From strict uptime promises reach flexible computing: Jensen Huang case for thinking about di AI power problem again
Introduction: Every computing task must demand perfect power availability?
To build more artificial intelligence [AI] infrastructure dey raise one clear question: where di electricity go come from? New power generation and stronger electricity networks matter. But another question dey wey no dey get plenty attention. Computing fit adapt more intelligently to di power wey already dey available?
Jensen Huang argument start with di difference between peak demand and normal demand. Electricity grid [Power Grid] must prepare for difficult conditions. Di maximum power wey e need no describe every hour of operation.
Di opportunity na to design some computing workloads around that difference. Na proposal about flexibility, no be claim say energy limits don disappear.
1. Peak requirement no be constant requirement
Hospitals, transport systems, homes and other essential services need reliable electricity. Grid must make allowance for periods when demand high pass normal.
Data centre [Data Centre] wey insist on di same uninterrupted supply under every condition dey add another demanding promise to that system. If plenty customers make di same request, providers must plan around all those obligations together.
Huang question na whether every computing task need that level of assurance. Some work urgent. Other work fit wait, run more slowly or move elsewhere. To treat all these tasks as di same fit make infrastructure harder to provide than e need be.
Unused capacity for one time no suppose mean guaranteed spare capacity everywhere. Di engineering opportunity depend on di location, di workload and agreement with di electricity provider.
2. Graceful degradation dey change di service promise
Graceful degradation [Graceful Degradation] mean say system fit reduce im service level in controlled way instead of failing suddenly.
Computing system fit respond to temporary power limit by reducing im processing rate. Some answers go take longer. Less urgent tasks fit wait. Where suitable infrastructure dey, dem fit move critical work to another location.
Di aim na to preserve di important parts of di service while reducing consumption. To protect data na different requirement from to maintain maximum speed every moment. Slower result no need mean result wey don lost.
This one need deliberate design. Workload no fit become flexible just because contract describe am like that.
3. Di customer na part of di power architecture
One important part of Huang reasoning concern di contract chain. Customer dey ask cloud service provider [Cloud Service Provider] for strong availability commitment. Di provider then need matching assurances from im own infrastructure and electricity suppliers.
Each negotiation fit look reasonable when you see am alone. Together, those promises fit create system wey leave only small room to adapt.
So business leaders suppose understand which services really need di strongest commitment. Critical operational system and analysis job wey fit wait no suppose automatically get di same requirements.
Na management decision as well as engineering decision. Somebody must take responsibility for wetin slower service go cause, define acceptable delays and make sure dem no promise customers wetin di system no fit deliver.
4. Di data centre need response wey e fit use when power signal come
Flexible demand need more than monitoring electricity consumption. Di data centre must know wetin to do when less power dey available.
That mean to classify workloads, understand their dependencies and decide which tasks fit pause safely. E also mean to know wetin fit move and wetin must remain where e dey.
To move computing no be di same as to move line for diagram. Data access, network capacity, security and service commitments all affect whether relocation make sense. Di architecture suppose make those limits clear before difficult period arrive.
Di test na practical one: di facility fit lower demand while keeping im essential promises?
5. Electricity providers need another kind of offer
Di third participant na di electricity service company [Utility]. Huang proposal dey ask providers to consider supply arrangements with different levels of assurance, instead of offering only di strongest uninterrupted commitment.
Customer wey ready to accept occasional limits fit use opportunity wey no suitable for load wey get strict specifications. Whether e go work for particular market need actual agreement, suitable equipment and clear operating rules.
Efficiency still matter too. More useful computing per unit of energy dey help, but e no guarantee say total consumption go fall when demand grow. Better efficiency and better demand flexibility dey address different parts of di problem.
Conclusion: Dem must design flexibility across di whole chain
Di AI power problem no fit get solution only inside processor or only for power station. Customer expectations, contracts, software, data-centre operations and electricity supply dey affect one another.
Huang useful challenge na to ask whether some of today strict promises necessary. Where flexibility acceptable, e suppose become clear service with clear limits, instead of response wey dem rush arrange because shortage happen.
Di goal na dependable computing wey understand when speed essential and when adaptation possible. That na better engineering question than to demand maximum performance from every workload every moment.
Powered By K7