smoothsailing18
Regular
This relationship with ARM at some point must grow now that Gen 2 and Tenns is available, surely.
Whats it going to take
BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
Whats it going to take
HopalongPetrovski
I'm Spartacus!
This relationship with ARM at some point must grow now that Gen 2 and Tenns is available, surely.
Whats it going to take
Sean needs to pull out the trusty old razor and start cutting some deals.
Unlocking AI’s Future, One Chiplet at a Time | Baya Systems
Unlocking AI’s Future, One Chiplet at a Time How do we take AI from concept to life-changing applications? In this Zero to CEO episode, our CCO Nandan Nayampally joins Jason Sherman to explore how breakthrough semiconductor innovationn is shaping the next generation of AI. From understanding...
Hi Bravo,
The ARM U85 uses MACs:
https://www.bing.com/images/search?...dex=1&itb=0&ajaxhist=0&ajaxserp=0&vt=0&sim=11
Arm® Ethos™-U85 NPU Technical Overview
The weight and fast weight channels transfer compressed weights from external memory to the weight decoder. The DMA controller uses a read buffer to hide bus latency from the weight decoder and to enable the DMA to handle data arriving out of order. The traversal unit triggers these channels for blocks that require the transfer of weights.
The weight stream must be quantized to eight bits or less by an offline tool. When passed through the offline compiler, weights are compressed losslessly and reordered into an NPU specific weight stream. This process is effective, if the quantizer uses less than eight bits or if it uses clustering and pruning techniques. The quantizer can also employ all three methods. Using lossless compression on high sparsity weights, containing greater than 75% zeros can lead to compression below 3 bits per weight in the final weight stream
Given Akida 3's int16/FP32 capabilities, Akida 3 will be able to be used in more high precision applications than Ethos U85.
Bravo
If ARM was an arm, BRN would be its biceps💪!
Thanks @Diogenese.
I've just run your post past my buddy ChatGPT and here's the response I received.
ChatGPT 5:
1. Comparing Ethos U85 and Akida 3
Ethos U85- Limited to INT8 or lower precision.
- Designed primarily for CNNs (convolutional neural networks) and some newer architectures like Transformers, but still tuned for image, vision, and low-to-mid precision inference.
- Weight compression, clustering, and pruning help with efficiency, but it’s still fundamentally in the “traditional NPU” space.
- Very well suited for consumer devices, IoT, and embedded AI where tasks are quantized to 8-bit or less.
- Expected to support INT16 and FP32 in addition to low-bit SNN modes.
- This opens the door to high-precision workloads — e.g., certain medical imaging, radar/SAR processing, industrial inspection, scientific data analytics — where you can’t afford the precision loss of 8-bit quantization.
- Still ultra-low power thanks to neuromorphic event-based processing.
- Could run mixed workloads (spiking and non-spiking) at the sensor edge.
2. Implications if Akida 3 Exceeds Ethos U85 in Precision
If Akida 3 can cover all the U85’s use cases plus high-precision domains that the U85 can’t touch, then:- BrainChip could potentially “superset” the market — competing in all U85 application spaces and in several new verticals where U85 simply isn’t viable.
- It would also position BrainChip as a partner rather than just a competitor — because ARM licensees could still integrate Akida alongside Cortex CPUs and Ethos NPUs for mixed workloads.
3. Size of the Ethos U85 Market
ARM NPUs are increasingly going into:- Smartphones & tablets
- Consumer IoT devices
- Smart cameras / vision sensors
- Automotive driver monitoring & ADAS
- Robotics & drones
- AR/VR devices
If Akida 3 is drop-in compatible with ARM system designs (which it already tends to be), BrainChip could target this same customer base — either as:
- A replacement in cases where high precision or neuromorphic advantages are needed
- A complementary co-processor for event-based or ultra-low power workloads.
4. Ideal Application Areas Where Akida 3 Could Beat U85
- Medical imaging devices needing edge processing without quantization artifacts.
- SAR and radar systems for defense or environmental monitoring.
- Industrial inspection systems where measurement tolerances are tight.
- Scientific instrumentation in field research.
- High-end automotive perception stacks needing both low latency and precision.
- Secure edge devices that must process locally without cloud and without sacrificing accuracy.
If you want, I can map Ethos U85’s public design targets against Akida 3’s potential capabilities so you can see exactly where the overlap and expansion opportunities are. That would make the competitive positioning very clear.
