Artificial intelligence chip startup Cerebras Systems Inc
2
00:00:03,200 --> 00:00:05,540
is making bold claims about its
3
00:00:05,540 --> 00:00:09,140
ability to rival NVIDIA Corporation in powering AI workloads
4
00:00:09,500 --> 00:00:10,740
But can they deliver
5
00:00:11,020 --> 00:00:11,760
Cerebras
6
00:00:11,760 --> 00:00:16,460
announced groundbreaking achievements in molecular dynamics simulations and sparse training
7
00:00:16,720 --> 00:00:22,280
potentially revolutionizing materials research and making powerful large language models more
8
00:00:22,280 --> 00:00:22,720
accessible
9
00:00:23,020 --> 00:00:25,940
But how exactly did they achieve these breakthroughs
10
00:00:25,940 --> 00:00:27,440
The company partnered with
11
00:00:27,440 --> 00:00:34,260
European AI powerhouse Aleph Alpha GmbH to create secure sovereign AI solutions for government
12
00:00:34,260 --> 00:00:34,740
agencies
13
00:00:35,000 --> 00:00:37,820
with the German armed forces as their first client
14
00:00:38,120 --> 00:00:40,120
This partnership is a significant
15
00:00:40,120 --> 00:00:41,580
endorsement for Cerebras
16
00:00:41,820 --> 00:00:45,880
Why did Aleph Alpha choose Cerebras over established players like
17
00:00:45,880 --> 00:00:46,400
NVIDIA
18
00:00:46,600 --> 00:00:48,360
Cerebras' flagship product
19
00:00:48,540 --> 00:00:50,460
the WSC-3 processor
20
00:00:50,760 --> 00:00:52,820
is built on a 5-nanometer process
21
00:00:52,820 --> 00:00:54,940
and boasts over 900,00
22
00:00:54,940 --> 00:00:57,360
compute cores and 44 gigabytes of
23
00:00:57,440 --> 00:00:58,040
onboard memory
24
00:00:58,320 --> 00:01:02,860
That's 52 times more cores than a single NVIDIA H100 GPU
25
00:01:03,340 --> 00:01:06,380
How does this translate to real-world performance gains
26
00:01:07,100 --> 00:01:09,960
Cerebras demonstrated the power of its WSC-2
27
00:01:09,960 --> 00:01:14,620
chips by performing atomic-scale material simulations at an unprecedented speed
28
00:01:15,400 --> 00:01:18,960
179 times faster than the world's number one supercomputer
29
00:01:19,480 --> 00:01:20,140
Frontier
30
00:01:20,280 --> 00:01:22,000
which uses 39,00
31
00:01:22,000 --> 00:01:22,760
GPUs
32
00:01:22,820 --> 00:01:26,140
What are the implications of this for material science research
33
00:01:26,460 --> 00:01:27,340
Not limited
34
00:01:27,440 --> 00:01:28,360
To traditional AI
35
00:01:28,680 --> 00:01:32,920
Cerebras achieved a 70% parameter reduction when training large
36
00:01:32,920 --> 00:01:35,360
language models on its CS-3 system
37
00:01:35,580 --> 00:01:39,160
This means companies can build more efficient and cost-effective
38
00:01:39,160 --> 00:01:39,900
LLMs
39
00:01:40,080 --> 00:01:43,040
What are some potential applications for these smaller
40
00:01:43,260 --> 00:01:44,460
more powerful models
41
00:01:45,000 --> 00:01:45,760
Aleph Alpha
42
00:01:45,960 --> 00:01:48,400
known for its luminous generative AI models
43
00:01:48,580 --> 00:01:50,600
will be the first European organization
44
00:01:50,600 --> 00:01:53,400
to deploy the Cerebras CS-3 platform
45
00:01:53,720 --> 00:01:56,640
This partnership aims to create a new class of
46
00:01:56,640 --> 00:01:57,420
compute-efficient AI systems that can be used to generate and generate data
47
00:02:07,240 --> 00:02:11,940
Cerebras is making waves with its innovative technology and strategic partnerships
48
00:02:12,360 --> 00:02:17,400
Only time will tell if they can truly disrupt the AI landscape and dethrone NVIDIA