Your article you posted doesn’t mention Neuromorphic, it actually says “AI-based robotic systems that are independent of specific hardware.”Ranting on about Megachips but see the connection below from the Megachips website between neuromorohic AI and robotics.
Contribution to Society | MegaChips Corporation
www.megachips.co.jp
"The robotic projects undertaken by MegaChips (see the press release dated Sep. 17, 2025) aim to promote the use of scalable, AI-based robotic systems that are independent of specific hardware.
Building on the expertise cultivated through joint research – such as AI-based robotic motion planning and high-speed control, as well as initiatives to enhance the safety and operational efficiency of robotic systems – we seek to develop these efforts into a competitive business."
" Through this collaboration with the Nara Institute of Science and Technology (NAIST), we will build on the experience gained from the joint research conducted through FY2024, “High-speed robotic control using a Spiking Neural Network (SNN) applied to a high-rate 3D sensor,” and continue activities to deepen our technical expertise toward commercialization."
My bold above.
Megachips likely have early access to the 1500M.2 CARD (Andes has access).
As discussed earlier the robotic software used is Acumino which is hardware agnostic so its likely that the will use both AKIDA and Quadric in order to get the high human like dexterity they aim for in robotics.
BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
Contribution to Society | MegaChips Corporation
www.megachips.co.jp
Its says Spiking Neural Network (SNN) which is the core of Neuromorphic.
Quadric is neural only which means on chip but Traditional AI in nature.
There is room for both in Megachips/Acumino robots depending on the tasks allocated to each. Quadric would take the high volume intense calculations that AKIDA could struggle with.
That is why IMO they have chosen both as partners.
See Section:
Collaboration with the Nara Institute of Science and Technology (NAIST), then Purpose.
Hi Manny,"" Through this collaboration with the Nara Institute of Science and Technology (NAIST), we will build on the experience gained from the joint research conducted through FY2024, “High-speed robotic control using a Spiking Neural Network (SNN) applied to a high-rate 3D sensor,” and continue activities to deepen our technical expertise toward commercialization.""Contribution to Society | MegaChips Corporation
www.megachips.co.jp
Its says Spiking Neural Network (SNN) which is the core of Neuromorphic.
Quadric is neural only which means on chip but Traditional AI in nature.
There is room for both in Megachips/Acumino robots depending on the tasks allocated to each. Quadric would take the high volume intense calculations that AKIDA could struggle with.
That is why IMO they have chosen both as partners.
See Section:
Collaboration with the Nara Institute of Science and Technology (NAIST), then Purpose.
"High rate 3D sensor" sounds like an ideal application for see-in-the-dark radar.
Yes, i think a lot of what AKIDA does will be used elsewhere. Steve Brightfield said that he has had 'early' discussions with robotic companies concerning the RTX/AKIDA radar project.
I bet Megachips cannot wait for the 1500 to arrive. It chews up less than 300 milliwatts (way less than AKIDA 1000) and goes a way to address the issues of heat and battery life which is crucial for Robotics.
Its almost as if Megachips made some very pertinent suggestions after their development experience with its development of AKIDA 1000 under license.
Our IBM friend Kevin Johansen have maybe got inspiration from below Swedish university white paper called:
Comparison of Akida Neuromorphic
Processor and NVIDIA Graphics
Processor Unit for Spiking Neural
Networks
CARL CHEMNITZ
MALIK ERMIS
Degree Project in Computer Science and Engineering, First Cycle 15 credits
Date: June 9, 2025
Supervisor: Jörg Conradt
Examiner: Pawel Herman
Swedish title: Jämförelse av neuromorfisk processor Akida och NVIDIA grafikkort för Spiking
Neural Networks
School of Electrical Engineering and Computer Science
Some snippets:
Key Observations
• Neuromorphic Akida demonstrates 99.520 % (MNIST) and
95.956-99.699 % (YOLO) energy reduction compared to GPU for the same
or similar networks.
• For simpler networks, Akida processes 76.733 % faster (1.622 ms vs
7.014 ms), proving its suitability for latency critical real-time processing
tasks. However, for more complex models, the AKD1000 is outperformed
by the GTX 1080, 73.769 ms to 160.872 ms.
• Akida’s adaptive clocking (35-179 MHz) reduces clock speeds by an average
of 86.610 % versus GPUs’ relatively fixed 1746 MHz operation, reflecting its
dynamic power engagement through sparse computing.
• The sparse input patterns, utilizing Akida’s neuromorphic architecture,
achieves up to 58 times fewer clock cycles through spike-based,
asynchronous processing, demonstrating significant improvements in
computational efficiency and suitability for edge AI systems.
