On Monday at CES 2025, Nvidia unveiled a desktop computer called Project DIGITS. The machine uses Nvidia’s latest “Blackwell” AI chip and will cost $3,000. It contains a new central processor, or CPU, which Nvidia and MediaTek worked to create.
Responding to an analyst’s question during an investor presentation, Huang said Nvidia tapped MediaTek to co-design an energy-efficient CPU that could be sold more widely.
“Now they could provide that to us, and they could keep that for themselves and serve the market. And so it was a great win-win,” Huang said.
Previously, Reuters reported that Nvidia was working on a CPU for personal computers to challenge the consumer and business computer market dominance of Intel, Advanced Micro Devices and Qualcomm.
↫ Stephen Nellis at Reuters
I’ve long wondered why NVIDIA wasn’t entering the general purpose processor market in a more substantial way than it did a few years ago with the Tegra, especially now that ARM has cemented itself as an architecture choice for more than just mobile devices. Much like Intel, AMD, and now Qualcomm, NVIDIA could easily deliver the whole package to laptop, tablet, and desktop makers: processor, chipset, GPU, of course glued together with special NVIDIA magic the other companies opting to use NVIDIA GPUs won’t get.
There’s a lot of money to be made there, and it’s the move that could help NVIDIA survive the inevitable crash of the “AI” wave it’s currently riding, which has pushed the company to become one of the most valuable companies in the world. I’m also sure OEMs would love nothing more than to have more than just Qualcomm to choose from for ARM laptops and desktops, if only to aid in bringing costs down through competition, and to potentially offer ARM devices with the same kind of powerful GPUs currently mostly reserved for x86 machines.
I’m personally always for more competition, but this time with the asterisk that NVIDIA really doesn’t need to get any bigger than it already is. The company has a long history of screwing over consumers, and I doubt that would change if they also conquered a chunky slice of the general purpose processor market.
All the crapiness of non-standardized bootloaders of the smartphone space, now on your desktop and laptop, experience… ARM PC™
Server and desktop tier ARM hardware uses UEFI if it wants to comply with ARM’s own specifications.
Doesn’t seem to apply to laptops though. Neither Snapdragon nor Apple Silicon laptops “just works” with a generic Linux ARM image.
Not sure they are targeting much “desktop” at US$3000 for piece, other than the space it takes up on some Devs desktop!
I get it if you are developing with a rack full Jetson, it might be of interest. But not ever as the alt consumer desktop that some are trying to spin.
NVIDIA’s CEO made a comment about the 5090’s $2000 price tag being affordable because computer desktops costed $10k anyway. Unfortunately I didn’t have the mind to take down the link at the time, but his swing was a hard miss and it shows just how far disconnected he is from consumers. Not that it matters much since normal consumers aren’t where the big money is at.
>”it’s the move that could help NVIDIA survive the inevitable crash of the “AI” wave it’s currently riding”
No company is going to make remotely enough money selling desktop cpu chips in a rapidly declining desktop market to offset the tsunami of losses that are going to overwhelm the tech industry when the AI bubble pops.
Thom Holwerda,
It’s clear that there are applications of AI that will fail. And there will always be AI gismos that we can all point to and laugh at together. But at the same time I worry that way too many people and employees are underestimating AI’s potential to render human employees redundant in the coming years. They are not seeing AI the way employers do. Employers are looking at the exorbitantly high costs of labor, including all the sick time, family leave, health insurance, social security taxes, and so on, and these employers are very eager to cut costs to increase their profits, which is all shareholders care about. AI has room to improve, but once jobs go to AI, humans looking for jobs are going to find it extremely difficult to compete against the AI. I don’t think employees are ready to accept the harsh reality that employers don’t really care about them over AI to do the job cheaper.
The real problem with AI as implemented today (LLMs) is that it makes mistakes in a way that cannot be diagnosed and debugged (in the way rules-engines can) and cannot be held accountable like humans can:
https://yro.slashdot.org/story/23/05/30/2049253/lawyer-cited-6-fake-cases-made-up-by-chatgpt-judge-calls-it-unprecedented
Better keep those humans who can be held accountable employed if you want to avoid fines and other consequences.
kurkosdr
Well, I don’t think that using NN as a source of legal data demonstrates the technology at it’s best. Storing factual information in an LLM is neither efficient nor is it LLM’s strong point, Where LLMs can do well is interact with humans and processing data. LLMs trained on internet sources like chatgpt were never going to be optimal for the applications we’re talking about. However LLMs developed to be an interface for a legal database (or any other database) could prove to be extremely useful and it can provide verifiable citations.
I also expect LLMs to become much more specialized, like an assembly line of AI specialists that work together, each being trained to optimize some task. These AIs will prove valuable because they can analyze millions/billions of case permutations to measure the statistical outcomes even optimizing down to the judge and possibly even jury members. It’s almost scary what can be optimized for when you have enough computing power, but regardless both defendants and prosecutors will find that AI can not only be cheaper than human labor, but may ultimately be more effective too because it’s too much data for humans to process. I’m not saying we’re there today, but I do think law firms that don’t employ AI technology will become increasingly disadvantaged over time.
