The stuff Hacker News actually recommends.

GPUs and CPUs · NVIDIA GeForce RTX

NVIDIA GeForce RTX 4090

74owners with an opinion
75%of verdicts positive
24/11positive/negative since 2024
3.5 yrslongest happy owner
2happy after 3+ years

Praised for performance and value; criticized for noise.

What people wrote

Likes itOwned 3.5 yearsStill going
PSA: Fast RAM isn't going to be getting cheaper anytime soon. Acquiring inference hardware is a really good way to own an appreciating hard asset. Learning how to use it and cool it is a hacker's journey worth taking. My 4090 I bought in late 2022 for $1600 is selling for a cool $3,489.95 right now, and going strong under nominal use. My DRR5 has tripled in value, my nvmes almost doubled. I grabbed a 128GB M5 Max MacBook Pro when they were still available and told all my friends to buy at least one. With that and a base M4 Studio 36GB, HuggingFace rates that hardware as: > Amazing! You have a total of 128.94 TFLOPS of computing power. 71.3% percentile on scale of "GPU Poor" to "GPU Rich" The way I see it, these are amazing machines that the richest folks are hovering up. I think they should be in the hands of regular people as much as possible. They depend on an incredibly global, increasingly fragile supply chain. If the become impossible to produce, their value would increase tremendously. I think they will become really valuable to you to use the tokens directly, but if that isn't the case, they can be rented out or resold. Please don't just buy any hold. Let's try to get as many people that can use them for decent things that help humans. For example: https://spectrum.ieee.org/small-language-models-ai-pharmaceu...
NeutralOwned 3.5 yearsStill going
It's a thought provoking article and I felt the pain when I shopped around for a new GPU lately to replace a 4090 I thought was faulty (eventually a cleaning of the PCIe connector solved those crashes). I bought it at the end of 2022 and three and a half years it seems like we've gone backwards, not forward on GPUs available for end users. They cost more and do less. But also consider that PCs have been an anomaly for very long. I don't think there's an equivalent market where you, as a consumer, can buy off-the-shelf cutting-edge technical pieces in your local mall and piece them together into a working device. It's a fun model, for sure, but I'm not sure it's an efficient model. It was just profitable enough to keep the lights on, thanks primarily to a bunch of Taiwanese companies in that space but it wasn't growing anywhere and the state of software is a mess. Apple the PCs collective lunch before DCs did. So have gaming consoles. So I weep for consumer choice but as things become more advanced maybe PCs and their entire value chain don't make a lot of sense any more. Obviously at the end there will still be consumer devices, because someone needs to consume all of this AI (at least people are thrown entirely out of the loop, but then all those redundant meat sacks will need entertainment to keep them content). We have the consumer device hyperscaler Apple doing rather OK even with these supply crunches although I'm not sure for how long.
Likes itOwned 3 yearsStill going
I don't know about that. I've had my RTX 4090 for nearly 3 years now. If I had a script that provisioned and deprovisioned a rented 4090 at $0.70/hr for an 8 hour work day for 20 work days per month. Assuming I get 2 paid weeks off per year + normal holidays over 3 years. 0.7 * 8 * ((20 * 12) - 8 - 14) * 3 = $3662 I bought my RTX 4090 for about $2200. I also had the pleasure of being able to use it for gaming when I wasn't working. To be fair, the VRAM requirements for local models keeps climbing and my 4090 isn't able to run many of the latest LLMs. Also, I omitted cost of electricity for my local LLM server cost. I have not been measuring total watts consumed by just that machine. One nice thing about renting is that it give you flexibility in terms of what you want to try. If you're really looking for the best deals look at 3rd party hosts serving open models for the API-based pricing, or honestly a Claude subscription can easily be worth it if you use LLMs a fair bit.
Neutral
I actually think one might? They do not have display outputs, but, at least with my 4090, if I plug the monitor into the iGPU instead, Windows still renders frames on it and just pushes them to the GPU that has a monitor attached. So the lack of output ports of those DC GPUs might not prevent that? Or maybe there are other limitations there I am not aware of, of course. Maybe the silicon cannot do the stuff needed for games or something like that. Yeah, probably that. But something like the L40 - which is just a 4090 with more VRAM - might be usable like that? Maybe? You SV guys go use your stock options and try it out for science.
Neutral
