
Wernstrom! A Paradigm Shift in AI... Maybe!
Good news everyone! It appears that Nvidia that silicon slinging behemoth has unveiled something called 'NVLink Fusion.' At first glance it seems they're letting others play in their NVLink sandbox. NVLink as you may or may not know (frankly I've forgotten myself) is a fancy way for Nvidia chips to chat with each other. Now they're saying it can talk to *other* chips even those made by competitors. A bold move I say! Or perhaps a foolish one like that time I tried to invent a hat that makes you smarter only to discover it attracted space wasps. Wernstrom!
Semi Custom... Sounds Like Something Zoidberg Would Order
Jensen Huang the CEO (a very smart cookie) proclaimed it will allow building 'semi custom AI infrastructure.' Semi custom? Sounds like something Zoidberg would order! But I digress. The idea is that you can mix and match Nvidia processors with CPUs and ASICs (Application Specific Integrated Circuits) from other manufacturers. Apparently MediaTek Marvell and a host of other companies are already on board. But what does this mean? Does it mean Nvidia is finally playing nice? Or is this a cunning plan to control the entire AI pie even the bits that aren't strictly Nvidia flavored? My gut tells me it's a bit of both. Though my gut also tells me to eat more Slurm.
Data Centers of the Future: Now With 20% More Confusion!
Ray Wang some semiconductor analyst (probably not as smart as me) suggests that NVLink is an attempt by Nvidia to grab a bigger slice of the data center pie especially those using ASICs. These have traditionally been seen as competitors and Nvidia wants to play them all to their advantage. Let us hope it works out because in the wrong hands technology could be used for evil! Like building bigger bombs or creating robot underlords!
A Calculated Risk or a Horrifying Mistake?
However Rolf Bulk another analyst person points out the obvious: Letting customers use alternative CPUs might decrease the demand for Nvidia's own CPUs. Doh! Seems like a bit of a double edged sword doesn't it? Still flexibility at the system level will improve Nvidia's GPU based solutions by comparison and in the end Nvidia will maintain their position at the center of the AI or so they say but in this world nothing is set in stone except for rocks.
Grace Blackwell and the GB300: The Next Generation of... Something!
Beyond the NVLink hullabaloo Nvidia also talked about its next generation Grace Blackwell systems specifically the 'GB300' coming out later this year. Apparently it will offer 'higher overall system performance.' (I'll believe it when I see it!) They also announced the DGX Cloud Lepton a platform to connect AI developers with GPUs. More GPUs are always good unless they're used to power doomsday devices. On a side note I think I can teach a supercomputer to play the theremin.
Taiwan Foxconn and an AI Supercomputer: A Recipe for... Innovation? Or Doom?
Lastly Nvidia plans to build a new office in Taiwan and is partnering with Foxconn to build an AI supercomputer. Partnering with Foxconn? That's bold! Hopefully they'll treat their employees better than they treat their robots. Seriously though it is good to see people are making new things even if this new thing is just a fancy computer. As I always say "When will they learn that science mustn't interfere with things man was meant to know?
Comments
- No comments yet. Become a member to post your comments.