• Akida correlation for MNIST model show that between clock cycles and both
energy and time were essentially zero (0.0153, with a p-value of 0.497). In
contrast to the YOLOv2 model where it shows a small but definite correlation
between clock cycles and both energy consumption (0.2119) and inference
time (0.2022), with p-value below 0.05 (making the correlation plausible).
• Quantization has a major effect on all key metrics except for the clock speed
on the GPU, reducing energy consumption, throughput, and latency by 85.9-
99.9 %. This demonstrates the suitability and importance of quantization
for deploying neural networks on resource limited hardware. However, for
more complex neural networks, this comes at the cost of reduced accuracy,
highlighting critical trade-off between computational efficiency and predic-
tive performance.
Whole white paper:
Comparison of Akida Neuromorphic
Processor and NVIDIA Graphics
Processor Unit for Spiking Neural
Networks
CARL CHEMNITZ
MALIK ERMIS
Degree Project in Computer Science and Engineering, First Cycle 15 credits
Date: June 9, 2025
Supervisor: Jörg Conradt
Examiner: Pawel Herman
Swedish title: Jämförelse av neuromorfisk processor Akida och NVIDIA grafikkort för Spiking
Neural Networks
School of Electrical Engineering and Computer Science
Some snippets:
Key Observations
• Neuromorphic Akida demonstrates 99.520 % (MNIST) and
95.956-99.699 % (YOLO) energy reduction compared to GPU for the same
or similar networks.
• For simpler networks, Akida processes 76.733 % faster (1.622 ms vs
7.014 ms), proving its suitability for latency critical real-time processing
tasks. However, for more complex models, the AKD1000 is outperformed
by the GTX 1080, 73.769 ms to 160.872 ms.
• Akida’s adaptive clocking (35-179 MHz) reduces clock speeds by an average
of 86.610 % versus GPUs’ relatively fixed 1746 MHz operation, reflecting its
dynamic power engagement through sparse computing.
• The sparse input patterns, utilizing Akida’s neuromorphic architecture,
achieves up to 58 times fewer clock cycles through spike-based,
asynchronous processing, demonstrating significant improvements in
computational efficiency and suitability for edge AI systems.
• Akida correlation for MNIST model show that between clock cycles and both
energy and time were essentially zero (0.0153, with a p-value of 0.497). In
contrast to the YOLOv2 model where it shows a small but definite correlation
between clock cycles and both energy consumption (0.2119) and inference
time (0.2022), with p-value below 0.05 (making the correlation plausible).
• Quantization has a major effect on all key metrics except for the clock speed
on the GPU, reducing energy consumption, throughput, and latency by 85.9-
99.9 %. This demonstrates the suitability and importance of quantization
for deploying neural networks on resource limited hardware. However, for
more complex neural networks, this comes at the cost of reduced accuracy,
highlighting critical trade-off between computational efficiency and predic-
tive performance.
Whole white paper:
Bravo
Meow Meow 🐾
I'm trying to ascertain whether the information in this Robocloud article (below) is correct. It says that this quadruped inspection robot called ANYmal D by ANYbotics is due for shipping in Q3 2026 utilizing Intel's Loihi 3.
The only credible information I can find from Intel re Loihi 3 is Mike Davies Linkedin post from 8 months go, in which he said Loihi was still under development.
Maybe by now they're close to completion???...
Excerpt from Robocloud article describing ANYmal shipping in Q3 2026 with Intel's Loihi 3.
robocloud-dashboard.vercel.app
www.linkedin.com
This June 2025 research paper discusses ANYmal running on Loihi 2.
ieeexplore.ieee.org
Anymal works without an internet connection as shown in this video.
The only credible information I can find from Intel re Loihi 3 is Mike Davies Linkedin post from 8 months go, in which he said Loihi was still under development.
Maybe by now they're close to completion???...
Excerpt from Robocloud article describing ANYmal shipping in Q3 2026 with Intel's Loihi 3.
Learn Robotics Online | RoboCloud Hub - Free Robotics Courses, Jobs & News
Master robotics with free ROS/ROS2 tutorials, find robotics engineering jobs, read latest robotics research & news. Join 10,000+ robotics engineers learning career skills.
RoboARCH @ ICRA 2025 | Mike Davies
All the recent breakthroughs in generative AI provide great optimism that longstanding algorithmic challenges in robotics, autonomy, and embodied AI will soon be solved, but there’s one catch: conventional computer architectures stand no chance of delivering the TOPS needed to fulfill this...
This June 2025 research paper discusses ANYmal running on Loihi 2.
IEEE Xplore Full-Text PDF:
Anymal works without an internet connection as shown in this video.
Similar threads
