[CMU AI Seminar] September 27 at 12pm (NSH 3305 & Zoom) -- Jian Zhang (SambaNova Systems) -- MLSys Innovations Beyond the Hardware Goldilocks Zone -- AI Seminar sponsored by SambaNova Systems

[Asher Trockman]

Dear all,

We look forward to seeing you *this Tuesday (9/27)* from *1**2:00-1:00 PM
(U.S. Eastern time)* for the next talk of this semester's *CMU AI Seminar*,
sponsored by SambaNova Systems <https://sambanova.ai/>. The seminar will be
held in NSH 3305 *with pizza provided *and will be streamed on Zoom.

To learn more about the seminar series or to see the future schedule,
please visit the seminar website <http://www.cs.cmu.edu/~aiseminar/>.

On 9/27, *Jian Zhang* (SambaNova Systems) will be giving a talk titled
*"**MLSys
Innovations Beyond the Hardware Goldilocks Zone**" *to share
SambaNova's recent advances in software/hardware co-design for sparse
training (e.g., of large language models).

*Title*: MLSys Innovations Beyond the Hardware Goldilocks Zone

*Talk Abstract*: The Goldilocks zone, or habitable zone, limits life to a
restricted part of the solar system. In a similar way, the limitations of
conventional hardware impose a Goldilocks zone for ML system innovations.
In recent years, the emerging deep learning accelerators have launched an
unprecedented opportunity for ML systems advance at different layers of the
software and machine learning stacks. In this tech talk, I am very excited
to share a case study for building large scale models on the reconfigurable
dataflow units (RDU) at SambaNova Systems. At the ML application layer, I
want to highlight the 0-1 accuracy breakthrough in High Resolution 3D
segmentation enabled by the large memory capacity in RDU. In the layer of
ML training algorithms, I will unbox our recent advance in SW/HW codesign
for RDU sparse training; this endeavor leads to a 6X faster
time-to-accuracy for pretraining large language models over standard dense
training on A100 GPUs. Lastly, I will showcase how our team built the
prototype of learned RDU performance optimization methods, which pushes
further towards fully unleashing the hardware capability of RDU with
significantly reduced engineering cost than rule-based methods. We hope
that the emerging accelerators will trigger many more ML system innovations
from the community.

*Speaker Bio*: Jian Zhang is the Director of Machine Learning at SambaNova
Systems. He leads the ML team which builds the deep learning foundations
for SambaNova’s large-scale enterprise AI solutions. With the mission of
democratizing modern foundation model systems, the ML team at SambaNova
innovates on both the machine learning and the system aspects,
including productionalizing large foundation models and ML/hardware
co-design on emerging hardware. Before joining SambaNova Systems, Jian got
his PhD in machine learning from Stanford University focusing on machine
learning and natural language processing systems.

*In person: *NSH 3305
*Zoom Link*:
https://cmu.zoom.us/j/99510233317?pwd=ZGx4aExNZ1FNaGY4SHI3Qlh0YjNWUT09

Thanks,
Asher Trockman
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/ai-seminar-announce/attachments/20220925/3318de5d/attachment.html>