Intel Myriad X AI Chip: Specs and Reality

The Intel Myriad X is a vision processing unit (VPU) designed to run artificial intelligence workloads—especially neural networks for computer vision—at the edge, meaning on the device itself rather than in the cloud. You plug it into a system via USB or M.2, feed it camera or sensor data, and it returns inference results like object classifications or depth maps. This article covers what the chip actually does, its specifications, how it compares to GPUs and other AI accelerators, and where it fits today.

Formula

Throughput (inferences per second) = 1000 ÷ inference time in milliseconds.
Compute load (GOPS) = operations per inference × (1000 ÷ inference time).

• t – inference time in milliseconds (ms), the time to process one frame or input.
• o – operations per inference, in giga-operations (billions of operations).
• perf – how many inferences the chip can complete in one second.
• gops – total giga-operations per second required for that workload.

This is a general arithmetic tool: it converts a measured latency into a throughput number and multiplies by the per-inference operation count to estimate required compute. It does not replace benchmarking.

Worked Example

Using the defaults: inference time = 10 ms, operations per inference = 1 GOPS (1 billion operations).

1. Throughput = 1000 ÷ 10 = 100 inferences per second.
2. Compute load = 1 × (1000 ÷ 10) = 100 GOPS.

So a workload with 1 Giga-op per inference and a 10 ms latency needs 100 GOPS of sustained compute and can process 100 frames per second. That is the kind of arithmetic you would use when comparing a Myriad X (rated at roughly 1 TOPS, or 1000 GOPS) against a GPU.

Background: What Is the Myriad X?

Intel introduced the Myriad X in 2017 as the third generation of its Myriad VPU line, after Myriad 1 and Myriad 2. It was designed for on-device deep learning inference, not training. The chip combines a set of general-purpose CPU cores with a dedicated neural compute engine—a hardware block that accelerates convolutional neural networks (CNNs). It also includes image signal processing and a hardware encoder, making it attractive for drones, smart cameras, and robotics.

Intel marketed the Myriad X at around 1 TOPS (trillion operations per second) of compute for deep learning, with a power envelope of roughly 1–2 watts. That made it far more power-efficient than a typical GPU for edge vision tasks. The chip was sold as part of the Intel Neural Compute Stick 2 (NCS2), a USB dongle, and also appeared in embedded systems like the Intel Vision Accelerator Design products.

How the Myriad X Works

The Myriad X integrates several components on a single die:

• Neural Compute Engine – a dedicated accelerator for CNN inference, handling operations like convolutions and activations.
• 2× SHAVE processors – vector processors that run custom vision and signal-processing code.
• 4× ARM Cortex-A53 cores – general-purpose CPUs for control and non-neural tasks.
• Image Signal Processor (ISP) – handles camera input, including lens correction and HDR.
• Hardware encoder – encodes video streams, useful for drones and cameras.

In practice, you write a neural network in a framework like TensorFlow or Caffe, convert it to an intermediate representation using Intel’s OpenVINO toolkit, and then compile it to run on the VPU. The OpenVINO runtime loads the compiled graph onto the Myriad X and manages communication over USB or PCIe.

How to Choose: When Is Myriad X the Right Tool?

Myriad X is best for low-power, always-on vision applications that need real-time inference without a cloud connection. Typical use cases include:

• Smart security cameras that detect people or vehicles locally.
• Drones that need obstacle avoidance and object tracking in flight.
• Industrial inspection systems that classify defects on a production line.
• Retail analytics that count customers or analyze shelf stock.
• Prototyping AI applications on a Raspberry Pi or similar single-board computer.

If your project runs on battery, needs to process video at 30 fps, and cannot rely on internet, the Myriad X is a reasonable choice. If you need to train models, run large language models, or process high-resolution video with complex networks, a GPU or a newer AI accelerator will be far faster.

Common Problems and Limitations

The most frequent complaint with Myriad X is its limited support for neural network layers. Because the hardware is fixed-function, any operation not implemented in the neural compute engine must run on the SHAVE processors or CPU cores, which can slow inference dramatically. Some modern layers—like certain attention mechanisms or advanced transformers—may not be supported at all by OpenVINO’s Myriad plugin.

Another issue is that the Myriad X is now old. Intel has moved on to newer VPUs like the Keem Bay (Myriad 3) and more recently the Intel AI Boost integrated into Core Ultra processors. The NCS2 is discontinued, and software support for Myriad X is in maintenance mode. If you are starting a new project, you may find better performance and longer support from a newer platform.

Also, the 1 TOPS figure is peak, not sustained. Real-world throughput depends heavily on the network architecture, batch size, and input resolution. A lightweight MobileNet might run at 100+ fps, while a heavy ResNet-50 could drop to 10–20 fps.

