BRN Discussion Ongoing

Frangipani

Top 20

"the MATRIX AI Consortium for Human Well-Being at UT San Antonio plans to launch a new initiative that establishes a national hub for “neuromorphic” computing available for public use."

"The initiative, called THOR: The Neuromorphic Commons, is funded by the National Science Foundation. THOR will make the promising technology available for researchers nationwide to explore and conduct experiments, serving as the largest-ever full-stack neuromorphic platforms to be open to the public."

"Historically, access to large-scale neuromorphic hardware has been exclusive to well-funded industry labs or select universities.

THOR upends that paradigm by making these powerful systems accessible to researchers and students across the United States, said Pandit.

“Our goal is to host neuromorphic hardware developed by the THOR team and our partners, building a specialized center where the community can learn and develop the next generation of neuromorphic systems,” he said.

Much like a public library, anyone can apply for access, and the resources will be free to use. Researchers will be able to enter a queue to run their experiments on the hardware. Once a user finishes his or her work, the system becomes available for the next person, allowing for high utilization of the resource."

"The official launch of THOR is slated for February 23 in the UT San Antonio San Pedro I building, with plans for a live demonstration to showcase the hardware capabilities. Beyond the hardware, the initiative focuses heavily on training and education."

Brainchip is a part of the ecosystem:


Hi @SharesForBrekky,

while the THOR project website lists Akida alongside 16 other examples of neuromorphic hardware, only three of them are currently marked as being accessible through THOR: SpiNNaker 2, BrainScaleS-2 as well as HiAER-Spike.


F207550D-3677-4649-8FED-F7B68DBBF803.jpeg
1D3A1535-1940-483D-9FD6-F332D15362C1.jpeg
f747dde6-85f1-49ea-a4e4-ebf0bef1ff1e-jpeg.95360



The idea behind the THOR project is to give the public access to large-scale neuromorphic hardware.


1EE5667F-06B5-4EB0-98B6-8448F889AE01.jpeg



4F8B21FB-ACBD-4B67-ABC8-90974A60AABD.jpeg



While BrainScaleS-2 may look out of place at first glance, the linked paper explains “This single-core system can serve as the unit of scale for larger-scale designs involving multiple neuromorphic cores.”





From the article you shared:

“Library for supercomputing​

Historically, access to large-scale neuromorphic hardware has been exclusive to well-funded industry labs or select universities.

THOR upends that paradigm by making these powerful systems accessible to researchers and students across the United States, said Pandit.

“Our goal is to host neuromorphic hardware developed by the THOR team and our partners, building a specialized center where the community can learn and develop the next generation of neuromorphic systems,” he said.

Much like a public library, anyone can apply for access, and the resources will be free to use. Researchers will be able to enter a queue to run their experiments on the hardware. Once a user finishes his or her work, the system becomes available for the next person, allowing for high utilization of the resource.

The team estimates that around 50 researchers could use the resource at once, but the precise number will vary depending on the complexity of the programs they are testing.

Powering the future​

At the core of THOR is the SpiNNaker2 system, developed in partnership with SpiNNcloud.

SpiNNcloud is a massive computing platform that utilizes roughly 400,000 highly parallel processing elements, making it one of the largest neuromorphic systems available for public research. Parallelism refers to a system’s capacity to complete a complex task quickly by breaking it down into smaller tasks and executing them simultaneously.

The system uses ARM-based cores — the same energy-efficient technology found in smartphones — specifically designed to simulate neurons and synapses. The architecture allows researchers to build “spiking neural networks” that process information in pulses, similar to how the brain signals muscles to move or eyes to process light.

Recent coverage from the publication HPCwire notes that the deployment places UT San Antonio among the top tier of institutions globally hosting such large-scale neuromorphic capacity.”
 

Attachments

  • F747DDE6-85F1-49EA-A4E4-EBF0BEF1FF1E.jpeg
    F747DDE6-85F1-49EA-A4E4-EBF0BEF1FF1E.jpeg
    349.7 KB · Views: 99
  • Fire
  • Like
Reactions: 2 users

Frangipani

Top 20
Hi Bravo,

The synchronous/asynchronous thing ... I think the big takeaway is that asynchronous has much lower latency, the response being practically instantaneous. It's like waiting for a bus compared with having an Uber driver as a neighbour, or better yet, having your own e-bike.

Synchronous is a hangover from von Neumann. It is very familiar to engineers and programmers. Old school AI programmers cut their teeth on CNN running on synchronous processors (ARM/Intel/Nvidia) But that is why BRN has MetaTF. It provides an almost seamless transition using software familiar to such programmers. In addition, BRN does provide models which they prepared earlier, and there are BRN partners who provide model development support.

Hi @Diogenese,

you seem to be under the impression that Akida is asynchronous digital?

That’s not what BrainChip says, though…

Here are two videos, according to which Akida is synchronous digital - one from November 2023 featuring Ian Cutress and Sally Ward-Foxton, followed by Jonathan Tapson’s presentation during last week’s EDGE AI FOUNDATION Neuromorphic Livestream:






DB3FA6F0-3BBF-401F-B535-D8856A8B62F9.jpeg







9E73CC55-B05B-4665-922A-4A8779703E04.jpeg



From 1h 30 min:
“So BrainChip’s circuits are synchronous and digital and deterministic, because our customers demand that […] But nonetheless we use event-based, we use graded spikes […].”
 
  • Love
Reactions: 1 users

itsol4605

Regular
Protests against DataCenter
 
  • Wow
Reactions: 2 users

itsol4605

Regular
Top Bottom