Searching...
18 results for “compute”
seven powers, and I don't know if he would consider this one. But there is, like, an ecosystem aspect here because the EDM companies Yeah. And the IP companies are so deeply integrated with TSMC. If you wanna be using Arm, for instance, or, you know
1:59:55 / 2:01:06
from transcript1:59:55 – 2:01:06
characteristics that being able to plow your CapEx into manufacturing capability does. Yeah. Should we do power now? Absolutely. Let's do it. Let's do it. So for folks new to the show, this is one of the discussion topics we do for every episode as w
1:56:16 / 1:57:19
from transcript1:56:16 – 1:57:19
none of TSMC's customers really benefit from other customers being on it. No. I do think there actually is I don't think Hamilton captures this in his seven powers, and I don't know if he would consider this one. But there is, like, a ecosystem aspec
1:56:16 / 1:57:21
from transcript1:56:16 – 1:57:21
...where the compute required in the future is unable to be provided by anything that TSMC is good at today. If all the crazy laser molten tin ASML stuff we were talking about,
1:56:50 / 1:58:08
from transcript1:56:50 – 1:58:08
there's not a lot of power, you know, as defined by sustainable, you know, economic profit, you know, operating cash flow coming out of this thing. So then as we talk about power here, what power do they have? And for listeners who are newer, this is
1:34:40 / 1:35:44
from transcript1:34:40 – 1:35:44
computer out there that is very performant for AI and runs on 20 watts, and we're using it to converse right now. It's our brains. Exactly. Right? So so that's very encouraging. And again, the brain to us instead of how you know, we we build it with,
42:42 / 43:46
from transcript42:42 – 43:46
You know you know, maybe you could look up what the power of consumption of a human a person is because the numbers are gonna get silly. But, like, that 15 to 20 megawatts was standard data center size. It was just unprecedented that was all GPUs run
3:50:12 / 3:51:16
from transcript3:50:12 – 3:51:16
...and compute perfectly, there may be a time period where your GPUs are just idle and you're exchanging weights and you're like, hey. The model's updating. So you're exchanging the radiance. You do the model update, and then you you start training agai
4:00:42 / 4:01:48
from transcript4:00:42 – 4:01:48
and a half thousand cores that are capable of running CUDA software. It's got 640 tensor cores, which are highly specialized for matrix multiplication. They have 80 streaming multiprocessors. So what are we up to here? Close to 20,000 unique cores on
1:16:31 / 1:17:34
from transcript1:16:31 – 1:17:34
...compute, which is the human brain. So, our brains use about 20 watts, of power, but and all that, only about 10 watts is higher brain function. Most of it's you know, half of it is just housekeeping functions, you know, keeping your heart going and b
56:59 / 58:08
from transcript56:59 – 58:08
...a computer it's when your monitor is sleeping. That's the amount of energy that your brain is consuming as it does all these crazy calculations. It's one blade of one GPU fan in one of these data centers. That's what I think of it. I feel like Noam S
19:31 / 20:39
from transcript19:31 – 20:39
and their strategy is what I think it is of not retailing TPUs at any point, then your customer is only yourself, so you're constrained by the amount of people you can get to use Google Cloud. Well, and at least with Google, they have Google Cloud th
1:48:43 / 1:49:45
from transcript1:48:43 – 1:49:45
if you were to scale this this chip, this tiny, tiny chip right here to a whole wafer, it would run on 20 watts and have, about 1,500,000,000 p bits and, 17 or 20,000,000,000 parameters per layer, and then you could do multilayered programs. So you g
43:29 / 44:36
from transcript43:29 – 44:36
state will you need to bring in from memory either like on chip SRAM or HBM from the accelerator attached memory or DRAM or over the network. And then how expensive is that data motion relative to, the cost of say an actual multiply in the matrix mul
32:16 / 33:20
from transcript32:16 – 33:20
...computer, if I own the storage, then I own the data. So the chip wars have accelerated primarily because what's sitting on top of it or what's being manipulated by it is becoming more and more,
10:30 / 11:40
from transcript10:30 – 11:40
...not just compute limited, but electricity limited. So both vectors really, really matter. Yeah. So I'm glad I'm glad you brought up the the ASIC versus,
34:53 / 36:15
from transcript34:53 – 36:15
Abilene, you know, two gigawatts. Right? You know, you have all these different sites Yeah. That they're they're signing up and discussions with, and we're we're noting them. And then we have the timeline because we're tracking entire supply chain. W
1:10:22 / 1:11:32
from transcript1:10:22 – 1:11:32
having people with interesting ML research ideas of things we think will start to work in that timeframe, or will be more important in that timeframe, really enables us to then get, you know, interesting hardware features put into, you know, TPU n pl
36:49 / 37:52
from transcript36:49 – 37:52
Have a podcast?
Run your podcast through Clypt
Get ranked clips, hooks, and ready-to-post copy from your own episodes. Free to try.