T ThinkEdge Learn

DGX Spark: What the Hardware Actually Is

DGX Spark is a small box whose whole reason to exist is memory: 128 GB that the CPU and the GPU both address directly. Understanding why that matters tells you more than the petaFLOP figure does.

dgx spark nvidia dgx spark project digits ai dev kit
AI and Electronics · Capacity versus bandwidth, and what each number really means

The hardware, plainly

The DGX Spark is a desktop box with one main chip in it. That chip, the GB10 Grace Blackwell Superchip, was co-designed with MediaTek and puts a 20-core Arm CPU, 10 Cortex-X925 and 10 Cortex-A725, next to a Blackwell GPU with 6,144 CUDA cores and fifth-generation tensor cores. The CPU and GPU share 128 GB of LPDDR5x memory through a coherent interconnect, so a program on the CPU can point at the same bytes the GPU is using, with no copy in between.

It was shown as Project DIGITS at CES on 6 January 2025 with a $3,000 price and a May release date, renamed DGX Spark at GTC on 18 March 2025, and shipped on 15 October 2025 at $3,999 for the 4 TB version. A 64 GB version sold through OEM partners came later. It uses a 240 W power supply, the chip itself is rated at 140 W, and the whole unit is 150 mm square and weighs 1.2 kg.

Why memory, not FLOPs, decides what you can run

The headline number is up to 1 petaFLOP of FP4 performance, and it is a theoretical figure that assumes 4-bit numbers and sparsity. The number that decides your daily experience is 273 GB/s of memory bandwidth. Every token the model generates needs the weights read from memory, so tokens per second is set by how fast those bytes arrive, not by how many multiply-adds the tensor cores can perform.

That is why the box is described as a capacity machine rather than a speed machine. Reviews of a 120B-parameter open model on a Spark report around 35 to 40 tokens per second, and one unit can hold roughly 200B parameters at 4-bit in its 128 GB, or about 405B with two units linked over the 200 Gbps ConnectX port. For a school the capacity is the argument: a model that fits entirely in memory, with no request leaving the building and no usage fee.

Step by step

  1. Separate the two numbers before you buy

    Capacity is the 128 GB of shared memory. Speed is the 273 GB/s bandwidth and the tokens per second you actually get. Quoting the petaFLOP figure to a parent instead of the bandwidth figure is the most common mistake in this conversation.

  2. Work out what fits

    At 4-bit, budget about 0.7 GB per billion parameters, so 128 GB of shared memory holds a model of roughly 180 to 200 billion parameters, with room left for the context and the operating system.

  3. Check the power and the heat

    The chip is rated at 140 W with a 240 W supply, and a box that size needs airflow around it. In a lab with 20 of them, the power and cooling budget is a bigger number than the purchase price.

  4. Budget the price honestly

    It launched at $3,999 for 4 TB, a 64 GB version appeared around $3,150 excluding VAT through OEM channels in late 2026, and NVIDIA's own marketplace listing for the 4 TB unit shows $6,950. Prices and stock move, so check the verified retail price rather than the launch figure.

  5. Decide what the classroom uses it for

    Local inference, fine-tuning a small model on your own data, and dataset work are realistic. Training a large model from scratch is not, and a petaFLOP figure will not change that.

  6. Pair it with a microcontroller

    Use the Spark for the slow, fuzzy decisions and an Arduino or Mega for the millisecond loop. That split is the one you would design on a factory floor, and it is the part students can actually build.

What you'll need

ItemQtyWhy
NVIDIA DGX Spark1The 128 GB unified-memory box. Choose the 64 GB or 128 GB version based on the model size you actually need, not on the headline.
Monitor with USB-C or USB-C to HDMI adapter1The unit has one HDMI 2.1a and up to three DisplayPort over USB-C, so check what the classroom already has before ordering.
Arduino Mega 25601Handles the sensors and the real-time control, and has the extra serial ports for logging sensor data to the Spark.
HC-SR04 ultrasonic sensor1The observation the model reasons about before the decision comes back over serial.
LED bar graph module1Shows the model's decision as a number you can check rather than a word you have to trust.
10 GbE network cable1The Spark has 10 GbE and a ConnectX-7 port at 200 Gbps, which is how two units share one large model.

Practise it before you build it

Every step above has a matching circuit in the ThinkEdge Learn simulator. The simulator runs a real Arduino sketch against a modified-nodal-analysis solver, so digitalWrite actually drives the LED on screen and the readings change when you change the wiring — no hardware, no burnt components.

Common questions

Is DGX Spark worth it for a school lab?

For most schools, no. It launched at $3,999 and, even at the 64 GB size, it is still several thousand dollars for hardware that most class projects never approach. A used workstation with 24 GB of graphics memory, or even a shared cloud credit, will carry a school project further for a fraction of the cost. The case for the Spark is capacity and privacy: a large model resident in memory, running on a machine inside your building.

Can a DGX Spark replace a data centre?

No, and the difference is worth teaching. A Spark has one chip and a few hundred gigabytes per second of bandwidth. A training cluster has thousands of graphics processors linked together. What a Spark can do that a laptop cannot is hold a model of a few hundred billion parameters in memory and serve it locally, which is a genuine capability for a lab, just a narrow one.

How fast is it in practice?

Around 35 to 40 tokens per second for a 120B-parameter class model running locally on a 273 GB/s memory bus. Since generation is bound by reading the model weights, a smaller model is not automatically faster, and the software releases since launch have improved throughput noticeably. Benchmark the model you intend to run, not the headline figure.

Related guides

RTX 5080: The Card, the Numbers, the CaveatRTX 5080 specs and price: 16 GB of GDDR7, 360 W of board power and a $999 launch price that no longer matches the street. Here is who should buy one. Local LLM: Your Own Model, Your Own MachineRun a language model on your own hardware: the RAM maths for a Raspberry Pi, the real token speeds, and how an Arduino can act on what the model says. Edge AI: Decisions Made on the DeviceEdge AI runs the decision on the device instead of in the cloud. Here is what that buys you in latency and privacy, and what a 32 KB microcontroller can really do. What an AI Agent Actually IsAn AI agent is a loop of model calls, tools and memory. Learn how that loop maps onto the sense, decide, act cycle of a robot you can build this term. GPU Repair: Why Graphics Cards Fail and Whether Fixing One PaysGraphics cards fail from solder fatigue, VRM breakdown, dead fans and ageing capacitors. How to diagnose the cause and whether a repair is worth the money.

Try the full simulator with your class

103 guided circuit projects, 21 lessons and an in-browser Arduino simulator. The first 15 projects are free, no account needed. NCERT Science mapping on 47 projects, with Arabic and English copy in the regional builds.