The Training Engine of the AI Era.
2.0×
Token Efficiency
AERO requires at least 2× fewer training steps to reach the same model quality, on any hardware — directly measured, not a projection.
3.7×
Wall-Clock Speedup
Combines 2× fewer steps with faster multi-node communication. In our testing, AERO achieved the same loss target in 27% the wall-clock time. (Tested on 4×8×A100 GPU machines.)
AERO replaces the entire AI training loop. Not optimized. Not patched. Replaced.
Training is the single biggest bottleneck in AI today.
Coordination overhead, memory traffic, and synchronization delays waste capacity at every scale.
Clusters often run well below full utilization.
Frontier training runs are heading toward nine-figure budgets.
SourceThe current pipeline cannot scale efficiently
The market has been polishing parts. AERO rebuilds the engine.
- Better kernels
- Better schedulers
- Better tuning
- Better optimizers
Rethinking the engine itself.
Shifting focus from individual components to the redesign of the entire architectural execution model.
A unified execution model built for efficiency.
Super-efficient communication · Optimized training path · Fused execution · Same math
Reduce Overhead
Sharply cut communication and synchronization drag.
Optimize Memory
Minimize unnecessary trips to keep data moving.
Maximize GPU Focus
Keep compute units focused on useful computation.
Faster Outcomes
Reach model targets faster with reduced friction.
AERO transforms the performance, cost, and deployment of large-scale model training
Hardware agnostic.
Quickly deployed.
Many deployment configurations can be completed in under 24 hours, so you can begin evaluating performance gains quickly.
Hardware Agnostic
Compatible with any existing infrastructure — no specialized hardware required.
Rapid Adoption
Integration-oriented design built for onboarding without lengthy platform transitions.
As training grows costlier,
efficiency becomes strategy.
AERO is the AI Training Strategic Layer.
- 1Frontier training economics
- 2Stranded / idle infrastructure use
- 3Extended GPU lifespans
- 4Capital-efficient scaling
A resilient IP strategy securing long-term platform value.
Core training-loop innovations are protected by a pending patent and by trade secrets.
Strategy 1: Trade Secrets
Proprietary internals never disclosed externally.
Strategy 2: Follow-on Patent Families
Layered protection extending the original filing.
1.36B parameter model,
validated results.
Measurable edge across every training metric.
AERO requires at least 2× fewer training steps to reach the same model quality, on any hardware — directly measured, not a projection.
Combines 2× fewer steps with faster multi-node communication. In our testing, AERO achieved the same loss target in 27% the wall-clock time.
WHITE PAPER · 1.36B SWIGLU C4 PRETRAINING · JUNE 2026
AERO Training Loop — 2× Tokens, 3.7× Wall-Clock
Apples-to-apples vs well-tuned AdamW at 1×–8× Chinchilla, benchmarked against Stanford's published optimizer landscape
01
BACKGROUND
- –AERO is a drop-in training loop from DevOpt Labs — full forward/backward/step replacement, not just an optimizer swap.
- –Same code path on H100/A100, single- and multi-node fabrics.
- –Methodology follows Stanford's Fantastic Pretraining Optimizers (Wen et al., 2025).
02
METHODS
- –1.36B SwiGLU transformer; packed C4-EN, GPT-2 BPE; seq 2,048; bf16 mixed precision.
- –Cosine LR (peak 1.9e–3, 8k warmup); 1× C = 27B tokens, ~415k steps.
- –Test 1: 1×8×H100 (token efficiency). Test 2: 4×8×A100 (wall-clock, comms-bound).
- –AdamW tuned to Stanford-consistent slope α=0.32 — well-tuned, not under-tuned.
03
KEY FINDINGS
- –AERO@1×C = 2.55 = AdamW@2×C — 2.0× tokens-to-target, directly measured.
- –Cluster wall-clock: 47 h vs 175 h — 3.7× faster, 73% GPU-hours saved.
