You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
$4,000 or $5,000 for an N1 or N1X computer may be too expensive for individual consumers, but for large corporate users, it could be justifiable and affordable because of the potential AI capabilities and productivity gains.
I know some major financial firms are deploying AI extensively. Two friends of mine who work in the financial industry told me that they don't know how they could perform their jobs with the same level of efficiency without AI.
Yes definitely agreed it's indispesable for some use cases, but the models doing the real work are far beyond the capability of even N1X -- they're trillion+ parameter models requiring terabytes of RAM to run. That's why I'm curious what this is going to do.
I'm a huge hardware nerd - and want to see this useful for something, but I'm honestly struggling to see where it's going to help. For 'strong home AI' - Apple likely offers better value/$ at this stage. (CUDA is not required to run models locally). AMD also has some really decent offerings with "Strix Halo". And then on the higher end - servers and GPUs are readily available.
I think 'toe in the water' is probably an accurate take on this for now..
The frontier models may still be in that range, but smaller models are becoming quite good. I was very impressed by the performance of GPT-OSS 20B, which is pretty dated at this point. I believe you can find rather good coding agents that will fit into 128GB of memory.
Not sure about Apple being better value. Their 128GB products appear to be in the $5000+ range.
I'm a huge hardware nerd - and want to see this useful for something, but I'm honestly struggling to see where it's going to help. For 'strong home AI' - Apple likely offers better value/$ at this stage. (CUDA is not required to run models locally). AMD also has some really decent offerings with "Strix Halo". And then on the higher end - servers and GPUs are readily available.
If I look at where all the previous NVIDIA SPARK boxes have gone, it’s primarily universities, startups (like OpenAI was once) and in-house AI developers in enterprises. What do all of those have in common - research and development leveraging the NVIDIA ecosystem, without having to live inside the limitations of neoclouds / cloud providers.
I do think this product services a different market than the high end Macs or x86/Strix, that don’t have access to all the NVIDIA libraries. But I do wonder how much new TAM Windows brings over SPARK w Linux. Or maybe I’m wrong about who will buy.
If I look at where all the previous NVIDIA SPARK boxes have gone, it’s primarily universities, startups (like OpenAI was once) and in-house AI developers in enterprises. What do all of those have in common - research and development leveraging the NVIDIA ecosystem, without having to live inside the limitations of neoclouds / cloud providers.
I do think this product services a different market than the high end Macs or x86/Strix, that don’t have access to all the NVIDIA libraries. But I do wonder how much new TAM Windows brings over SPARK w Linux. Or maybe I’m wrong about who will buy.
Because NVIDIA's N1 and N1X will support both Linux and Microsoft Windows, their addressable market is much larger than it would be with Linux only support.
If I look at where all the previous NVIDIA SPARK boxes have gone, it’s primarily universities, startups (like OpenAI was once) and in-house AI developers in enterprises. What do all of those have in common - research and development leveraging the NVIDIA ecosystem, without having to live inside the limitations of neoclouds / cloud providers.
I do think this product services a different market than the high end Macs or x86/Strix, that don’t have access to all the NVIDIA libraries. But I do wonder how much new TAM Windows brings over SPARK w Linux. Or maybe I’m wrong about who will buy.
Do you remember SUN SPARC workstations. They never took off exactly because of a too elitarian, academic spin around them.
An average high end user was never explained why he had to choose it over just a faster x86 box.
Most powerful users don't care of "advanced capabilities," and niche features, as surprisingly as it sounds. They care for it being fast, over it being "advanced"
There is a whole genre of Chinese hardware dedicated for desktopifying old server parts sold at rock bottom prices, that still have amazing cost-performance ratio.
The frontier models may still be in that range, but smaller models are becoming quite good. I was very impressed by the performance of GPT-OSS 20B, which is pretty dated at this point. I believe you can find rather good coding agents that will fit into 128GB of memory.
Not sure about Apple being better value. Their 128GB products appear to be in the $5000+ range.
That's fair - the Apple priicng has gone up a lot. But you get a whole portable computer for that and a fully working OS .. (Vs Windows on ARM). The CPU performance is also significantly higher on the Apple side, too.. I suspect Nvidia has a bit of an advantage on GPU though.
Because NVIDIA's N1 and N1X will support both Linux and Microsoft Windows, their addressable market is much larger than it would be with Linux only support.
I can see a lot of use cases for edge inference, but these seem most likely going into specific devices like cars, cameras, glasses, appliances.
I saw someone had mentioned there is academic use/preference. But for the general public, be it consumer or enterprise, I fail to see where the demand is for "local" general purpose inferencing.
Most powerful users don't care of "advanced capabilities," and niche features, as surprisingly as it sounds. They care for it being fast, over it being "advanced"
The new Nvidia RTX Spark N1 and N1X are obviously too expensive and unnecessary for most mainstream PC users. However, in terms of building a native developer network (rather than relying on x86 translation) and penetrating client environments, they represent a measured starting point. Starting without mass market volume can be a disadvantage, but it also allows Nvidia and MediaTek to cultivate their own market within a smaller and more controllable audience, such as edge AI and client AI developers.
Gaining developers' support is the first step toward building Nvidia's long term client hardware ecosystem. Developers come first then mass market adoption follows.
I've been thinking, Nvidia is on a similar route to Apple taking more control in the design of the processors and having more "arm" over their systems. Apple has their own developer system, parted with x86 and is on its way to replace communication silicon, and has an OS to gel it all together, albeit in a different type of end product for different people.
Is Nvidia's own OS even worth considering? Something that integrates the most out of their distinguished CUDA library and has more control over their hardware than ever before.
