AI Infrastructure • Power Efficiency • Networking
The Latest DGX Spark Update Cuts Idle Power by 32%+ — Thanks to Smarter ConnectX NIC Power Management
NVIDIA’s newest DGX Spark software update introduces deeper idle power states for its 200 Gbps ConnectX-7 NIC via hot-plug support. In one widely reported set of wall-power measurements, idle draw drops from ~37W to ~25W (and even ~22W headless), making Spark far more practical as an always-on AI workstation.
What to know in 30 seconds
- What changed: DGX Spark now supports ConnectX-7 “hot-plug” power management, allowing the NIC to save power when it isn’t in use.
- Why it matters: A 200 Gbps NIC can be a meaningful chunk of idle power. Getting that power back improves 24/7 usability and lowers idle thermals/noise.
- Headline result: Reported idle power reductions reach 32% or more under certain conditions.
DGX Spark sits in an unusual category: it’s not a “gaming PC with an AI sticker,” and it’s not a rack server either. It’s a compact, developer-centric system built around NVIDIA’s GB10 platform, designed to bring serious local AI workloads—prototyping, inference, and even multi-node experiments—into a desk-friendly footprint.
But as early adopters learned, performance isn’t the only thing that defines whether a system belongs on a desk or in a lab. When you leave a machine on all day (or all week), idle behavior becomes part of the product. That includes idle thermals, fan acoustics, and—more than ever—idle power draw.
This week’s DGX Spark software update targets that exact pain point. The change is not a vague “optimizations” bullet: NVIDIA explicitly calls out a new power-management behavior for the system’s 200 Gbps ConnectX-7 NIC. In short: if you’re not using the NIC, the system should stop paying a constant energy tax for keeping it fully awake.
Why this is news
Datacenter-class networking inside a desk-class AI box is a feature—until it becomes an idle power anchor. This update aims to keep the feature while reducing the penalty.
What changed in the update
NVIDIA’s DGX Spark release notes include a highly specific line item under power management: hot-plug support for the ConnectX-7 network adapter, which can save up to 18W when the adapter is not in use.
The key phrase here is “hot-plug support.” In everyday terms, this enables the platform to more intelligently detect whether the NIC is effectively “present and active” from the system’s point of view (for example, whether it is actually being used, linked, or needed) and to place it into a deeper low-power state when it isn’t.
That matters because a 200 Gbps NIC is not a typical desktop component. It’s designed to deliver massive throughput and low latency—capabilities that can be essential for multi-node workflows, fast storage fabrics, RDMA experiments, or “stacking” systems together. But those capabilities also come with power domains that, if left in a higher state at idle, can become an outsized share of the machine’s baseline draw.
What NVIDIA is explicitly claiming
- Feature: ConnectX-7 hot-plug support (power management)
- Impact: “Saving up to 18W of power when the adapter is not in use”
- Scope: DGX Spark software updates (also referenced in NVIDIA’s forum announcement)
NVIDIA’s own forum announcement for the February 2026 DGX Spark software updates reiterates the same point: “Measurable power efficiency gains” from the new hot-plug behavior for the ConnectX-7 adapter.
The idle power numbers (and why they vary)
In coverage of the update, one widely circulated set of measurements reports a dramatic improvement in wall-power idle draw: roughly ~37W idle before versus ~25W idle after with a display connected—an improvement of about 32%+. In a headless setup (no display connected), idle reportedly dropped further to ~22W.
Those numbers are important for two reasons:
- The magnitude aligns with the release notes. NVIDIA’s “up to 18W” NIC savings fits the scale of an idle reduction in the low-teens (or mid-teens) watts, depending on how the system is configured.
- Idle is a small budget. When a machine idles at a few dozen watts, reclaiming 12–18W is not a rounding error. It’s the difference between a “always-warm mini server” and something that can sit quietly on a desk.
But there’s a crucial nuance: idle power is famously context-sensitive. Two DGX Spark-class systems can show different results depending on: connected displays, refresh rates, OS power policy, NIC link state, attached peripherals, background services, and OEM firmware defaults.
Don’t treat one number as a guarantee
The best way to interpret the “32%+” headline is as an example of what is possible under certain conditions—not a universal promise that every Spark-based system will land on the same idle wattage.
