-
Notice regarding image hosts and attachments. See here.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Nvidia Spark
- Thread starter Villan
- Start date
Soldato
- Joined
- 28 Sep 2014
- Posts
- 4,018
- Location
- Scotland
Nvidia launched Shield TV Pro back in 2019 with updated X1+ CPU with same Maxwell GPU, hopefully they will launch new Shield TV Spark RTX soon.
Soldato
- Joined
- 28 Sep 2014
- Posts
- 4,018
- Location
- Scotland
Yes everything is low end Amlogic that are over 10 years old due to $50-$100 prices. I spend all day struggled to find high end Android/Google TV box with high end GPU but none of it existed on internet and asked Google AI about it.
Soldato
- Joined
- 28 Sep 2014
- Posts
- 4,018
- Location
- Scotland
HONOR to launch 2L AI workstation with NVIDIA RTX Spark and 128GB memory - VideoCardz.com
HONOR's Tiangong AXB35 Ultra combines NVIDIA RTX Spark with 128GB of memory and a 2TB SSD. The entire workstation fits inside a roughly 2-liter chassis.
videocardz.com
Thought Nvidia confirmed RTX Spark has 9 OEM partners Acer, ASUS, Microsoft, MSI, Dell, Gigabyte, HP, Lenovo and MSI.
Now it gain another new OEM partner Honor so it seems there could be more OEM partners to support RTX Spark soon.
Nice!
It's turning into gold, prices keep going up https://videocardz.com/newz/dgx-spark-prices-are-already-climbing-before-rtx-spark-mini-pcs-launch
Last edited:
Precisely, those prices go up and down for dgx spark all the time, it's not the first time they went up with some sellers. People try to scalp on them all the time but there seem to be enough for everyone, datacenters do not really use them and for single person using LLM at home they are too slow and too expensive to justify purchase.
And yes, Mac seem to be a better choice in at least some cases, as at least one has whole Apple ecosystem with it, whereas this is Linux (custom to that), which most people still prefer not to touch.
And yes, Mac seem to be a better choice in at least some cases, as at least one has whole Apple ecosystem with it, whereas this is Linux (custom to that), which most people still prefer not to touch.
Last edited:
As they should, there's really not any super advanced tech in them to be that expensive, the only excuse that have is "memory expensive". That aside it's just a small pc with more tensors than average soc but nothing really groundbreaking that would dictate such price normally. But this market isn't really normal at all...
Microsoft releases new Nvidia-chip AI PCs with revamped Windows 11
Microsoft releases new Nvidia-chip AI PCs with revamped Windows 11 | TechCrunch
Microsoft revealed the specs and price for its Surface Laptop Ultra, AI PCs that run on Nvidia chips that are designed to run AI models and agents.
techcrunch.com
It runs decently fast due to improvements in inference engines. TensorFlow now > vLLM by quite a bit for the models it supports. Also with Sparks you get great concurrency. Because CUDA you get nice prefill. And they effectively sip power, making them a bit more cost efficient than Nvidia GPUs if you are running multi agent workloads constantly. I have GLM 5.3 Flash running across 2x DGX Sparks. Tempted to grab a 3rd to get a bit more speed and a bit more concurrency.
For context I am using approx 500m tokens a day across the two Sparks I own.
Last edited:
IMHO nobody "normal" buys Spark (or Apple or AMD one - both got recently very well optimised engines too with nice prefil and decode speedups) for concurrency and inference, it's not what those machines have been really designed for IMHO - training, fine-tuning, general "fun" with models, not serious inference as such. Sure, you can do it, but it's just not cost-effective at all for a serious work. Most people that work and earn on whatever LLM produce will just buy some fancy frontier sub and be done with it. No, IMHO Spark and the likes are just for absolute enthusiasts, as a hobby more than anything. Those are very niche machines as such, with crazy (these days) pricing. Personally, I'd love to put my hands on RTX Pro 6000 but, having experienced Qwen3.8 Flash on my 5090 running with very sensible up to 50T/s, I am fine whilst pricing is silly.
Plus, plenty new inference engines show up optimised for specific models (like said Qwen3.8 Flash) allowing to run them even on 12GB vRAM GPUs these days - and still faster than Spark can do it in some cases, for a fraction of the cost.IMHO nobody "normal" buys Spark (or Apple or AMD one - both got recently very well optimised engines too with nice prefil and decode speedups) for concurrency and inference, it's not what those machines have been really designed for IMHO - training, fine-tuning, general "fun" with models, not serious inference as such. Sure, you can do it, but it's just not cost-effective at all for a serious work. Most people that work and earn on whatever LLM produce will just buy some fancy frontier sub and be done with it. No, IMHO Spark and the likes are just for absolute enthusiasts, as a hobby more than anything. Those are very niche machines as such, with crazy (these days) pricing. Personally, I'd love to put my hands on RTX Pro 6000 but, having experienced Qwen3.8 Flash on my 5090 running with very sensible up to 50T/s, I am fine whilst pricing is silly.
While I agree with some of your points I also think some of these things are not true at all from my experience!
Im pulling approx 500m input tokens worth of requests per 24 hours running multiple concurrent requests (typically sitting at 3-5 but my current config supports 8x 272k context window agents all prefill/decoding at the same time! I also do qwen 3.8 on my rtx 5090 but there is a big difference between having single turns on a GPU and being able to service 100% of a very mature multi agent, multi scheduled job, kanban ticket setup in Hermes agent.
But yes I also use frontier, but solely as my orchestrator / main agent. Sparks pick up the rest of the work. And while I got these at near original retail price, while actual list price is seemingly increasing, the power efficiency where I don't see each spark break 50w each while doing this, you win back the cost in electricity particularly Vs a GPU. I know, I also get the most out of both my 5090.and 4090, pulling a combined 750w just from the GPUs in aggregate, and getting nowhere near the same level of work our of them since they can only service one request each, despite being faster.
I just jumped in and bought 2, I would be interested in any tips on how to get started.
Last edited:
First thing install your AI harness of choice. Whether that Claude or Codex etc. then give this repo to your agent to install exactly as it is https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold. I made 2 deviations from default path, 8 parallel instead of 4. 272k context window instead of 1m. Then whatever harness you want to use it in, point it to GLM 5.3 flash on your sparks. I presume you got a cable to connect them together? You just need one cable for the connectx 7 port.
I literally bought a 3rd spark earlier this week and it just arrived!
Thanks, and yes cable and the sparks hopefully arriving today.