The 4090 I bought in '23 is selling for ~3x what I paid for it now in late '26. There's a huge squeeze going on with computer hardware using relatively cheap borrowed money. Relatively cheap money is flowing into a (artificially?) limited GPU/memory asset class. And this asset class is growing much faster in value than the interest on that money. So if you borrowed heavily to buy GPUs in the prior 3 years, you're likely coming out well ahead even if you are paying 8 to 10% on the money. That dynamic will fuel further borrowing until we get to some kind of equilibrium, or some kind of washout occurs where interest rates spike higher and/or the value of GPUs and GPU services start leveling off or even declining. We might get a test soon as the FED has started hiking.
Doesn’t like it
I tried "write binary tree in rust and ensure it compiles with rustc" using the coder model with their chat it didn't produce the correct output and their chat can't call tools. Then with opencode it went on thinking loop until it exhausted the thinking budget. 4090 16gb, 96 gb RAM. But I like this overall approach of neutering experts and loading them from ssd to make the model run in low VRAM systems.
Doesn’t like it
Do you think someone who had a 4090 gpu isn't aware of their value? I don't need a $1500 video card sitting in a box nor do I care that it might potentially be worth $2k now. For what its worth, since you seem completely lost on the actual market- almost nobody offered to buy it for the $1400 I wanted, it took me months to sell it. It isnt consumers buying the 5090s. People who don't own and sell these cards constantly think you can just stick one on ebay and sell it for a fortune, that isn't reality.
Neutral
In my laptop 200% fractional scaling of the built-in QHD+ OLED display on KDE Plasma is great, but when I connect my portable 1080p displays or e-paper monitor (UXGA) the fractional scaling of it is not as good as when compared to when the same monitors are used with macOS (M4) in HiDPI mode under lower resolutions. As the other comment says, I suspect it to do with GPU as the built-in displays are usually driven by intel and external displays are always driven by my Nvidia 4090. But surprisingly, when I drive my e-paper monitor using original Jetson Nano (Ubuntu 20.04 ESM / Nvidia Maxwell); the fractional scaling is quite good perhaps even better than macOS.
Likes it
On my 4090 setup I'm getting 86t/s on a 12B Q6_K quantized model running entirely in VRAM. The current GPUs are optimized for processing huge numbers of triangles per second. There are three things at play here: 1. the organization of the data being sent to the GPU to optimize throughput; 2. the speed at which the GPU can read that data from its VRAM; 3. how many triangles it can process in parallel by using individual compute units. I suspect that given parallel improvements for neural networks, we'll see similar improvements: 1. optimizing the structure of the weights in the model for efficient access by the CPU/GPU/NPU/TPU; 2. efficient access of data strides (matrix rows) in the memory, e.g. being able to read multiple 2x2 matrix values in one clock cycle, or stepwise pairs of values (a(i,j), b(j,k)) needed for matrix multiplication; 3. parallel compute for matrix and tensor multiplication and other operations needed by neural networks.
Likes it
If you're using a MoE model, then why do you care about the larger RAM offered by these devices? That's the main problem with low bandwidth devices: they limit the effective ram you can make use. I do (and have historically done) quite a work with both local LLMs and local diffusion models. I have an M3 Max MBP at 400 GB/s and also a desktop with a RTX 4090 with 1,008 GB/s While the M3 Max MBP can serve up MoE reasonably fast (~60 token/sec)the RTX 4090 is an entirely different experience (~170 token/sec). I also do a fair bit of experimentation and am currently running a custom decoder that requires expensive look-ahead, but I'm still able to get a usable 25 token/s on the RTX. The raison d'etre for the DGX spark is not practical home inference, but rather offering the same fundamental architecture as data center cards for a affordable CUDA prototyping. If you want to build software to run on H100s, you probably can't justify buying (and running) a single card. The DGX spark solves this by having the same fundamental setup as what those cards have. That makes these non-NVIDIA DGX-like devices confusing to me. The entire benefit of the DGX series is the NVIDIA architecture itself. Anyone interested in home LLMs should decide whether a Mac or a dedicated GPU is the more sensible path based on their budget and other computer use. Each has their own benefits.
Likes it
