Meta puts GEM training efficiency at 20-25%, its paper says 37%
Meta credits the doubling of its ads foundation model’s training efficiency to “the latest GPU hardware” and never names the chip. The research paper that same post links does name it, and that paper reports a utilisation figure nearly twice as high.
The post went up on 13 October on Meta’s engineering blog, covering the 2026 generation of GEM, its Generative Ads Model. It says end to end training efficiency “is doubled to 20-25% model FLOPS utilization (MFU) with 4x total training FLOPs”, running on several thousand of what it calls the latest generation GPUs. MFU is the share of a chip’s theoretical peak compute a job actually lands, so a higher number means less idle silicon.
Kunlun, the architecture paper the post links for that work, puts the number at 37%. Its abstract credits its optimisations with raising “MFU from 17% to 37% on NVIDIA B200 GPUs”, and with doubling scaling efficiency on top of that. The same paper reports that recommendation models typically reach only 3% to 15% MFU, against 40% to 60% for large language models.
Those two numbers measure different things, but the gap is still the point. Kunlun’s appendix says its MFU “is measured on a single GPU and reflects only the model architecture’s intrinsic GPU efficiency, independent of system-level overheads”. Meta’s post supplies the bridge itself, defining end to end MFU as local MFU multiplied by a scaling ratio.
MFU is measured on a single GPU and reflects only the model architecture’s intrinsic GPU efficiency, independent of system-level overheads.
Kunlun paper, Appendix A.2, arXiv:2602.10016
Divide the blog’s 20 to 25% by the paper’s 37% and the implied scaling ratio is 0.54 to 0.68, our calculation from the two figures. That means a third to nearly half of single-GPU efficiency goes missing once the job spreads across thousands of chips. Meta doesn’t print that ratio, though it does say the split “lets us treat compute efficiency and scaling efficiency as related but distinct optimization problems”.
The kernel work behind the per-GPU number is published separately on the PyTorch blog, which says Meta’s GDPA kernel hits 1,145 BF16 Tensor Core TFLOPs and roughly 97% tensor core utilisation on B200s power capped at 750W. Across the full model those kernels deliver over 30% more training throughput. The Kunlun paper puts that one kernel’s share at 3 points of the 37%, since removing it “reduces MFU from 37.0% to 34.0%”. Custom GPU kernels are a contested lever right now, which Databricks leaned on with Proteus too.
The ad numbers cover different populations
On the advertising side, the post says GEM “has delivered 18.9% conversion improvement and 12% ad click-through rate (CTR) improvement” across Meta’s ad surfaces and objectives. But it then says GEM “additionally contributed” three smaller figures: 5.8%, 4.1% and 1.6% conversion uplift on Facebook, Instagram and off-site conversion. Those three are limited to “targeted users over the past year”, a qualifier the two headline numbers don’t carry.
Last year’s GEM post split the two apps out as well. Writing in November 2025, it reported a 5% conversion increase on Instagram and 3% on Facebook Feed in Q2, so Instagram led then. This year Facebook leads, at 5.8% against Instagram’s 4.1%, though the two posts aren’t measuring the same population. GEM is ads infrastructure, a different line from the enterprise AI revenue Meta has also been chasing.
What the linked papers quantify and the blog doesn’t
| Item | Meta’s GEM blog post | The paper it links |
|---|---|---|
| GPU named | “the latest generation of hardware” | NVIDIA B200 (Kunlun) |
| Model FLOPs utilisation | 20 to 25%, end to end | 17% to 37%, single GPU (Kunlun) |
| Knowledge transfer | “doubles the intelligence transfer” | 0.5%, then 1.03% and 1.22% conversion per launch (LoopFM) |
| Precomputation | “harvest the gains from the upstream model” | 0.67% gain on revenue-driving top-line metrics (SOLARIS) |
LoopFM, the post-training technique the blog credits with doubling intelligence transfer, carries per-launch results in its own paper: a 0.5% conversion improvement in what it labels Y1H1, then 1.03% and 1.22% from two individual launches in Y1H2. SOLARIS, the speculative precomputation system the blog calls Solaris, reports a 0.67% gain on “revenue-driving top-line metrics” while deployed across an ads system serving billions of daily requests. Neither number appears in the blog post, which is why the papers are worth opening.
One more claim rewards a close read. The post says GEM “scaled the long engagement context length by 2x in production”, then “further enabled the capability of scaling long context to 25x”. But only the 2x is serving ads today. The 25x is described as demonstrated infrastructure capability, so that’s the number to watch when the next generation ships.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