One report even notes that an OEM GB10-based variant did not show the same improvement in the same way, underscoring that platform variations can matter. The fix is real; the exact “before and after” depends on your system’s state at idle.
Why a 200 Gbps NIC can dominate idle power
To understand why this update is so impactful, it helps to translate “200 Gbps ConnectX NIC” into practical terms. This is not a typical 1GbE or 10GbE controller that sips power in the background. ConnectX-class NICs are engineered for high-performance fabrics, with features that can include advanced offloads, low-latency datapaths, and (in many deployments) RDMA capabilities.
That design focus brings performance benefits, but it can also make the NIC’s baseline power draw more noticeable—especially when you’re measuring the entire machine at the wall while the rest of the system is in a low-activity state.
The old “desktop networking” intuition—where the NIC’s idle draw is negligible compared to everything else—doesn’t always hold when the networking stack is datacenter-grade. If the NIC remains in a relatively active state even when there’s no meaningful network use, it becomes a persistent idle load.
Hot-plug support is one of the cleanest ways to address this. If the platform can accurately treat an unused or unlinked high-speed adapter as something that can enter a deeper sleep state, the system recovers real watts without removing the feature.
A simple mental model
Before: “Keep the high-speed NIC awake just in case.”
After:
“If it isn’t being used, let it sleep—wake it when needed.”
Who benefits most from this fix
Not every DGX Spark owner uses the same workflow. Some treat it like a personal AI workstation. Others treat it like a tiny lab node. The idle power reduction is meaningful for both—but it is especially valuable for a few specific use cases.
1) Always-on local inference and agent workloads
If your Spark acts as a local “AI appliance”—serving models, running background jobs, hosting RAG indexes, or powering agents—then it’s likely online 24/7. In that world, idle is not a transitional state; it is the default state between bursts of work. Lower idle draw directly improves the economics and the comfort of keeping the machine on.
2) Headless mini-server setups
Many owners run DGX Spark headless and connect via SSH, remote desktop, or a web UI. Headless operation often changes the system’s display pipeline behavior and can reduce baseline power draw further. If your workflow is headless most of the time, the update’s impact can stack with the headless configuration’s typical savings.
3) Multi-node “stacking” and bursty lab experiments
High-speed networking is a feature for multi-node workflows—yet multi-node jobs tend to be bursty. You might run a distributed workload, then let the nodes sit idle overnight. If the NIC can drop into a lower-power state between jobs, the idle windows become far cheaper.
Always-on AI: why idle watts matter more than you think
For years, PC power discussions focused on peak performance: “What’s the wattage under load?” But the AI workstation era changes the profile. Many AI systems are used like small servers: they wait, they respond, they run periodic tasks, they host services, and they handle spikes.
In that pattern, idle behavior becomes the main driver of total energy use. A reduction of 12–18W can look modest on paper, but it compounds across:
- Hours: a workstation left on all day
- Days: a model host that stays online through weekends
- Fleets: multiple nodes in a small lab or classroom
It’s also about comfort. Lower idle power typically means lower idle heat, which can mean lower fan activity and reduced acoustics in a work area. The experience difference between “always slightly warm and audible” and “mostly silent” is often driven by baseline watts, not by peak watts.
The practical win
This update doesn’t make DGX Spark faster. It makes DGX Spark easier to live with—especially if you treat it like a service node rather than a “turn it on only when needed” desktop.
How to get the savings (update + validate)
If you want to confirm whether your DGX Spark is actually benefiting from the ConnectX power-management change, approach it like a quick lab test. The goal is not to chase a specific wattage number; it’s to verify that the NIC can enter a lower-power behavior when it’s not in use.
Step 1: Install the latest DGX Spark software update
NVIDIA announced the February 2026 DGX Spark software updates as a rolling release, and the feature is described directly in the release notes. Update using the official DGX Spark update flow documented by NVIDIA.
Step 2: Reboot and let the system settle
Power-management changes often require a reboot to fully apply at the firmware/driver boundary. After reboot, let the system settle for a few minutes so background processes calm down.
Step 3: Measure idle at the wall
Use a wall power meter or smart plug that shows real-time watts. Measure “idle” consistently: close heavy apps, stop active downloads, and ensure no model inference is running.
