What happened
On October 19, 2026, Arduino moved the VENTUNO Q from its 299 dollar introductory price to an official retail price of $399. The stated reason was not a change in the silicon. It was memory. The company wrote that the cost of key components, “in particular memory modules,” had risen sharply across the industry and that it had absorbed those increases as long as it could. The board sold out twice at the lower price.
The VENTUNO Q is the first product from Qualcomm’s October 2025 acquisition of Arduino. It pairs a Qualcomm Dragonwing IQ8 processor with a Hexagon NPU rated at 40 dense TOPS, an STM32H5 microcontroller running Zephyr for sub-millisecond motor control, 16 GB of LPDDR5, 64 GB of eMMC expandable through an M.2 NVMe slot, and pre-loaded models including a 4B language model. It is a real step up in integration for a robotics board at this price. It is also a clean example of a spec sheet where the headline number and the number that determines your experience are two different things.
Two numbers on the spec sheet, and neither one decides token speed
40 TOPS is a compute rating. It describes how many operations the NPU can perform per second on a workload that is compute-bound, which is what a vision model or a dense convolution pipeline is. Token generation on a language model is not that kind of workload. A single stream reading a model one token at a time is limited by how fast weights can be moved to the compute units.
Qualcomm’s documentation for the IQ-8275 gives the memory configuration: six independent 16-bit channels, LPDDR5 or LPDDR5X, up to 3200 MHz. That is 96 bits of total width. Qualcomm does not publish a bandwidth figure for this part, so the arithmetic is mine. At LPDDR5’s double data rate, a 16-bit channel at 3200 MHz carries 12.8 GB/s. Six channels carry 76.8 GB/s. That is the theoretical ceiling for anything read from memory per token.
For comparison, NVIDIA publishes 102.4 GB/s for the Jetson Orin Nano Super, which is a 128-bit LPDDR5 bus at 6.4 Gbps effective per pin. The Orin has 1.33 times the width and about 1.3 times the theoretical bandwidth. It rates 67 TOPS against the VENTUNO Q’s 40. Neither gap explains what actually happens on this board, because the measured result is nowhere near either ceiling.
The measured result: 0.774 tokens per second
The most useful thing published about this board is not a vendor benchmark. It is a project-hub write-up from October 14, 2026, in which someone ran a mixture-of-experts model on the board through Colibri, an engine that loads experts from storage and keeps a cache in RAM.
Three results from that run matter.
- A 32-slot expert cache produced about 0.522 tokens per second.
- A 96-slot cache produced about 0.740 tokens per second, roughly 42 percent faster.
- Switching the CPU frequency governors to performance added about 5 percent more, to 0.774 tokens per second.
The cache size moved the number by 42 percent. CPU tuning moved it by 5 percent. The NPU was not used at all, because the Colibri revision tested has no Qualcomm QNN backend. On the earlier core-pinned tests, the CPU was about 1.74 times faster than the Adreno GPU on that workload.
Here is what the numbers say when you do the arithmetic. The tested checkpoint is a 35B model with about 3B active parameters per token. At int4 that is roughly 1.5 GB of weights touched per token. On a 76.8 GB/s bus, if every active weight were resident in RAM, the ceiling would be about 51 tokens per second. The measured rate was 0.774, which is 1.5 percent of that ceiling. The effective data rate implied by the measurement is about 1.16 GB/s, which is below even the sequential read rate of the Gen4 NVMe the project used.
The model file is 23.1 GB on a board with 16 GB of RAM. Most of the work is not on the memory bus. It is on the storage tier and the cache policy. That is not a bandwidth story. It is the same shape as the case where a 48 GB model ran on 32 GB of RAM: the memory wall is a cache policy problem before it is a capacity problem. A larger cache moved this board from 0.522 to 0.774 tokens per second without changing a single piece of hardware.
The same class of engine on a machine that fits its cache
I measured the box that serves this model on the same day. It is a Ryzen 9 5900X, 64 GB of DDR4 and an RTX 3090, with the model served by llama.cpp through llama-swap. The model is a 125B mixture-of-experts with 24,576 experts, resident in about 12 GB of VRAM with the rest split across RAM, CPU and storage. It does not fit in memory either. The request was sent from a different machine, a Ryzen 7 2700X, so the wall-clock numbers include the network round trip.
A 1,200-token generation completed in 12.3 seconds wall-clock. The engine reported 11.0 seconds of generation time, which means 1.3 seconds of fixed request and tokenization overhead, about 10.5 percent of the total. Engine-reported decode was 109.1 tokens per second. Wall-clock decode through the API was 97.6. Speculative decoding accepted 559 of 720 proposed tokens, 77.6 percent.
