FRESH
Hacker News
Hacker News Hacker NewsThe reality is even cutting edge games and consumer workloads don’t actually take full use of the PCIe bandwidth of the GPU or the bandwidth of its GDDR memory. Even local AI use cases don’t substantially or meaningfully benefit from faster memory, at least to average consumers.
A unified memory pool does two things:
1) Lets systems optimize utilization based on need, rather than be confined to specific pools
2) Reduce overall memory cost, by letting system builders purchase a single type of memory in bulk instead of having to figure out GDDR vs DDR memory placement (important for SFF/portable machines)
So at a time when memory is expensive, unified pools make more sense. Even when memory becomes cheap and plentiful again, it’s just practical at this point to allocate a larger overall pool instead of managing discrete sets.
The one big drawback is security. A shared memory pool means side-channel attacks against memory from the GPU or CPU could potentially compromise the other as well, meaning memory-safe designs are going to be critical to security going forward (which is good for Rust adherents, I figure).
I don't know who will be the winner but with some of the recent releases from gemma it seems more probable that you may run some models locally if only from a cost perspective, not even considering business security. Not sure how this type of architecture would make for good gaming though, puts into question the whole statement.
"Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers" - side note but this guy puts this everywhere, gives me probably the inverse of what he is marketing for.
1. Yes it has the same number of cores as a 5070 mobile. It’s also running at a shared peak of 2/3 the bandwidth and a shared peak of 2/3 the TDP. The GPU by itself will likely perform at half the dedicated units performance
2. Apple may not have SVE2 but they do have the AMX (private) and SME. I don’t see why he thinks the SVE2 will give him more performance than the SME.
3. He mentions a single core type but doesn’t mention the total makeup. We already have known for a year how the DGX Spark compares to Apple chips. For CPU it’s roughly equivalent to an M3 Pro and for GPU compute (not rasterization) it’s between an M4 Pro and M4 Max without considering bandwidth.
The real advantage to these is that they run CUDA. That’s it. Otherwise when they launch they’ll be 2-3 generations behind where Apple is and 1 gen behind AMD.
The other super power of the DGX Spark was the NIC for pairing them together. But that’s been removed here too.
Before we get local AI, we'll be using hybrid AI.
Running big models locally is unrealistic ($$$$$) but, if you imagine an Agentic Workflow where some bits run on the cloud and other smaller tasks locally, it's an amazing deal. You don't need Opus/Code/DeepSeek/Kimi/etc to do basic stuff that models like Gemma4:12b/Qwen-27b can do locally with much less latency.
Having a laptop where I can use a remote big model and combine it with 5 local domain specific models, is something I would love to do today. Imagine using OpenCode and you've a small model deciding which tasks run locally, then decides if you've a good local model for XYZ task or if we use a cloud model.
My main concern is: Is this hardware powerfull enough to allow local quick models switch? Unlikely but I hope I'm wrong
It's just a personal computer. It normally runs multiple operating systems just fine.
Windows PC sounds like people talking about tech who are either payed by M$, or embed pictures into Word documents to send them.
Nobody has to kill the fun those OS agnostic machine allow, by artificially bind them to a shitty OS.
https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-...
I have been somewhat surprised at the lack of commentators observing that this is Microsoft and above all NVIDIA launching a device that is fundamentally at odds with the metered cloud model of AI.
When you look at the other announcements and murmurings (better offline BYOK for Copilot, talk of an unmetered AI future) I think it’s clear that these two firms understand that cloud-only AI is not sustainable or inherently in their interests. But their willingness to undermine OpenAI with a product like this is notable.
I've found it very useful for running big models, but it's not a screaming powerhouse in terms of raw compute.
A RTX Pro 6000 has ~24K 5th generation tensor cores, I'm guessing this would then be 1/4 of the count but 6th generation? Wasn't clear from the images.
(HN reaction to Vision Pro back in 2024 is almost hilarious if not ridiculous, looking at it today. I knew it would be a flop and I was so right.)
nvidias master plan may be making it the new normal to have "only" 400GB/s bandwidth, thus gatekeeping local model usage further behind "more memory but not as fast as the cloud can do it"
> The memory is not as fast as dedicated GPU memory, but it is cheap enough while delivering enough bandwidth to run AI models locally.
