PyTorch 2.15 tackles the 'one bad GPU' problem
AI summary
PyTorch 2.15 addresses the "one bad GPU" problem, where a single failing accelerator can disrupt large-scale AI training runs. The release involved clearing away accumulated technical debt so that training at scale becomes more resilient. For teams running distributed workloads, that points to fewer costly interruptions and less wasted compute when hardware misbehaves.
Headline and excerpt are shown for attribution. Full text remains with the publisher.
Request a consultation
Talk to a vetted AI-governance partner
News, not engineering or legal advice. Verify vendor guidance and the rules that apply to you.
Ask one question
Ask a specialist one question
No call, no form interview. Send the question you would ask a colleague, and a vetted specialist in this field answers in writing.
Related coverage
Sponsored: Water stewardship starts with design, not offsets
A sponsored article argues that water stewardship for data center development should be built into project design rather than handled later through offsets. It asks how new facilities can both limit their impact on local water resources and support the long-term resilience of the watersheds they rely on. The point matters to professionals because data center growth can intensify local water stress, making early planning more effective than compensating after construction.
Fireworks says you’re wasting money on AI unless you’re using all the models
Fireworks is arguing that companies overspend on AI by defaulting to a single model for every task, and it is making the case for managed inference that automatically routes each query to the model best suited to it. The pitch is that matching workloads to the right model can cut costs while improving results, instead of paying premium prices for capability a given task doesn't need. For teams deploying AI at scale, the takeaway is that model selection and routing are becoming as important as the models themselves.
DCD Talk: Rethinking data center battery validation with Tong Xu, CSB
Tong Xu of CSB spoke with DCD's Emma Strutton about how data center batteries are currently validated and why that approach falls short. The discussion makes the case for moving past marketing labels and conventional runtime figures toward validation methods that are replicable and capture dynamic performance under real-world conditions. For professionals specifying backup power, the implication is that familiar benchmarks may not reliably predict how a battery will behave in actual service.
The process
How this works
- 1
Send the form
Name, email, and what you need. One request, one channel.
- 2
We route it
A vetted independent partner for this channel receives it. No one else does.
- 3
You get a reply
Within one business day, at no cost and with no obligation.
One partner per channel. Your details are never sold, and no partner sees another partner's requests.
The weekly brief
Get the weekly brief
One email a week with the stories that moved this field, plus what to watch next.
- One email a week, with the Monday refresh.
- Stories first — at most one labelled note, usually our own.
Fields marked * are required.