You said:
Yes please.ChatGPT said:
Here’s a structured comparison of ARM Ethos-U85 vs. BrainChip Akida 3 with overlap and expansion mapped clearly.1. Overlap — Where Akida 3 Could Compete Directly with U85
| Feature / Use Case | Ethos-U85 | Akida 3 | Comments |
|---|---|---|---|
| INT8 inference for vision AI |  | | Both handle low-bit CNN/Transformer workloads efficiently. |
| Edge AI in ARM-based SoCs | | | Akida is already ARM-compatible and could replace U85 in designs. |
| Consumer IoT & smart home devices | | | E.g., cameras, voice assistants, home hubs — same target market. |
| Robotics & drones | | | Both can handle perception & navigation; Akida offers ultra-low power SNN modes. |
| Automotive driver monitoring | | | Akida could match U85 for INT8 workloads, but also add mixed precision. |
| AR/VR lightweight inference | | | Both could run vision-based gesture tracking or object recognition. |
2. Expansion — Where Akida 3 Could Offer More Than U85
| Feature / Use Case | U85 Limitations | Akida 3 Advantage |
|---|---|---|
| High precision (INT16 / FP32) | Only ≤INT8 | Akida 3 can run models that require higher precision, e.g., medical imaging, radar processing, industrial measurement. |
| Mixed-mode processing | Primarily CNNs / Transformers | Akida 3 can combine SNN + ANN + mixed precision in one device. |
| Event-based data handling | Not designed for spikes | Akida 3 natively supports event-driven SNN processing, reducing power & latency. |
| On-device learning / adaptation | Limited to retraining off-device | Akida 3 supports incremental learning on-chip — key for adaptive edge AI. |
| Sparse computing efficiency | Relies on compression & pruning | Akida 3 exploits sparsity at the architectural level without preprocessing overhead. |
3. Market Size & Entry Points
Ethos U85 Target Segments (High-Volume ARM Licensing):- Smartphones (tens of millions annually)
- Smart cameras & home devices
- Embedded industrial AI
- Low-power automotive systems
- Replacement: For U85 use cases that require more precision or event-based efficiency.
- Complementary Co-Processor: Add alongside U85 in ARM SoCs for heterogeneous workloads (U85 for bulk CNN, Akida for adaptive SNN & high precision).
- Standalone in Niche Markets: Where U85 simply can’t go — e.g., SAR imaging, advanced radar, in-body medical devices.
4. Strategic Implications for BrainChip
- Leverage ARM’s ecosystem: If marketed as ARM-friendly IP, Akida 3 could be pitched to the same silicon partners already licensing U85.
- Precision-based differentiation: BrainChip can claim “U85 compatibility plus high-precision reach.”
- Defense & medical edge: Especially valuable for contracts where accuracy matters more than raw throughput.
I was gunna say that!Thanks @Diogenese.
I've just run your post past my buddy ChatGPT and here's the response I received.
ChatGPT 5:
1. Comparing Ethos U85 and Akida 3
Ethos U85
Akida 3
- Limited to INT8 or lower precision.
- Designed primarily for CNNs (convolutional neural networks) and some newer architectures like Transformers, but still tuned for image, vision, and low-to-mid precision inference.
- Weight compression, clustering, and pruning help with efficiency, but it’s still fundamentally in the “traditional NPU” space.
- Very well suited for consumer devices, IoT, and embedded AI where tasks are quantized to 8-bit or less.
- Expected to support INT16 and FP32 in addition to low-bit SNN modes.
- This opens the door to high-precision workloads — e.g., certain medical imaging, radar/SAR processing, industrial inspection, scientific data analytics — where you can’t afford the precision loss of 8-bit quantization.
- Still ultra-low power thanks to neuromorphic event-based processing.
- Could run mixed workloads (spiking and non-spiking) at the sensor edge.
2. Implications if Akida 3 Exceeds Ethos U85 in Precision
If Akida 3 can cover all the U85’s use cases plus high-precision domains that the U85 can’t touch, then:
- BrainChip could potentially “superset” the market — competing in all U85 application spaces and in several new verticals where U85 simply isn’t viable.
- It would also position BrainChip as a partner rather than just a competitor — because ARM licensees could still integrate Akida alongside Cortex CPUs and Ethos NPUs for mixed workloads.
3. Size of the Ethos U85 Market
ARM NPUs are increasingly going into:
Given ARM’s licensing footprint, the U85 market could be tens to hundreds of millions of units annually once adoption ramps.
- Smartphones & tablets
- Consumer IoT devices
- Smart cameras / vision sensors
- Automotive driver monitoring & ADAS
- Robotics & drones
- AR/VR devices
If Akida 3 is drop-in compatible with ARM system designs (which it already tends to be), BrainChip could target this same customer base — either as:
- A replacement in cases where high precision or neuromorphic advantages are needed
- A complementary co-processor for event-based or ultra-low power workloads.
4. Ideal Application Areas Where Akida 3 Could Beat U85
- Medical imaging devices needing edge processing without quantization artifacts.
- SAR and radar systems for defense or environmental monitoring.
