00:00 (upbeat music)
00:01 Generative AI, as it implies in the name,
00:04 is actually generating new content from data.
00:06 So it's trained on very large data sets,
00:09 but the output that you get is brand new content.
00:12 You can even ask it to create content
00:14 like presentation materials or marketing slides.
00:18 So in that context, we'll see generative AI
00:21 in pretty much every industry.
00:23 (upbeat music)
00:28 Hi, my name is Monica Livingston
00:30 and I lead the AI Center of Excellence at Intel.
00:33 Generative AI is very, very new.
00:36 And a lot of the models out there,
00:38 we've not really gotten to the optimization phase.
00:42 And so Intel is focusing on that optimization
00:45 for those models as well to make AI accessible.
00:48 For training, we have our Habana GAU-82 processors
00:53 and we have our discrete GPUs, the GPU Max family.
00:57 And then for inference, for the smaller models,
01:00 we use Intel Xeon scalable processors,
01:03 and specifically the fourth generation
01:05 Intel Xeon scalable processors
01:07 because of that advanced matrix extension
01:10 or AMX engine inside that processor.
01:13 And then we're spending a lot of time
01:15 actually optimizing software for our hardware.
01:19 So for example, we have a tool out of Intel Labs,
01:22 it's called SetFit.
01:24 And what it actually does is enables smaller models
01:28 to run with the same type of accuracy,
01:30 but be trained on a much smaller data set.
01:33 Being able to train on less data,
01:36 again, makes these models much more accessible.
01:39 Most install bases, most data centers
01:40 have fourth generation Intel Xeon scalable processor
01:44 in their infrastructure.
01:45 For us optimizing these types of models on Xeon
01:49 means that they are accessible to those enterprises
01:51 because you already have this infrastructure.
01:53 You're not having to stand out new boxes.
01:56 The good thing about AI is that all open source
01:58 and it's online.
01:59 And so upskilling to AI is really not limited or selected.
02:04 It's very accessible.
02:06 Just think about all of the different
02:08 types of possibilities.
02:10 You can already see very context specific models out there
02:15 for law, for example, for medicine.
02:18 That's extremely helpful to those industries
02:20 because they're heavy paperwork focused.
02:22 And so when you have to do a lot of paperwork,
02:25 doing that with some of these large language models
02:28 saves significant amount of time.
02:30 As generative AI becomes more prevalent
02:32 and AI in general becomes more prevalent,
02:35 Intel continues to explore ways to optimize these models
02:40 and allow customers access to a wide range of hardware
02:44 and software tools in order to run these models
02:47 much more accessibly and much more cost effectively.
02:50 (upbeat music)
02:53 (electronic music)
02:56 (electronic music)
02:59 (electronic music)
03:02 [MUSIC PLAYING]
Comments