Step 4: Test the two biggest variables
Variable A: display attached vs headless
Compare idle watts with a display connected vs disconnected. A headless state can reduce display pipeline activity, which may reduce idle watts further.
Variable B: NIC “in use” vs “not in use”
If your ConnectX port is linked to active hardware, you may see different idle behavior than if it is not linked. The update’s “save power when not in use” implies that link/usage state matters.
Step 5: Record before/after for your environment
The most useful outcome is your own baseline: “My Spark idles at X watts in my normal setup; after update it idles at Y watts.” If you manage multiple nodes, do this once and apply it as a sanity check across the rest of your fleet.
Caveats: OEM variance, link state, and peripherals
If you update and don’t see a dramatic drop, that doesn’t automatically mean the update failed. Idle behavior can be held up by a few common factors:
OEM firmware defaults can differ
DGX Spark-class systems based on the same underlying platform can ship with different defaults—BIOS/UEFI settings, firmware revisions, and power policy choices. That can affect whether a power feature expresses itself the same way across vendors.
Link state can keep high-speed hardware awake
A NIC that is actively linked (especially at high speed) can remain in a more awake state than one that is physically disconnected or unused. If your Spark is connected into a network configuration that keeps the adapter “busy enough,” you may not see the full “not in use” savings.
Displays and refresh rates can influence idle
The headless vs display-attached difference is a reminder: the GPU and display subsystem can alter idle draw. A high refresh rate display, multiple monitors, or certain resolutions can keep parts of the pipeline active.
Background services matter
Indexing, updates, container pulls, telemetry, or any continuous workload can prevent a “true idle” state. If your baseline is noisy, consider measuring idle after a clean reboot and a short wait period.
What not to do
Don’t chase a single published idle watt number. Validate your baseline and confirm that the system’s “NIC unused” state is cheaper than before.
The bigger picture: efficiency is now software
The most interesting aspect of this update is what it signals about modern AI systems: efficiency improvements are no longer only about silicon. They’re about coordination across firmware, drivers, and OS policy.
DGX Spark didn’t become a different machine overnight. The hardware didn’t change. What changed is the platform’s ability to place a high-performance component—the ConnectX-7 adapter—into a more appropriate low-power state when it isn’t doing useful work.
This is the direction the entire industry is moving:
- More datacenter features in smaller boxes (high-speed networking, advanced accelerators, specialized interconnects)
- More mixed usage patterns (bursty inference, background agents, occasional distributed training)
- More demand for always-on behavior without always-on costs
In that world, the best products are not just the ones with the highest peak performance. They are the ones that behave intelligently at idle. This update moves DGX Spark in that direction.
What you should take away
Peak performance sells headlines. Idle performance determines whether you enjoy owning the machine.
FAQ
Is the 32%+ idle power reduction guaranteed for every DGX Spark?
No. The 32%+ figure comes from reported measurements under specific conditions. NVIDIA’s own release notes state “up to 18W” savings when the ConnectX-7 adapter is not in use. Your exact result depends on display attachment, NIC link state, system services, and OEM firmware defaults.
What does “hot-plug support” mean here?
In practical terms, it enables more robust detection and handling of the NIC’s presence/usage state so the platform can transition the adapter into deeper low-power behavior when it isn’t being used.
Why does headless operation sometimes reduce idle power further?
With no display connected, the system may not need to keep parts of the display pipeline active at the same level. That can change baseline behavior. The extra headless savings is not necessarily the NIC—it can be display-related power state differences.
Will this affect performance when I actually use the NIC?
The intent is to reduce power when the adapter is not in use while preserving full capability when it is. When your networking workload starts, the NIC should wake into its active state as needed.
What else is included in the update?
NVIDIA’s release notes and forum announcement also mention Bluetooth audio support and additional security controls (including the ability to disable Wi-Fi/Bluetooth in UEFI), plus improvements to monitor/TV compatibility. The power management improvement is the headline change for this story.
Sources
- NVIDIA DGX Spark Release Notes (Power Management: hot-plug support for ConnectX-7, up to 18W savings): docs.nvidia.com
- NVIDIA Developer Forums announcement: “DGX Spark Software Updates 02/2026” (mentions measurable power efficiency gains and other changes): forums.developer.nvidia.com
- Reported measurements and summary coverage of the idle power reduction: tomshardware.com