SiMa.ai’s Palette 3.0 Targets 3.5x Throughput for Multi-Model Edge AI

Share

SiMa.ai’s Palette 3.0 lets several AI models run on a single chip, with throughput gains of up to 3.5 times according to the company. The release targets developers of drones, robots and other autonomous machines that need to run AI directly on the device rather than in a data center. The sources provide no detailed measurement protocol behind the figure.

๐Ÿ”‘ Key takeaways

  • Palette 3.0 runs several models concurrently on the MLSoC Modalix module.
  • The up to 3.5x throughput gain is tied to speculative decoding, according to the company.
  • Modalix is rated at 50 TOPS at 10 W, with an operating range of โˆ’40 ยฐC to 85 ยฐC.
  • No detailed measurement protocol accompanies the 3.5x figure in the available sources.
  • SiMa.ai raised $150 million in a Series C at a $1.45 billion valuation in September 2026.

Concurrent execution sits at the core of Palette 3.0

The main addition in Palette 3.0 is concurrent model execution on the MLSoC Modalix module. According to SiMa.ai, a single chip can process complex reasoning workloads and several models at once while keeping a low-power profile. Modalix can also host separate Docker containers on the chip. Supported workloads include large language models (LLMs) and vision-language models.

Palette 3.0 complements Palette Neat, an agentic development environment launched in June 2026. SiMa.ai also says perception, depth and reasoning models used in drones and robots run faster with the new release. However, it does not publish detailed latency or throughput figures for each workload type.

MetricReported figureStatus
Peak throughputup to 3.5xCompany claim
Module compute50 TOPS at 10 WPromotional data
Operating temperatureโˆ’40 ยฐC to 85 ยฐCPromotional data
Module memory16 GB, runs GPT-OSS-20BCompany announcement
CompatibilityJetson Orin NX and Nano pin layoutCompany announcement

Speculative decoding and mixture-of-experts support

The 3.5x figure is tied to speculative decoding, a technique in which a small draft model proposes several tokens that a larger model then verifies in a single pass. Palette 3.0 relies on EAGLE3, DFlash and Gemma 4 MTP. The sources do not detail the exact test configuration or the baseline used to calculate the gain.

The release also supports large mixture-of-experts (MoE) models. The example cited is GPT-OSS-20B from OpenAI Group PBC, which can now run on the 16 GB Modalix MLSoC module. SiMa.ai says the open capabilities are published on GitHub for developers to adapt freely, and that version 3.0 documentation is available in its Developer Center. The commercial name of the open-source license is not specified in the sources.

Model Registry, compiler and toolchain

A new Model Registry lists model performance. SiMa.ai states that each result is measured under power conditions representative of real-world use, and that comparison details are shared on request. On LinkedIn, the company also refers to an “enhanced Model Browser” for assessing latency and power consumption in watts. The sources do not confirm whether this is the same component as the registry.

The Model Compiler supports DNN models, including generative AI models, with mixed INT8 and BF16 precision. A 2025 product brief describes a compiler that distributes models across the chip’s heterogeneous resources using optimized kernels. The document states:

“Palette’s integrated ML compiler accepts any model from an ONNX framework.”

SiMa.ai, Palette SDK product brief, 2025

The suite combines several complementary modules:

  • Application Development: samples, pre- and post-processing plugins, data ingestion connectors and support for up to 16 camera streams.
  • Device Manager: secure communication with boards, packaging, deployment and maintenance of applications.
  • Debugging: GDB for the multicore MLSoC, LogCat and PerfCat tools, and application pipeline visualization.
  • Platforms: Linux, Windows and macOS, with Yocto or eLXr embedded operating systems.

The AWS use case: from SageMaker AI to the edge

A joint post by AWS and SiMa.ai, written by Manuel Lopez Roldan of SiMa.ai and Jason Westra, senior solutions architect at AWS, links Amazon SageMaker AI with the Palette Edgematic platform. The example uses YOLOv7 to detect people and personal protective equipment (PPE) in warehouses, construction sites and production facilities. Inference runs on MLSoC devices to avoid cloud latency, keep data on site and generate real-time alerts.

The process has two phases. Training and export use SageMaker to train and validate the model, then SiMa.ai SDKs to load, quantize, test and compile it from PyTorch, TensorFlow or ONNX. The compiled artifacts are then transferred to Edgematic, deployed on the device and evaluated on target hardware. The low-code, no-code approach is presented as reducing update management, toolchain maintenance and testing effort.

  1. Prepare the AWS environment: account, AWS CLI, Docker, Git, an IAM user with the required permissions, and access to the SiMa.ai Developer Portal, in the us-east-1 Region.
  2. Provision the Palette image for SageMaker, which exceeds 25 GB and is stored in a private Amazon ECR repository before being attached to SageMaker JupyterLab. On a 20 Mbps connection, the Docker push can take several hours.
  3. Train on a GPU instance with at least 16 GB of memory, such as ml.g4dn.xlarge with Fast Launch or ml.g4dn.2xlarge, with at least 30 GB of disk space.
  4. Retrieve 8,000 labeled images per class from the open-images-v7 dataset using the fiftyone library.
  5. Adjust YOLOv7 to recognize heads and PPE, then run graph surgery, quantization, compilation, transfer to Edgematic, deployment and validation of real-time performance and accuracy.

The source states that 75% of the images are used for training, but that sentence is truncated, so the exact split must be verified in the notebook. The AWS post stresses fully local execution:

“This workflow makes sure AI applications run entirely on the edge without needing continuous cloud connectivity, decreasing latency issues, reducing security risks, and keeping data in-house.”

Manuel Lopez Roldan and Jason Westra, AWS blog post

SiMa.ai adds that a new Edgematic application can be created in a few minutes. The cloud remains involved in model preparation and deployment management; only the resulting application runs on the edge without continuous connectivity.

Private funding, partners and a measured reception

SiMa.ai is privately held. It raised $150 million in a Series C round in September 2026 at a $1.45 billion valuation. Fidelity Management & Research and Amplify Partners co-led the round, bringing total fundraising to $500 million. Named customers and partners include Robert Bosch GmbH, Emerson Electric Co. and Synopsys Inc. As the company is not listed, the sources report no stock-market reaction.

On LinkedIn, where SiMa.ai has 27,561 followers, the Palette 3.0 announcement showed 253 comments at the time of capture, with no breakdown or sentiment analysis. Phrases such as “ANY Physical AI Model. REAL Machine Autonomy.” are promotional language, not independent results. A webinar held the same day, November 17 at 9 a.m. Pacific time, featured Manuel Lopez Roldan and the Edge AI and Vision Alliance on model support, compilation and deployment of physical-AI applications.


Outlook and scenarios

Palette 3.0 reflects a broader shift: moving AI inference onto the device itself, for autonomous machines operating under tight power and temperature constraints. If the 3.5x gain holds up on independent benchmarks and real workloads, integrators of drones and robots could lower their total cost of ownership.

Several scenarios remain open. In the base case, gains are partly confirmed and vary by workload, with adoption depending on the maturity of Palette Neat and its documentation. In a cautious case, results depend heavily on the models and configurations tested, with no public protocol. In a favorable case, the edge AI ecosystem and industrial partners accelerate deployments. Investors and customers should wait for independent measurements before drawing firm conclusions.

Sources

Cet article est publiรฉ ร  titre informatif et รฉducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute dรฉcision.

Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
Telemac
Telemachttp://cryptoinfo.ch
Passionnรฉ de nouvelles technologies, jโ€™explore lโ€™univers de la blockchain et des cryptomonnaies pour partager lโ€™actualitรฉ et les innovations du secteur.

Read More

Items