I’m afraid you are exactly right about this. Just look at what happened to IS/IT/CS jobs long before AI was a thing: It all got outsourced to workers in countries where the companies could get away with lower salaries, no healthcare, no taxes, etc. On the one hand I’m glad there are jobs at all for people in those countries, but it’s not out of any altruism on the part of the corporate world, it’s pure greed driving such decisions. And of course, now AI is poised to take those jobs from the lower paid overseas workers, so they will suffer the effects as well.
I know my job is secure because AI can’t do the physical and social aspects of being a sysadmin for a small business, in fact I could probably leverage AI for the more mundane parts of my job that are too much for a script but not important enough for another human hire. But give it time and even hands-on jobs might be at stake.
Morgan,
Yeah, very similar dynamics are at play with AI. Employees are a means to an end for corporations. Once they find a way to replace employees with something cheaper, they will. And even though some people will make the argument that offshoring yielded worse results, which is a view I sympathize with, it didn’t matter. The quality argument didn’t put a stop the widescale offshoring. And now those jobs that have been offshored are extremely difficult to bring back once they are lost. If we do nothing but deny AI’s potential to displace employees, by the time we realize it’s happening it will be too late to do anything about, we’ll be past the tipping point.
I haven’t jobs in my field go to AI yet, knock on wood…however quite a lot, perhaps even a majority, got offshored.
So, it’s as overpriced as nvidia, and as crappy as mediatek. Great combo
Fairly sure I made a comment recently that they only needed to add a raspberry pi. They listened fast :p
Windows on ARM is so awful.
I’d be all for nvidia joining the fray in the CPU space for consumer parts, but they are woefully behind there. They threw in the towel in mobile because they didn’t have the chops to build a full platform, which is what the industry wanted from SoC suppliers. Much of the benefit folks are seeing in laptops (really long battery life) again, have nothing to do with ARM ISA, and everything to do with the whole package – which includes even things like better cameras, and sleep/wake sub-systems.
It makes sense for nvidia to tap a partner like MediaTek, because they simply don’t have the history behind them to make a play here, without a LOT of investment and time they don’t have. But, they don’t seem to be targeting the consumer or laptop space? It’s a little unclear to me who they are targeting with $3k desktop hardware.
What I would love to see from essentially any ARM SoC (and platform) maker for Windows devices is more of a concerted effort to add x86 extensions in silicon. Not everything in x86 is patented, and they could seriously accelerate x86 translation the way Apple did with their memory modes. I’m not sure why none of these chip makers have done that yet, and it makes me skeptical that this is really going to work longer term. As evidence, see how the hype over (Snapdragon) ARM laptops, which was very favorable at first, has completely fizzled (and it’s not the first time for Windows on ARM).
A very nice side effect though of ARM moving in on x86 turf, is that Intel and AMD are finally looking seriously at efficiency in the consumer mobile space (they already have efficiency in the high power scenario – again, I don’t know who nvidia is targeting with a $3k desktop machine on ARM). Now we just need Intel/AMD to stop shipping in such cheap mass produced low end hardware platforms (think HP laptops), which is where I would bet they get most of their crappy reputation…
The Tegra 4i had the whole package, since it had an integrated modem (something missing from previous Tegra SoCs). The real issue of Tegra was Nvidia’s inability to keep up with Android’s breakneck pace of development when it came to drivers (board support packages), with phones featuring Tegra chips known for taking forever to take new versions of Android, back when new versions of Android mattered. I know, I own an LG Optimus 2X. Even the Nexus 9 received new Android versions with a delay, which was a bit embarrassing for team Nvidia.
Windows, with its backwards compatibility when it comes to drivers, should be a better fit.
Because those ARM SoC vendors really think x86-64 is “legacy”. Don’t get high on your own marketing BS, I guess.
NVidia GPU.. no thanks. The market does not need more closed hardware.
0brad0,
This bugs me a lot about nvidia. They make good products for GPGPU, but I hate that everything they make is proprietary…but at the same time it’s hard to find a better open solution for GPGPU. It sucks when the market isn’t competitive enough to provide solutions that check all the boxes that we’d like.
Until NVidia changes their businesses practices and developers completely open source drivers and stack I won’t give them a penny. They always do everything in bad faith.
0brad0,
I understand. However then many genuinely good FOSS projects become off limits because they’re built on nvidia’s cuda platform. Not for nothing but I’m often faced with these kinds of dilemmas over 1) the availability of FOSS options and 2) the compromises involved. For example: printers, remote management hardware, network power controllers, thermostats, oscilloscopes, solar controllers, bluetooth lighting & speakers, projectors, etc. It’s very frustrating but sometimes I’m forced to compromise especially when I am seeking a professional working turnkey product instead of a DIY project that I have to build myself. Keep in mind I’m saying this as someone who hates proprietary and loves FOSS. I am strongly put off by proprietary dependencies, which often cause problems later down the line because then we’re dependent on someone else. More power to you if you are able to do FOSS 100% of time, but I have to concede that I sometimes struggle to achieve it.
