ONNX Format Announcement
Microsoft developers unveil the ONNX (Open Neural Network Exchange) format, a groundbreaking open standard for AI model interoperability, to a room of tech journalists and AI researchers, signaling a
Setting
Microsoft's Redmond campus, a modern conference hall with floor-to-ceiling windows overlooking lush greenery, equipped with state-of-the-art AV systems and ergonomic seating
Characters
The figures in this scene as an entity network — co-presence links everyone in the moment; speakers who trade lines are bound tighter. Turn the resolution dial to reveal depth the engine actually computed.
TNGF
SELECTED
Microsoft Lead Developer
primary
A middle-aged man of average height with a slightly stocky build, short-cropped dark hair with hints of gray at the temples, and rectangular glasses that reflect the presentation screen's glow. His face shows the focused intensity of someone deeply immersed in technical work, with faint smile lines suggesting frequent engagement in problem-solving discussions.
Tech Journalist
secondary
A sharp-eyed woman in her early 30s with a lean build, wearing rectangular glasses that reflect the screen's glow. Her dark brown hair is pulled back in a professional ponytail, and she has a slight furrow between her brows from years of scrutinizing technical details.
AI Researcher
secondary
A middle-aged man of average height with a slightly hunched posture from long hours at the computer. His dark brown hair is thinning at the crown, and he wears rectangular wire-frame glasses that frequently slide down his nose. His face shows signs of fatigue around the eyes, but his gaze remains sharp and analytical.
AV Technician
background
A young to middle-aged man in his late 20s to early 30s, with a lean build and slightly tousled dark brown hair. He has a focused demeanor, with sharp eyes that quickly scan the equipment for any issues. His hands are nimble, accustomed to handling delicate AV components.
Dialog
Microsoft Lead Developer
What ONNX provides is a shared computational graph representation—think of it as a universal translator for neural networks across frameworks.
Tech Journalist
But how does that handle custom ops? If I've got PyTorch extensions running CUDA kernels, does that just break the portability promise?
AI Researcher
Assuming the operator semantics can be preserved through serialization—which isn't trivial with stateful layers—this could actually solve our reproducibility headaches. Right?
Microsoft Lead Developer
We've implemented extensibility hooks precisely for those cases—the ONNX runtime will pass through unrecognized ops to the native framework with full tensor metadata.
Tech Journalist
So now we're back to framework lock-in for anything nontrivial. That's not the 'write once, run anywhere' story you led with.
AI Researcher
If we consider this as version one, the coverage for standard architectures is actually quite... (checks notes) 87% of TorchVision models convert losslessly according to your whitepaper.
Microsoft Lead Developer
Exactly—and that percentage grows as more ops get standardized through community contribution. This is why we open-sourced the spec on day one.
Chat with Characters
Causal neighbors · 347 linked moments
A
2011
· same figure
G
2005
· same figure
R
1975
· same figure
T
2017
· same figure
R
2017
· same figure
M
2007
· same figure
2
2015
· same figure
M
2014
· same figure
G
2016
· same figure
G
2015
· same figure
R
2018
· same figure
I
1981
· same figure
T
1977
· same figure
F
2006
· same figure
C
1977
· same figure
A
1977
· same figure
A
2018
· same figure
A
2017
· same figure
A
2019
· same figure
G
2012
· same figure
S
2011
· same figure
A
2014
· same figure
W
2014
· same figure
I
2015
· same figure
G
2018
· same figure
G
2019
· same figure
2019
· same figure
2018
· same figure
A
2011
· same era
H
2008
· same era
I
2011
· same era
S
2007
· same era
G
2018
· same era
N
2017
· same era
T
2017
· same era
C
2017
· same era