The memory channel on the serving box measured 16.95 GB/s of copy bandwidth on a 128 MB buffer, which is 33.9 GB/s of memory traffic because a copy touches memory twice. That is 33 percent of the theoretical 51.2 GB/s for dual-channel DDR4 at 3200 MHz counted as data moved, or 66 percent counted as memory traffic.
Now put the two measurements next to each other. If a dense 125B model at int4 had to be read from that channel, 62.5 GB per token at 16.95 GB/s gives a ceiling of 0.27 tokens per second. The measured rate is 360 times that. It is only possible because the weights that matter for this workload are not on the DDR4 channel. They are on the GPU tier, and the cache policy decides how much of the model ever has to leave it.
The difference between 0.774 tokens per second and 97.6 is not TOPS, and it is not core count. It is how much of the model sits on a tier fast enough to feed the compute units, and how much has to be fetched from a tier that cannot.
The tiers, in one table
| Tier | Bus | Bandwidth | What it can serve |
|---|---|---|---|
| VENTUNO Q, IQ-8275 | 6 x 16-bit LPDDR5, 96-bit total | 76.8 GB/s theoretical, 1.16 GB/s measured effective | 16 GB. A 4B int4 model fits. A 23 GB MoE does not. |
| Jetson Orin Nano Super | 128-bit LPDDR5, 6.4 Gbps per pin | 102.4 GB/s published | 8 GB. Small dense models, vision pipelines. |
| Desktop dual-channel DDR4-3200 | 2 x 64-bit | 51.2 GB/s theoretical, 16.95 GB/s measured copy | 64 GB. Large models, but at a rate the channel caps. |
| HBM3E stack | 1024-bit, 16 channels | 1.18 TB/s per stack (SK hynix, 8-Hi); 1.22 TB/s for the 12-Hi | 24 GB per 8-Hi stack, 36 GB per 12-Hi. The tier that actually feeds a data center accelerator. |
A 256-bit LPDDR5X bus at 273 GB/s is the number that sets the ceiling on local decode for a board-class part. The VENTUNO Q is just over a third of that width (96 bits against 256). But the measured 1.16 GB/s on this board shows that bus width is not the binding constraint when the model is larger than RAM. The gap between a board that fits a model in memory and a board that streams it from storage is larger than the gap between any two TOPS ratings on the same shelf.
What to buy for this workload
If the workload is a small dense model that fits in memory, the bus decides everything, and the VENTUNO Q’s 16 GB is not an advantage over the Orin Nano Super’s 8 GB. Extra RAM buys model size. It does not buy tokens per second. The hardware ladder from budget to enterprise is built on that distinction, and a robotics board is a different decision from a local inference box.
For the Orin Nano Super, the developer kit is the part to get. Current Amazon price is about $399 against a $249 list price, and it has traded at $245 earlier this year. The VENTUNO Q itself is not on Amazon; Arduino sells it through its own store and distributor network, so there is no listing to point at.
NVIDIA Jetson Orin Nano Super Developer Kit (B0BZJTQ5YP), about $399 on Amazon as of October 2, 2026, against a $249 list price.
If the workload is a model larger than RAM, storage becomes part of the inference path, not just a place to keep files. The tested project used a 1 TB Gen4 2230 NVMe. That is the correct form factor for this board, and the drive’s random-read behaviour matters more than its sequential rating once the engine is pulling individual experts per token.
SABRENT Rocket 2230 1TB NVMe, PCIe Gen4 x4, DRAM-less HMB design (B0BQG6JCRP), about $119 as of October 2026.
And the reason the board’s price moved is visible in the retail memory market. A TEAMGROUP 16 GB DDR4 kit that sold for $27 in February 2025 is listed at about $130 now. The mechanism is the one already covered in what a 3x DRAM price means for local LLM builds: wafer capacity moved to HBM, and the modules on a retail shelf are what is left.
TEAMGROUP T-Force Vulcan Z DDR4 16 GB kit, 2 x 8 GB, 3200 MHz CL16 (B07T637L7T), about $130 as of October 14, 2026, against a $27 low in February 2025.
What the price change actually tells you
A 40 TOPS rating is not wrong. It describes a different workload than the one a language model runs. For single-stream generation on a board with 16 GB of RAM, the numbers that matter are the width of the memory bus, whether the model fits on it, and how much of the model the engine can keep resident. The VENTUNO Q is a well-built platform for robotics, and its dual-brain architecture solves a real integration problem. As an inference machine it is a 96-bit bus with a cache policy on top, and the published measurement shows what that produces when the model is bigger than the memory.
The price change is the honest part of the story. Arduino said plainly that memory module costs moved the price. Anyone buying hardware for local inference in 2026 is buying into the same allocation problem, and the spec sheet does not mention it.
EOF


Pingback: A 16 TOPS Board on a 34 GB/s Bus: What the OrangePi O1 Price Says About Arithmetic Intensity - GenX AI Tools