- –Per-step time on 32×A100: 408 ms vs 760 ms (Lever 2 throughput advantage).
- –130M proxy reproduces gap to two decimals across seeds.
04
IMPLICATIONS
- –~10× the best published optimizer gain (Muon/Soap/NAdamW ~1.0–1.1×).
- –Levers compound — wall-clock gain grows on slower fabrics.
- –Stable single-config across 415k steps; no spikes or restarts.
- –C4-EN at 1×C favors AdamW — 2× headline is a lower bound.
HEADLINEAERO halves training tokens to a given loss and cuts cluster wall-clock by ~3.7× — measured, seed-invariant, and a lower bound on production-typical gains.
Source: DevOpt Labs — AERO Performance White Paper, 1.36B SwiGLU C4 Pretraining, June 2026
RESEARCH BRIEF · PRETRAINING OPTIMIZERS · 2026-06-05
AERO at 1×C beats AdamW at 2×C
130M-parameter LLaMA-style pretraining on SlimPajama-6B and OpenWebMath, one RTX 5090 per run
01
BACKGROUND
- –Extends the 1×-Chinchilla study by doubling AdamW's compute (278k steps, ~4.55B tokens).
- –Decisive question: can AERO at 1×C still beat AdamW given 2×C?
- –All else held fixed — model, data, tokenizer, batch, schedule, seed.
02
METHODS
- –130M LLaMA-style decoder; SwiGLU, RoPE, RMSNorm; GPT-2 BPE, seq 2,048.
- –Batch 16,384 tok/step; cosine LR → 0 with 1k warmup.
- –Budgets: 1×C = 139k steps; 2×C = 278k steps.
03
KEY FINDINGS
- –SlimPajama: AERO 1×C 3.259 vs AdamW 2×C 3.385 — AERO wins by 0.126 nat at half the compute.
- –OpenWebMath: AERO 1×C 2.205 vs AdamW 2×C 2.264 — AERO wins by 0.059 nat at half the compute.
- –At equal 2×C compute, gap widens to 0.242 (SPJ) and 0.166 nat (OWM).
04
IMPLICATIONS
- –AERO delivers better loss for half the compute against a fully LR-tuned AdamW baseline.
- –Lead grows with data entropy — larger margin on the noisier corpus.
- –Unused stability headroom suggests reported margins are a lower bound.
HEADLINEAERO at 1×C beats AdamW at 2×C on both corpora — better loss for half the compute, lead widens at matched 2×C budget.
Source: AERO at 1×C vs AdamW at 2×C — SlimPajama-6B & OpenWebMath pretraining study, June 5 2026
DIAGNOSTIC STUDY · GRADIENT-CLIP INTERNALS · 2026-06-04
Untapped AERO Potential — Gradient-Norm Headroom
AERO sits 4–5× below AdamW's gradient norm and almost never clips — a saturated AdamW vs an unsaturated AERO
01
BACKGROUND
- –Diagnostic from SlimPajama, OpenWebMath, and C4 pretraining runs.
- –Same model, data, schedule, seed; only the optimizer differs.
- –AdamW clips at norm 1.0; AERO at 4.0 — pre-clip grad_norm_raw compared.
02
METHODS
- –TensorBoard scalars: grad_norm_raw, clip_coef, clipped.
- –Rolling-mean smoothing (window 800 at ebsz 8, 2000 at ebsz 32).
- –Runs span ebsz 8 (1-GPU) and ebsz 32 (8-GPU) across all three corpora.
- –% clipped = rolling fraction of steps with clip_coef < 1.
03
KEY FINDINGS
- –AdamW clips ~100% of steps by mid-training (86% even at ebsz 32).
- –AERO clips ≈0% — raw norm stays ~4–5× below AdamW at matched batch.
- –Both follow the 1/√B noise-floor law; AERO sits in a much lower band.
04
IMPLICATIONS
- –AdamW is saturated; AERO retains large unused stability margin.
