From Newsgroup: comp.lang.prolog
Hi,
On small clusters, a failed GPU process
usually crashes visibly. On large TPU systems,
you can get rCLsilentrCY or rCLpartialrCY failures
where a chip is alive but unhealthy rCo
producing NaNs, stalling on the interconnect,
or returning corrupted data thatrCOs
not immediately flagged.
At hyperscale, training becomes a kind of
rCLdistributed computer architecture problem,rCY
and engineers do end up inventing their own
rCLparity-checkrCYrColike systems rCo not for
individual bits (as in DRAM ECC), but for
entire tensors, gradients, and replicas.
See also:
Gemini 2.5: Pushing the Frontier
https://arxiv.org/abs/2507.06261
Bye
Mild Shock schrieb:
Hi,
Prolog has become the Useful Fool.
Over the years it was trained that
constraints are importanter than
anything else. Even this talk blends
out natural language processing:
Manuel Hermenehgildo - 50 Years of Prolog and Beyond https://prologyear.logicprogramming.org/videos/PrologDay_Session_1_talk.mp4
And if you give them some IBM Roadrunner
precursor, all they can do, is play
some stupid constraint games:
Parallel local search for solving Constraint Problems on the Cell
Broadband Engine (Preliminary Results) https://www.researchgate.net/publication/220481722
And mentioning in passing how useless
blunt constraint solving is as implemented
by most CLP(X), can't even solve
magic square of 10 x 10.
Bye
Mild Shock schrieb:
Hi,
With their stupid fixation on Emacs Prolog
missed a Billion Dollar Business Model.
Now people will start paying 4'000 USD per
Year and GPU to Red Hat to have some AI
model deployment and training enviroment:
DGX Spark Arrives at SpaceX
https://www.youtube.com/watch?v=peaIkB0NzS0
But these environments are nothing then
Python virtual environments. Can Prolog not
also push some evaluation graphs to
a GPU or NPU? Well it seems with its current
philosophy only if there is a spare Key
Combination in Emacs for it.
Bye
--- Synchronet 3.21a-Linux NewsLink 1.2