RTX Spark: October launch, PAIR routing, a free RTX speed-up
NVIDIA confirmed at IFA in Berlin that RTX Spark PCs would ship in October, and released PAIR, a free tool that spreads AI work across the PCs on a home network. Nobody outside NVIDIA had benchmarked an RTX Spark, and I think the free speed-up for RTX cards mattered more on the day.
Plays here, from YouTube
Corrected on 13 September 2026. In the video of 6 September I suggested NVIDIA's PAIR router would combine the memory of several machines. It sends each request to one machine as a whole and does not split one model across them, so it will not. I said so in M5 Ultra Mac Studio wait times: 16 to 18 weeks for the full chip.
NVIDIA confirmed at IFA in Berlin that RTX Spark PCs would ship in October. They use the DGX Spark chip, moved into Windows laptops and small desktops. NVIDIA says the chip has 20 CPU cores, a Blackwell GPU, up to 1 petaflop of AI compute and up to 128GB of shared memory.
NVIDIA's post names Lenovo and Acer as makers of the new PCs. Lenovo had two laptops at that point, the Yoga Pro 9n and a 2-in-1, and there were no prices yet.
How fast it might be
As we know from the Sparks, memory bandwidth is what really decides how fast these machines will write text day to day, and petaflop numbers can be misleading on their own. As a reference point, NVIDIA's spec for the DGX Spark is 273GB a second of memory bandwidth, and on two of them I measured a 27 billion parameter model writing at 42.8 tokens a second.
Lenovo's spec sheet for the Yoga Pro 9n, as Tom's Hardware reported it, lists memory at up to 9,400 MT/s, against 8,533 MT/s in NVIDIA's spec for the DGX Spark. That is a touch faster than my desktop boxes, so I would expect similar speeds on a model that fits, for one person at a time.
PAIR
Alongside the hardware, NVIDIA released PAIR, which stands for Personal AI Router. It is a free and open source tool that spreads AI work across the PCs already on a home network. Digitimes and PCWorld reported it as Apache 2.0.
A few caveats
- By 6 September nobody outside NVIDIA had benchmarked an RTX Spark.
- NVIDIA's footnote for the DGX Spark says its 1 petaflop is a 4-bit figure that assumes sparsity, which is the most generous way of counting. I am assuming the RTX Spark figure is counted the same way.
- Tom's Hardware reported that Lenovo's machines run Windows on Arm, so I would check that the applications I rely on will run.
The free speed-up
The quieter announcement probably mattered more on the day. NVIDIA put new optimisations into llama.cpp and vLLM, and says local models run up to 1.9 times faster on RTX cards. The engine underneath LM Studio and Ollama, the two apps most people run models with, is llama.cpp. NVIDIA said the optimisations were available straight away, directly and through both apps, so anyone who already owns an RTX card gets the speed-up with an update.
What I would do
With an RTX card already in the machine, I would update first and shop later.
Read more
- NVIDIA's IFA announcement, with the October date, PAIR and the speed-upblogs.nvidia.com
- NVIDIA's post on what is in the RTX Spark chipblogs.nvidia.com
- Tom's Hardware on the specs of Lenovo's two RTX Spark laptopstomshardware.com
- LMSYS on the DGX Spark, with the Spark's memory bandwidthlmsys.org
- Digitimes on NVIDIA's IFA launch, where PAIR is given as Apache 2.0digitimes.com
- NVIDIA's DGX Spark page, with the footnote on how the petaflop is countednvidia.com
- LM Studio's documentationlmstudio.ai
- Ollama's page on GitHubgithub.com
The numbers in this story
| Figure | What it is | From | Kind |
|---|---|---|---|
| October 2026 | When NVIDIA said RTX Spark PCs would ship | blogs.nvidia.com | NVIDIA's statement |
| 20 cores, up to 1 petaflop, up to 128GB | The RTX Spark chip's CPU cores, AI compute and shared memory | blogs.nvidia.com | NVIDIA's claim |
| 1.9 times | The most NVIDIA says local models speed up on RTX cards | blogs.nvidia.com | NVIDIA's claim |
| 9,400 against 8,533 MT/s | Memory in Lenovo's Yoga Pro 9n, against the DGX Spark | tomshardware.com | the maker's claim |
| 273 GB/s | Memory bandwidth of the DGX Spark | lmsys.org | the maker's claim |
| 42.8 tok/s | A 27 billion parameter model on two DGX Sparks | my own test | measured by me |
My Sparks would run a little faster than these figures. I cap their graphics chips at 2,200MHz, against about 2,400 uncapped, to cut their power use. When I tested the cap on one box in August, writing speed stayed within 1 per cent, reading was 2 to 4 per cent slower, and the chip's peak power fell from 55.5W to 43.3W.
If a number on this page is wrong, I would rather know. Tell me in the comments under the video and I will correct it here and pin the correction there. What has been corrected so far is on the corrections page.