They are kinda sorta doing that. They moved some of the proprietary bits out of the driver system, and in to some firmware which runs on an embedded RISC-V CPU on the GPU card. This solved their concerns about people seeing the source code of whatever they put in that binary blob, and made it so they don’t have to ship binary blobs to all the various operating system and platforms to support their GPUs. It has improved things in meaningful ways. Was it all necessary? Probably not, but whatevs.
Edit: Why would I say “whatevs”? Because at a certain point, hardware is proprietary (it hasn’t been documented well for ages, and you can’t look in side the silicon and see what it’s doing), and I don’t have an issue with pairing proprietary hardware with a small kernel of proprietary software – it just moves the line. I’d prefer open hardware, like RISC-V (and the gall of nvidia to use RISC-V). Maybe one day we’ll see competitive open hardware CPUs and GPUs, but I don’t expect to see that for many decades, not at scale anyway.
CaptainN–,
The drivers are still proprietary and run in userspace. The nvidia linux kernel driver is just a shim, you can check it out. I will say this solved the driver breakages because previously nvidia’s latest downloadable drivers broke fairly regularly against the linux unstable ABI. Now that nvidia has mainlined a stable ABI for itself the breakages don’t happen any more, but it’s fairly useless for FOSS purposes. I personally don’t give nvidia credit here, it’s not still not open source. Even if you are a dev willing & able to add/fix features, you can’t because it’s still proprietary.
No, not really. Not even close. Hardware is NOT proprietary and to say it hasn’t been documented well shows incredible ignorance, but I’m used to that nowadays.
@obrado Does nvidia explain how their GPU hardware works in great detail?
@alfman Maybe I’m speaking out of turn here – I haven’t used nvidia in a long while. But I was under the impression they moved the binary/proprietary part to a co-processor on their boards, with a stable ABI to that, and that opened the door for a reasonable open source implementation similar to RADV for AMD. Is that not what NVK is?
CaptainN–,
When you write an open GL or cuda program, you have to link in nvidia’s huge binary blobs. Libcuda for example is 34MB. The shim isn’t that large but it doesn’t do much either beyond providing an IO interface.
I think the processor is there to do housekeeping, firmware updates, bootloading, monitoring, leds, fans and the like. I don’t think it’s part of any graphics pipeline, it’d be too slow for any type of synchronous operation having to do with graphics/compute, but if anyone knows different let us know!
0brad0,
In the case of nvidia, there are nouveau drivers that have been reverse engineered by 3rd parties We all know this work is open source and documented, but this what you meant here? If so I wouldn’t call that open hardware.
Maybe CaptainN– and I have both misundersood, but can you clarify?
I would think if the hardware was well documented, all of this would be moot – since then someone could read that hardware documentation, and figure out how to write their own binary blob. That doesn’t seem to be the case? I can imagine there may also be signing requirement that prevents running custom firmware, and the like, but I don’t know if that’s the case or not. But it’s hard to imagine the hardware being “not proprietary” – or well documented in the way that I’m thinking – documented enough for someone to completely replace that blob with custom code. In fact, it seems the entire purpose of that blog is to keep people from understanding how some aspects of the hardware works. Why keep it proprietary other than specifically for obfuscation?
But again, the effort around NVK – it looks like this is indeed based on binary blobs, more than just firmware level, with some APIs called GSP – that sucks. I’ll continue to steer clear of nvidia. AMD has great Linux support, and Intel doesn’t seem too far behind.
CaptainN-,
I’m testing out one of my cuda programs. Here are the runtime dependencies from userspace.
I couldn’t tell you what nvidia is doing with all these handles to nvidia0. Nvidia also launched 3 extra threads in my process and talks to it’s own threads with pipes. I don’t know what these threads are doing. Threads are often used in linux to compensate for the lack of asynchronous kernel interface (ie to avoid blocking) but I’m not sure if that’s the case here.
BTW this program links in a binary kernel that was built using nvidia’s nvcc compiler.
I found numerous posters asking about open source alternatives to nvcc, noting that parts of nvcc are derived from “open64”, which is actually open source. But all of them come to the same conclusion that the open source code of open64 is insufficient to replace nvcc. It seems like nvcc performs additional post processing to compile GPU kernels.
https://forums.developer.nvidia.com/t/is-it-possible-to-compile-nvcc-myself/8762
It seems so many companies want to use FOSS themselves to speed up their development, but then they don’t want their own works to be FOSS. This relates to the FOSS washing concept OS-SCi talked about in their presentation yesterday. Did anyone else join?
Edit: Er, this was supposed to be threaded – moving…
Why do so many of these comments read like people, who use smoke signals to communicate, trying to make sense of what gigabit ethernet is, does, and enables.
In any case. A 1 petaflop FP4 level machine on that form factor and price point is a clear marker as to where and how fast things are moving.