Comparison: Myriad X vs. GPU vs. Other VPUs

Compared to a desktop GPU like an NVIDIA GTX 1650, the Myriad X is far slower: a GPU can deliver tens of TOPS, but consumes 75 watts or more. The Myriad X uses about 1–2 watts, making it over an order of magnitude more power-efficient per watt for simple vision models. However, the GPU can handle much larger and more complex networks, and is better for training.

Against other edge AI accelerators, the Myriad X was competitive at launch but has been surpassed. Google’s Coral (Edge TPU) offers similar power efficiency and is easier to use with TensorFlow Lite. NVIDIA’s Jetson Nano provides more compute and flexibility at higher power. The Myriad X’s advantage is its integrated ISP and video encoder, which the Coral lacks.

Costs and Availability

The Intel Neural Compute Stick 2 originally retailed around $70, but it has been discontinued and is now only available through resellers, often at inflated prices. Embedded modules with the Myriad X are also scarce. If you are buying new, you will likely pay a premium for old stock. For new designs, consider the Intel Vision Accelerator Design with Keem Bay or a competing product.

Development costs include the OpenVINO toolkit, which is free, but you may need to purchase a supported camera and a host board. The NCS2 works with any Linux, Windows, or macOS machine with a USB 3.0 port, so a Raspberry Pi 4 is a common low-cost host.

Safety and Security Considerations

When deploying Myriad X in a product, keep in mind that the VPU is not a secure enclave. It does not have hardware security features like encrypted storage or secure boot. If your application processes sensitive video, you must encrypt the data before it reaches the chip and secure the host system.

Also, the Myriad X is not certified for safety-critical applications like autonomous driving or medical devices. For those, you need a processor with functional safety certification (e.g., ISO 26262). The Myriad X is intended for consumer and industrial vision, not life-safety systems.

Alternatives to Myriad X

If you are evaluating edge AI chips today, consider these options:

• Google Coral Edge TPU – USB stick or module, great for TensorFlow Lite models, low power.
• NVIDIA Jetson Nano / Xavier NX – full Linux computer with GPU, more powerful and flexible.
• Intel Keem Bay (Myriad 3) – successor to Myriad X, better performance, but also discontinued in favor of integrated AI.
• Intel Core Ultra with AI Boost – integrated NPU in laptop CPUs, good for lightweight AI on the host.
• Hailo-8 – dedicated edge AI chip with high TOPS per watt, used in some automotive and industrial products.

Each has its own SDK and model support, so the best choice depends on your framework and deployment constraints.

Frequently Asked Questions

Is the Intel Myriad X still supported?

Intel continues to provide OpenVINO support for Myriad X, but it is in legacy status. Newer versions of OpenVINO may drop or deprecate the Myriad plugin. For new projects, assume limited future updates and plan for migration.

Can I use Myriad X for training neural networks?

No. The Myriad X is designed for inference only. Training requires backpropagation and large memory, which the VPU does not support. You train on a GPU or cloud, then deploy the trained model to the Myriad X.

What is the difference between Myriad X and Myriad 2?

The Myriad X adds a dedicated Neural Compute Engine, which accelerates CNN inference significantly. Myriad 2 relies entirely on SHAVE processors for all neural network operations, so it is slower and less power-efficient for deep learning. Myriad X also has a higher clock speed and more memory bandwidth.

Does Myriad X work with TensorFlow?

Yes, through OpenVINO. You can convert a TensorFlow model to OpenVINO’s Intermediate Representation (IR) format using the Model Optimizer, then run it on the Myriad X. The same applies to Caffe, ONNX, and PyTorch (via ONNX).

How much power does the Myriad X consume?

The Myriad X is rated at about 1–2 watts under typical load, and can go lower in idle. The NCS2 draws power from the USB port, so it stays within USB 3.0’s 4.5W limit. Actual power depends on workload and host.

Can I use multiple Myriad X sticks together?

Yes, OpenVINO supports splitting a network across multiple NCS2 devices, but the performance scaling is not linear. Communication overhead and load balancing limit gains. For heavy workloads, a single GPU is usually more practical.

Related

• SNMPv2c vs SNMPv3 Security Differences & Upgrade
• MacBook Touch Bar Not Showing? Fixes That Work
• Thermostat Batteries Dead: What Happens & What to Do

in t | Inference time | ms | 10 | 0.1 | 60000
in o | Operations per inference | Giga-ops | 1 | 0.001 | 100000
out perf | Throughput (inferences/s) | inf/s | 1000 / t | 2
out gops | Compute load | GOPS | o * (1000 / t) | 2