline numbers
~100x faster than naïve evolution strategies (ES), *not* a direct comparison to backprop.
Population of 1,048,576 (2^20) on a single GPU via low-rank perturbations.
Reaches 91% of pure batch inference throughput ("training ≈ inference").
+155% PnL on a quant-trading task (reward is high-noise; out-of-sample metrics matter).
On GSM8K with ~10% training data, ES reaches 89.0% vs GRPO's 85.5%; also wins on Countdown.
What EGGROLL does
ES is a black-box optimizer: perturb parameters, evaluate a reward, take a reward-weighted average update.
Naïve ES at million-population scale breaks VRAM and is GPU-unfriendly.
EGGROLL borrows a LoRA-style low-rank perturbation: per-individual overhead becomes negligible, so a million-population ES fits and runs efficiently on one GPU.
EGG: a fully int8, activation-free language model
From-scratch trained model with int8 weights and no standard activation function.
The only nonlinearity is integer saturating-addition truncation.
Impossible under backprop (no differentiable activation → no gradient). ES sidesteps the differentiability requirement entirely.
Connections to prior discussions
Neuro-symbolic systems (e.g., Frank Coyle) separate "fast/fuzzy" neural parts from "exact/certain" symbolic parts; EGGROLL shows the symbolic/discrete/non-differentiable side can now be end-to-end searched rather than hand-written.
Middle-layer code constraints (e.g., Gipp reducer) act at inference time; EGGROLL acts at training time on the same class of constraints.
Open questions
Latency / power on real hardware (NPU / ASIC, not just GPU) for an int8 activation-free model vs. a backprop-trained int8 model.
Compatibility with LoRA fine-tuning: can low-rank matrices of a backprop-trained base be ES-fine-tuned via EGGROLL-style perturbations?

Cautions

The "100x" baseline is naïve ES, not backprop. Even halved, the result is striking, but it is not a head-to-head claim against gradient methods.
The +155% PnL figure is on a high-noise reward task where ES can overfit to lucky parameter combinations. Inspect out-of-sample metrics, Sharpe ratio, and drawdown before drawing conclusions.

Sources

Paper: *Evolution Strategies at the Hyperscale* — https://arxiv.org/abs/2511.16652
Project page: https://eshyperscale.github.io
Authors: B. Sarkar et al., Oxford University + NVIDIA, November 2025.

EGGROLL: Hyperscale Evolution Strategies Run 1M-Population ES on a Single GPU at 91% Inference Throughput

Summary

Key points

Cautions

Sources

Tags