6) Liquid cooling takes off
The warmth generated by AI techniques has already pushed the necessity for liquid cooling as a result of air cooling is solely not sufficient. Nevertheless it’s not simply GPUs that run sizzling – new CPUs from Intel and AMD are fairly toasty as properly. I anticipate to see a rise in self-contained liquid cooling that can go into present information facilities and never require a hefty retrofit. I additionally assume that HPE and Dell will lastly do their very own liquid cooling, much like Lenovo’s Undertaking Neptune. Up till now, HPE and Dell have been content material to depart liquid cooling to 3rd events, however they could must lastly do their very own factor.
7) Intel splits
There’s simply no means round it. Intel should spin off its fabrication enterprise similar to AMD did in 2008. It was costly, painful and obligatory for long-term success. Gelsinger merely didn’t have the bandwidth to handle Intel foundries and Intel merchandise, and all three suffered for it: The fab enterprise was sluggish to get began, the chip enterprise fell behind AMD, and Gelsinger’s tenure was minimize quick. It’s time to chop the fabs unfastened, Intel.
8) Continued predictions of on-prem exodus
This utterly inaccurate prognostication will proceed, and it’ll proceed to be fallacious. There are too many causes to maintain an on-premises information heart, beginning with information privateness and integrity. Repatriation of knowledge from the cloud to on-premises infrastructure goes on yearly. On-premises information facilities will die when mainframes do.
9) GPU utilization turns into paramount
Nvidia is displaying no indicators of reducing energy draw, so it’s as much as others to make this stuff work as effectively as potential. Which means most utilization and scaling of {hardware}. Consequently, maximizing GPU utilization will develop into the first design objective for contemporary information facilities. This may drive improvements in {hardware} and software program to maintain the infrastructure obligatory for coaching and decrease latency, lagging, and different points that trigger pauses in coaching. Success shall be measured by how successfully information facilities can maintain their GPU sources utilized.
10) Energy constraints impression information heart areas
With virtually 500 information facilities within the Virginia space, it’s secure to say that area is reaching its restrict. The identical applies for Texas and Santa Clara. The demand for large-scale processing of knowledge for AI, information analytics, and quantum computing will shift the place new information facilities are constructed. The excellent news is that these purposes, particularly AI coaching, usually are not latency delicate, to allow them to afford to be positioned in a distant location that has quite a lot of land and quite a lot of low cost energy. This may largely contain information facilities devoted to large computational processes, so there’s no want to fret that your colocation supplier will open up store within the Sierra Mountains.