Nvidia says it's more than doubled the DGX Spark’s performance since launch
(2026/01/06)
- Reference: 1767654012
- News link: https://www.theregister.co.uk/2026/01/05/nvidia_dgx_spark_speed/
- Source link:
Nvidia's DGX Spark and its GB10-based siblings are getting a major performance bump with the platform's latest software update, announced at CES on Monday. The AI mini PC is also getting access to the GPU giant's full suite of AI Enterprise apps, alongside integrations with RTX Remix and Hugging Face's Reachy robotics platform.
First teased at CES 2025 under the codename [1]Project Digits , the DGX Spark is a tiny AI workstation designed to support rapid prototyping, GPU-accelerated software development, and local AI training and inference workloads.
While it's [2]billed as the "world's smallest AI supercomputer," the machine isn't actually that powerful with the computational grunt equivalent to an RTX 5070. What sets it apart from the rest of Nvidia's lineup is the inclusion of 128 GB of unified memory, all of which can be allocated to the GPU. That's the most of an Nvidia workstation product, save for the DGX Station.
[3]
The DGX Spark's golden design is clearly inspired by the original DGX-1 system hand-delivered by Jensen Huang to Elon Musk at OpenAI in 2016 - Click to enlarge
Since the Spark's launch in October, Nvidia has been hard at work improving the system's performance by an average of 2.5x across a number of software libraries and frameworks, though we haven't had the opportunity to independently verify those claims just yet.
Check out our day one review of Nvidia's tiniest AI supercomputer [4]here
But before you get too excited, don't expect to see the Spark churning out tokens twice as quickly as before. The decode phase of LLM inference, during which tokens are generated, is bandwidth-limited. The Spark can't actually get much faster here.
Instead, Nvidia has confirmed that most of the performance gains in this software release are for the compute-intensive parts of the genAI pipeline. For LLM inference, these updates will predominantly improve prefill performance, which reduces the time from when a prompt is submitted to when the Spark begins generating a response.
[5]
Updates include enhancements to Nvidia's inference engine, TensorRT LLM, Llama.cpp, and PyTorch to name a few. The latter, we'll note, should help to improve other compute-intensive workloads like fine tuning and image or video generation.
AI Enterprise is coming to Spark
Nvidia also announced plans to make its full [6]AI Enterprise suite available on the Spark as a subscription service later this month.
The suite includes access to a host of enterprise-focused applications, frameworks, models, and microservices designed to streamline the development of AI apps and services.
[7]
[8]
Normally the suite runs for $4,500 a year per GPU or $1 an hour per GPU in the cloud, but we're told Nvidia does plan to offer special pricing for Spark, although that's not official yet. Nvidia does make the offering available at no cost to developers, but a paid plan is required to use applications and services built on it in production.
We've reached out to Nvidia for comment; we'll let you know if we hear anything back.
No official support for third-party Linux distros, yet
In any case, the offering should help to assuage concerns over long-term software support. When the Spark launched, many expressed fears that the $3,999 AI lab in a box could be rendered a golden paperweight in a few years if Nvidia failed to release updates to DGX OS, the company's custom spin of Ubuntu.
Some older development boards, like the Jetson Nano, have faced such a fate. The single-board computer never advanced past Ubuntu 18.04, which hasn't been supported since 24.04 came out a couple of years ago.
[9]
Nvidia tells us this shouldn't be the case for the Spark or other GB10-based systems.
"We're committed to support. In fact, we just released the latest kernel with some security patches," Allen Bourgoyne, director of product marketing for Nvidia's professional visualization business, said in response to our questions.
The real test will be whether we see Nvidia base DGX OS for the Spark on Ubuntu 26.04, which will be Canonical's next long-term support release.
[10]
Support for third-party Linux distros, like Red Hat Enterprise Linux (RHEL), would go a long way to mitigating concerns of obsolescence, but it doesn't appear to be on the table just yet.
Nvidia says it's focusing development on DGX OS for now. But the company could still release the drivers and firmware packages necessary to get GPU acceleration working on the GB10 on other distros.
[11]Nvidia shrinks Grace-Blackwell Superchip to power $3K mini PC
[12]AMD Strix Halo vs Nvidia DGX Spark: Which AI workstation comes out on top?