So, the reason "dedicated GPU memory" is fast, isn't because it's "dedicated"; it's because the types of memory built into GPU cards — GDDR and HBM — are designed for throughput over latency.
Which is to say, GDDR and HBM memory could be shared with the CPU in UMA while still being "fast" (for GPU use-cases.) In fact, the PS4/5 and Xbox 360 / One X / Series consoles have UMA architectures that use GDDR memory as their main memory, with no regular DDR memory to be found.
What I don't understand: why don't we see UMA architectures where there's both regular DDR and GDDR/HBM memory mapped into the address space of the CPU+GPU? That seems like the best of both worlds: you'd have some memory that's "tuned" for random-access CPU usage (regular DDR), and some memory that's "tuned" for streaming GPU usage (GDDR/HBM), but either type of memory can still be put to the use it wasn't "tuned" for, just with slightly-worse performance.
I guess you'd need to do a bit of software work:
1. a bit of work in the OS kernel / malloc library to get CPU workloads to "prefer" allocating DDR memory over the GDDR/HBM memory until they've exhausted DDR memory (or maybe not, if you just tell the kernel the GDDR/HBM memory is something like a zswap thinpool);
2. and a bit of work in supported ML frameworks, to teach them about a hybrid strategy between UMA "allocate anywhere, it's all the same" and NUMA "keep assets in VRAM if possible; if you spill assets to RAM, then they must stream into VRAM on access" (i.e. "at allocation time, allocate as if the system were NUMA, VRAM first then spilling to RAM; but at execution time, use the UMA codepaths, no need to copy RAM into VRAM.")
...but once that's done, it's done.
It's an interesting "newcomer" and the more the better but calling this a "beast" and a "game changer" is ridiculous to say the least.
Then there is the price..
The obvious comparison here is the M5 Max where you can buy a Macbook Pro with 128GB of also unified memory. Obviously CUDA cores are specific to NVidia so it's hard to directly compare but I've seen claims that the M5 Max is roughly equivalent to ~4000 CUDA cores. This obviously depends on workload and whether the CPU supports the precision you want to use (eg FP4).
The M5 Max has memory bandwidth of 819GB/s. The RTX Spark I believe is ~600. So it might be slightly better than the current generation of Macs but likely worse than the expected M5 Ultras of the new Mac Studios (likely Q3 2026).
For comparison, a 5090 has >20k CUDA cores and 1800GB/s memory bandwidth with 32GB of VRAM. The RTX 6000 Pro (at ~$10k) has 96GB of VRAM, same bandwidth and ~24k CUDA cores.
We have to see what RTX Spark systems sell for but the DGX Spark is in the Mac Studio price range (~$4k).
I do think Apple has a real opportunity here but there offerings aren't quite there yet. The M5 Ultras might be a really attractive option for local LLMs. I expect them to be in high demand.
Nvidia going from GPU to CPU now?
Up to $5000 because why not?
With that money you can build a real PC with rtx 5090!
A powerful new chapter for Windows PCs, accelerated by Nvidia RTX Spark
https://news.ycombinator.com/item?id=48352693
Nvidia RTX Spark
Decent single core (a long ways from Apple level, but decent), but it makes up for it in cores to provide M5 level performance, CPU wise. Memory bandwidth it is kind of starved, at 1/6th many GPUs.
They got Microsoft to customize Windows for the RTX Spark, and will likely have to brutally throttle it when running as a laptop (it's literally a 140W TDP chip), and that's neat. It's going to be a very expensive laptop.
Tech companies have strangled their own market.
Nvidia is milking the market now. We need more competition again - currently we have a mafia control the prices, not just Nvidia but all the AI companies. The price increases should be paid for them, not by us. "Free market" is being manipulated by them here.
I'd say this relates directly to the cost of running AI models remotely.
And we won't know what the actual cost will be until AI vendors recover the huge pile of cash they've dumped into development (plus interest).
Bill Gates had a quote some years ago...
People have still not learned how fast we improve our tech and how much cheaper thing gets I guess :)
Clip me :). You are currently living through the final stages of unrestricted computing in the hands of the 'public'. Our regimes are going to pull up the drawbridge in the name of 'safety'. Download the open models asap and prepare for an airgapped computing environment. That will be your frontier in not extremely neutered AI in the near future.
I am so hoping I'm completely wrong on this btw.