- Industrial inspection systems where measurement tolerances are tight.
- Scientific instrumentation in field research.
- High-end automotive perception stacks needing both low latency and precision.
- Secure edge devices that must process locally without cloud and without sacrificing accuracy.
If you want, I can map Ethos U85’s public design targets against Akida 3’s potential capabilities so you can see exactly where the overlap and expansion opportunities are. That would make the competitive positioning very clear.
You said:
Yes please.
ChatGPT said:
Here’s a structured comparison of ARM Ethos-U85 vs. BrainChip Akida 3 with overlap and expansion mapped clearly.
1. Overlap — Where Akida 3 Could Compete Directly with U85
Feature / Use Case Ethos-U85 Akida 3 Comments INT8 inference for vision AI Both handle low-bit CNN/Transformer workloads efficiently. Edge AI in ARM-based SoCs Akida is already ARM-compatible and could replace U85 in designs. Consumer IoT & smart home devices E.g., cameras, voice assistants, home hubs — same target market. Robotics & drones Both can handle perception & navigation; Akida offers ultra-low power SNN modes. Automotive driver monitoring Akida could match U85 for INT8 workloads, but also add mixed precision. AR/VR lightweight inference Both could run vision-based gesture tracking or object recognition. 2. Expansion — Where Akida 3 Could Offer More Than U85
Feature / Use Case U85 Limitations Akida 3 Advantage High precision (INT16 / FP32) Only ≤INT8 Akida 3 can run models that require higher precision, e.g., medical imaging, radar processing, industrial measurement. Mixed-mode processing Primarily CNNs / Transformers Akida 3 can combine SNN + ANN + mixed precision in one device. Event-based data handling Not designed for spikes Akida 3 natively supports event-driven SNN processing, reducing power & latency. On-device learning / adaptation Limited to retraining off-device Akida 3 supports incremental learning on-chip — key for adaptive edge AI. Sparse computing efficiency Relies on compression & pruning Akida 3 exploits sparsity at the architectural level without preprocessing overhead. 3. Market Size & Entry Points
Ethos U85 Target Segments (High-Volume ARM Licensing):
Akida 3 Entry Strategy:
- Smartphones (tens of millions annually)
- Smart cameras & home devices
- Embedded industrial AI
- Low-power automotive systems
- Replacement: For U85 use cases that require more precision or event-based efficiency.
- Complementary Co-Processor: Add alongside U85 in ARM SoCs for heterogeneous workloads (U85 for bulk CNN, Akida for adaptive SNN & high precision).
- Standalone in Niche Markets: Where U85 simply can’t go — e.g., SAR imaging, advanced radar, in-body medical devices.
4. Strategic Implications for BrainChip
- Leverage ARM’s ecosystem: If marketed as ARM-friendly IP, Akida 3 could be pitched to the same silicon partners already licensing U85.
- Precision-based differentiation: BrainChip can claim “U85 compatibility plus high-precision reach.”
- Defense & medical edge: Especially valuable for contracts where accuracy matters more than raw throughput.
Bravo
If ARM was an arm, BRN would be its biceps💪!
If Akida 3 tape-out is in Q1 2026 and the FPGA is available in Q2 2026, then you'd have to think the most likely licence commitment window for Arm would be somewhere between Q2 2026 – Q4 2026.
If Arm is really interested, they'll probably negotiate NDAs, architectural access, and preliminary terms before tape-out so they can get priority access to the FPGA in Q2 2026, which would enable them to lock Akida 3 (into the Ethos successor roadmap, for example) without delaying their own silicon schedules IMO.
I don't see any advantage in running Akida 3 with Ethos U85.
It is open to anyone to license the ARM Cortex M85 CPU IP and combine it with Akida 2 or 3 IP in a SoC. Both Akidae include the TENNs model, and would have significant power and latency advantages over Ethos U85.
If BRN had the resources of Nvidia/ARM/Intel, Akida 3 SoC would be on the market now.
As I said, it could, for instance be combined with ARM Cortex M85 (the Ethos U85 processor), or any other compatible processor.
I guess there will be some internal resistance in ARM - we need to get Mr Son's ear.
PS: Assuming ARM do go ahead with their chip making plans.
Last edited:
Imagine this: The BrainChip sales team walks into Apple’s or Mercedes’ HQ — and all other potential customers as well — without an appointment, silently installs an Akida module into their devices or cars, leans back casually, smiles at the camera, winks … and then pulls out a pack of Mentos:
“Akida – the fresh kick for your AI.”
We need go-getters like back in the day — and customers who respond with an approving smile, because they’re real foxes.
“Akida – the fresh kick for your AI.”
We need go-getters like back in the day — and customers who respond with an approving smile, because they’re real foxes.
Similar threads