[13]DGX Spark, Nvidia's tiniest 'supercomputer' , tackles large models at solid speeds
[14]Nvidia details its itty bitty GB10 superchip for local AI development
A CUDA coding assistant, RTX Remix, and a robotic collab
Alongside Nvidia's software updates and subscription services, the company plans to release a version of its Nsight CUDA code assistant capable of running entirely on the Spark. Previously, the models used by the assistant were too large to fit on Nvidia's consumer graphics offerings and therefore were limited to the cloud, making it less useful for privacy-conscious enterprises.
Nsight is expected to arrive on the Spark later this spring.
For gamers, Nvidia is extending RTX Remix support to the Spark. The platform is designed to support the development of game mods that take advantage of Nvidia's ray tracing accelerators. With the integration, activities such as text generation can be offloaded to the Spark.
Meanwhile, for robotics enthusiasts, Nvidia says it's working on a new guide that pairs the Spark with Hugging Face's Reachy robot. The desktop robot is designed to support the development of embodied AI frameworks and services.
Bigger Spark Clusters are on the way
Finally, we've learned that Nvidia could soon extend support for Spark clusters containing more than two systems.
One of the Spark's more interesting features is the inclusion of a ConnectX-7 NIC with a pair of QSFP+ ports capable of delivering 200 Gbps of bandwidth between them.
Nvidia currently supports linking up to two Sparks (or GB10 partner systems) using these ports, but in theory there's nothing stopping someone from building a whole cluster of them. Nvidia tells us that it's seen interest from customers for larger clusters and its engineers are actively exploring the possibility. ®
Get our [15]Tech Resources
[1] https://www.theregister.com/2025/01/07/nvidia_project_digits_mini_pc/
[2] https://www.theregister.com/2025/10/14/dgx_spark_review/
[3] https://regmedia.co.uk/2025/10/14/dgx_spark_front_view.jpg
[4] https://www.theregister.com/2025/10/14/dgx_spark_review/
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[6] https://www.theregister.com/2024/08/06/nvidia_software_empire/
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2025/01/07/nvidia_project_digits_mini_pc/
[12] https://www.theregister.com/2025/12/25/amd_strix_halo_nvidia_spark/
[13] https://www.theregister.com/2025/10/14/dgx_spark_review/
[14] https://www.theregister.com/2025/08/27/nvidia_blackwell_gb10/
[15] https://whitepapers.theregister.com/
First teased at CES 2025 under the codename [1]Project Digits , the DGX Spark is a tiny AI workstation designed to support rapid prototyping, GPU-accelerated software development, and local AI training and inference workloads.
While it's [2]billed as the "world's smallest AI supercomputer," the machine isn't actually that powerful with the computational grunt equivalent to an RTX 5070. What sets it apart from the rest of Nvidia's lineup is the inclusion of 128 GB of unified memory, all of which can be allocated to the GPU. That's the most of an Nvidia workstation product, save for the DGX Station.
[3]
The DGX Spark's golden design is clearly inspired by the original DGX-1 system hand-delivered by Jensen Huang to Elon Musk at OpenAI in 2016 - Click to enlarge
Since the Spark's launch in October, Nvidia has been hard at work improving the system's performance by an average of 2.5x across a number of software libraries and frameworks, though we haven't had the opportunity to independently verify those claims just yet.
Check out our day one review of Nvidia's tiniest AI supercomputer [4]here
But before you get too excited, don't expect to see the Spark churning out tokens twice as quickly as before. The decode phase of LLM inference, during which tokens are generated, is bandwidth-limited. The Spark can't actually get much faster here.
Instead, Nvidia has confirmed that most of the performance gains in this software release are for the compute-intensive parts of the genAI pipeline. For LLM inference, these updates will predominantly improve prefill performance, which reduces the time from when a prompt is submitted to when the Spark begins generating a response.
[5]
Updates include enhancements to Nvidia's inference engine, TensorRT LLM, Llama.cpp, and PyTorch to name a few. The latter, we'll note, should help to improve other compute-intensive workloads like fine tuning and image or video generation.
AI Enterprise is coming to Spark
Nvidia also announced plans to make its full [6]AI Enterprise suite available on the Spark as a subscription service later this month.
The suite includes access to a host of enterprise-focused applications, frameworks, models, and microservices designed to streamline the development of AI apps and services.