- –Low end-of-run norm is a flat-minimum signature — matches AERO's lower val loss.
- –Reported AERO margins are likely a lower bound.
HEADLINEAdamW runs pinned at its clip bound; AERO never clips and sits 4–5× lower — large stability headroom left to spend for more performance.
Source: Untapped AERO potential — gradient-norm headroom diagnostic, June 4 2026
EARLY RESULTS · DIRECTLY MEASURED VISION BENCHMARKS · FIVE TASKS
AERO Across Five Vision Tasks
The training-loop advantage carries across five public vision benchmarks — medical and natural images alike — with only the optimizer changed in every comparison
01
BACKGROUND
- –First proven on large-language-model pretraining; now measured directly on vision.
- –Five public benchmarks: medical segmentation, medical classification, and CIFAR.
- –Proxy-scale backbones — SAM ViT-B and ViT-Ti.
02
METHODS
- –Only the optimizer changes — data, model, seed, budget, schedule all identical.
- –Both sides tuned: AERO vs AdamW's own swept best.
- –Four seeds most tasks (three on NCT-CRC); public splits.
03
KEY FINDINGS
- –Wins 5 / 5 vision tasks vs tuned AdamW.
- –+1.0 to +3.3 pp higher accuracy / Dice at peak.
- –2–4× fewer steps to reach AdamW's peak.
- –CIFAR wins unanimous across all four seeds.
04
IMPLICATIONS
- –Signals a general training-loop effect, not a dataset artifact.
- –Margin grows on the harder task (CIFAR-10 → 100: +3.12 → +3.32 pp).
- –Still early — public, proxy-scale; clinical and RL not yet measured.
HEADLINEAERO wins all five vision benchmarks vs a tuned AdamW — +1.0 to +3.3 pp higher accuracy and 2–4× less compute to reach AdamW's peak, in every head-to-head so far.
Source: DevOpt Labs — AERO × Vision early results across VerSe'20, MURA, NCT-CRC-HE, and CIFAR-10/100. Datasets and backbones openly licensed; AERO configuration proprietary.
Built by People Who've Done This Before
Ralph has spent his career pushing the limits of what's computationally possible. He began in high-performance computing, pioneering 3D seismic algorithms and building the first commercial 3D seismic processing system — reducing months-long compute runs to hours. Mentored by signal-processing luminary Sven Treitel, Ralph invented ThinAnywhere in 2000, the first remote 3D visualization platform. Today at DevOpt Labs, he's applying that same frontier mindset to AI: replacing broken training architecture at the systems level.
Charles Black
Chief Developer
Charles Black is a veteran systems engineer known for precision, scalability, and innovation. He began as a programmer at NOAA's Princeton labs, then joined Ralph Gobeli to pioneer seismic data processing, designing the GUI for a 3D seismic platform with over five million lines of code. He led the development of ThinAnywhere, delivering near real-time, 100 fps 3D visualization across dual HD monitors through distributed computing and optimized graphics pipelines. Today he focuses on software integration, performance engineering, and system optimization.
Louis is a veteran energy and sustainability executive with 45+ years of global experience spanning 35+ markets. He has held senior leadership roles at Schlumberger (SLB), OVS Group (now Baker Hughes), Intel, and Earthen. A technologist at heart, Louis combines deep technical insight with extensive management experience to drive innovation, transformation, and growth.
A Harvard mathematics graduate and serial entrepreneur with deep roots in computer, networking, and agriculture tech. He helped establish undergraduate computing with the first UNIX system outside Bell Labs, wrote MH at Rand, and co-founded 3Com where he led an early TCP/IP stack. He contributed to workstations at SGI, oversaw Doré — a 3D graphics package later sold to Apple — and invented a desktop network a decade before USB. He holds several patents and has lectured at UCLA and UC Santa Barbara.
Get in Touch
Interested in AERO? We'd love to hear from you.