ICS Video Player

Faculty Mini-Talks - Johnsson Lennart's Computer Science Faculty Mini Talks 2022

Transcript ×

Auto highlight

Off

Font-size

00:03	thank you for watching this video about was hearing and my research lab,

00:11	advanced computing research lab, which the is on energy efficient computing, that

00:18	highly important in their wide range of . And I will try to convince

00:24	that that is the case. So you are later on looking for jobs

00:29	the silicon industry, like Intel AMG IBM or in NVIDIA or others or

00:36	the user industry and particularly the big companies or the oil and gas

00:42	Energy is one of the prime concerns . Again, whether you do use

00:49	or platforms or your design and build . So in order to address this

00:58	looks at all aspects of the so to speak, algorithm software and

01:04	and students in the group worked on aspect or the other depending upon their

01:11	. So first time I tried to you that why power and energy is

01:18	important and here's the first argument why important. So the heat density of

01:30	At about 2005 got to the point there was no cooling technology available or

01:37	the horizon that would allow the continued increase that has been going on until

01:47	2005. As you can see on little right hand diagram, the heat

01:52	is close to the same as in nuclear reactor and about 10 times more

01:57	on your stove plate where, you , boil your eggs or make tea

02:02	coffee or something similar. So the density ended up being unacceptable to continue

02:12	the other part why it's important is Economics. So since about the same

02:19	as well it turns out that for lifetime of the system the cost of

02:26	and cooling exceeds the cost of the hardware itself. For the last bigger

02:32	was involved in purchasing myself. It's 1300 old cluster. It ended up

02:39	about 1/3 of the total cost was actual computing system itself. And the

02:45	reason why it's an important aspect is electricity consumption for what's known as the

02:53	. C. T. They basically and computing technologies uh By the end

03:01	this decade is expected to be more for about 20% of the total electricity

03:08	. And as you can see it's exponentially growing. So that in itself

03:13	a concern but not necessarily negative one computing tend to replace in many aspects

03:21	more energy consuming activities. But the is that only about a quarter of

03:28	generation comes from clean and renewable energy and about 2/3 comes from combustion.

03:37	it has a big environmental impact. this slightly basically gives a little bit

03:43	the scales at this because as data consume about the equivalent of 100 million

03:48	. S. Homes in terms of electricity. So now what can be

03:53	about it? Well software depending upon you um use it can have a

04:03	range of efficiencies as this points points for a very trivial case, basically

04:08	multiplication and you know, popular programming like python and java tends to be

04:16	inefficient in terms of resource utilization. at this slide shows it's about more

04:23	four orders of magnitude in terms of , depending upon the efficiency of your

04:29	. For the very trivial type algorithm matrix multiplication of course this is performance

04:35	not energy but unfortunately energy is not to the performance. It's Has a

04:45	but it's not very large. So would claim that may not be 60,000

04:50	between the best and worst case but probably most likely at least 10,000 times

04:57	. It's also that depending upon how architecture you can get a widely different

05:04	efficiencies and what this line shows that to standard microprocessors, kind of general

05:12	like your typical x 86 whether it from in calorie MB or some other

05:19	and um be about up to five of magnitude less energy efficient than a

05:28	type architecture. So that's why, was pointed out by the touring award

05:35	a couple of years ago, john and the Paterson that many of you

05:41	know by name from their very popular book. But it's basically says the

05:49	purpose type architecture that implied convergence over couple of decades is kind of over

05:56	to the need for more energy efficient And another one is this example from

06:04	saw some of the big users of , they took their medicine their own

06:10	and um, Google designed what they the tensor processing unit. And according

06:17	their own claims, if they hadn't it and continued with standard General Gpus

06:22	Cpus, they would have actually required to double their data centers. So

06:28	we're doing in the group now is to be driven by application by vertical

06:35	co design if you like that, a popular word or domain specific

06:40	but one of my students is focused machine learning for in this case for

06:48	applications and machine learning is incredibly our high energy consumption. So the

06:58	graph towards the end of the left side shows you that some of these

07:03	networks and their training uh requires as energy as a lifetime energy or I

07:11	say energy but the emissions associated given a mix of sources for electricity,

07:21	about five times as um damaging in of carbon dioxide as five cars over

07:31	lifetime of the cars. Another student working on encoding up point clouds and

07:38	to you use it in fact for in his case to do computational

07:45	And so what we've come up with representation of point class that is Up

07:51	200, we have served up to times as compact if you like compared

07:57	a standard of representation. And a student has been working on understanding and

08:04	a novel architecture. I was originally for vision processing, but then it's

08:10	used for machine learning and we have gun up obtain a demonstrated quite a

08:18	improvement in energy efficiency with proper algorithmic choices. And here is finally a

08:27	where the students have ended up in past summer, ended up in being

08:33	, professors at some good universities, ends up in the computer industry and

08:40	and uses of the industry and others more entrepreneurial and ends up in startups

5999:59

Previous Next

00 : 01