[7]
[8]
Normally the suite runs for $4,500 a year per GPU or $1 an hour per GPU in the cloud, but we're told Nvidia does plan to offer special pricing for Spark, although that's not official yet. Nvidia does make the offering available at no cost to developers, but a paid plan is required to use applications and services built on it in production.
We've reached out to Nvidia for comment; we'll let you know if we hear anything back.
No official support for third-party Linux distros, yet
In any case, the offering should help to assuage concerns over long-term software support. When the Spark launched, many expressed fears that the $3,999 AI lab in a box could be rendered a golden paperweight in a few years if Nvidia failed to release updates to DGX OS, the company's custom spin of Ubuntu.
Some older development boards, like the Jetson Nano, have faced such a fate. The single-board computer never advanced past Ubuntu 18.04, which hasn't been supported since 24.04 came out a couple of years ago.
[9]
Nvidia tells us this shouldn't be the case for the Spark or other GB10-based systems.
"We're committed to support. In fact, we just released the latest kernel with some security patches," Allen Bourgoyne, director of product marketing for Nvidia's professional visualization business, said in response to our questions.
The real test will be whether we see Nvidia base DGX OS for the Spark on Ubuntu 26.04, which will be Canonical's next long-term support release.
[10]
Support for third-party Linux distros, like Red Hat Enterprise Linux (RHEL), would go a long way to mitigating concerns of obsolescence, but it doesn't appear to be on the table just yet.
Nvidia says it's focusing development on DGX OS for now. But the company could still release the drivers and firmware packages necessary to get GPU acceleration working on the GB10 on other distros.
[11]Nvidia shrinks Grace-Blackwell Superchip to power $3K mini PC
[12]AMD Strix Halo vs Nvidia DGX Spark: Which AI workstation comes out on top?
[13]DGX Spark, Nvidia's tiniest 'supercomputer' , tackles large models at solid speeds
[14]Nvidia details its itty bitty GB10 superchip for local AI development
A CUDA coding assistant, RTX Remix, and a robotic collab
Alongside Nvidia's software updates and subscription services, the company plans to release a version of its Nsight CUDA code assistant capable of running entirely on the Spark. Previously, the models used by the assistant were too large to fit on Nvidia's consumer graphics offerings and therefore were limited to the cloud, making it less useful for privacy-conscious enterprises.
Nsight is expected to arrive on the Spark later this spring.
For gamers, Nvidia is extending RTX Remix support to the Spark. The platform is designed to support the development of game mods that take advantage of Nvidia's ray tracing accelerators. With the integration, activities such as text generation can be offloaded to the Spark.
Meanwhile, for robotics enthusiasts, Nvidia says it's working on a new guide that pairs the Spark with Hugging Face's Reachy robot. The desktop robot is designed to support the development of embodied AI frameworks and services.
Bigger Spark Clusters are on the way
Finally, we've learned that Nvidia could soon extend support for Spark clusters containing more than two systems.
One of the Spark's more interesting features is the inclusion of a ConnectX-7 NIC with a pair of QSFP+ ports capable of delivering 200 Gbps of bandwidth between them.
Nvidia currently supports linking up to two Sparks (or GB10 partner systems) using these ports, but in theory there's nothing stopping someone from building a whole cluster of them. Nvidia tells us that it's seen interest from customers for larger clusters and its engineers are actively exploring the possibility. ®
Get our [15]Tech Resources
[1] https://www.theregister.com/2025/01/07/nvidia_project_digits_mini_pc/
[2] https://www.theregister.com/2025/10/14/dgx_spark_review/
[3] https://regmedia.co.uk/2025/10/14/dgx_spark_front_view.jpg
[4] https://www.theregister.com/2025/10/14/dgx_spark_review/
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[6] https://www.theregister.com/2024/08/06/nvidia_software_empire/
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aVyW7f2A38S0UGJNH_nk2AAAA0g&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://www.theregister.com/2025/01/07/nvidia_project_digits_mini_pc/
[12] https://www.theregister.com/2025/12/25/amd_strix_halo_nvidia_spark/
[13] https://www.theregister.com/2025/10/14/dgx_spark_review/
[14] https://www.theregister.com/2025/08/27/nvidia_blackwell_gb10/
[15] https://whitepapers.theregister.com/