This is a great post that covers a lot of the recent ground. I have a very similar setup after a very similar journey, minus the RTX6000. Worth noting though that a lot of the recent changes make a single 3090/4090 much more viable here too. MTP and the recent improvements to kv quantization in particular, as well as model-specific template & quant fixes. I run a 4090 with the 4-bit quantized variant of the same model now and have had a great experience. Qwen3.5 was already a big step up, but with 3.6 and the rest of the improvements it's substantially more reliable as a daily use tool and I find myself reaching for hosted models a lot less. Feels like I could work entirely without them if they were to disappear without going back to typing every line of code myself. To make 4-bit fit on one card with reasonable (100k+) context needs a bit more care though. And tuning can be highly specific to your machine, gpu and use-case. But I use a headless server, offload multi-modal to CPU, use fit-target to reduce wasted memory and use q8_0 kv since the 4090 performs well with it... In addition to most of the same config as the author elsewhere. I get 50-60tps generation with a power limit of 275W (450W is default), more than enough to offer a roughly an Opus-speed feedback loop. I haven't seen many of the issues with looping the author mentions. But I did with Qwen3.5 and in particular other 4-bit quants in the past. But the difference is probably a mix of the improvements above, as…
Neutral
80tp/s with 5080 3090 combo is wild. I’ve been working with a 4090 and two Tenstorrent p150 cards, and manage only about 30 tps utilizing all three for qwen3.6 27b q8. Guess I got more optimization to do. Would like to see the perf of their setup with and without mtp and ngram speculative decoding though, as well as parallel decode performance (once llamacpp mtp plays well with multiple slots). Being in California electricity alone puts this non-competitive with just paying a cloud though.
Doesn’t like it
Its nvidia attempt to gain additional market share and expected as well. If the whole ecosystem is around nvidia and its the easiest way of running stuff, Nvidia offering more enterprise infrastrcuture allows companies to just buy directly nvidia. Nvidia is also very very rich and pushes the boundaries of stuff. They stoped waiting for industry standards. You can see this in there network stuff. All nvidia. Next logical step (at least now, not something i thought about) was there CPU for their GPU racks/clusters/systems. Now they have everything anyway, RTX Spark is just logical. I don't think its specificly targeted at Apple at all. Apple has like 10-15% market share and just because some IT nerds buy themselves a mac mini doesn't mean much. Plenty of them actually just run openclaw without local models. Something which surprised me quite a lot. But i have two 4090 at home. They consume a lot of power and i had to research the proper Mainboardmodel and had to mod one 4090 to use water cooling because they run too hot. There Spark setup was at 3k, way to expensive for normal people. If they can get this down and sell more, great for their ecosystem (strengthening it) and getting more money from people. It does surprise me though that they have enough capacity for this chip and not just putting everyting in Rubin but perhaps the build out has slowed down a little or they start to diverse already for economic savety
Likes it
> GPU compute for self-study Those suggestions they make for a B200 start at $4.99 an hour. Is that really required, for starting out? I've been tinkering with my own from-scratch LLM, but in the early phases I don't need anything more than a 4090 on Vast.ai
Likes it
> Based on what? I take it you haven’t actually run any of the current gen local models? They all fit on fairly accessibility hardware, and their performance is at least on par with what I was paying for last year. I have one of my agents running entirely from a local model running on a MBP and it has repeatedly shown it’s capable of non-trivial tasks. Playing around with another, uncensored, local model on my 4090 desktop has me finally thinking about canceling my personal Anthropic subscription. Fully private, uncensored chat is a game changer. For work it’s still all private models but largely because, at this stage, it’s worth paying a premium just to be sure you’re using the best and it saves the time of managing out own physical servers. But if we got news tomorrow that Anthropic and OpenAI were shutting down, a reasonable setup could be figured out pretty quickly.
Neutral
I think that we're getting there. I put together a workstation in early 2023 with a single 4090 GPU. I did it to run things like BERT and YOLO image classifiers. At that point the only "open weights" LLM was the original Llama from Meta, and even that was open-weights only because it was leaked. It was a very weak model by today's standards. With the same hardware I now get genuine utility out of models like Qwen 3.5 for categorizing and extracting unstructured data sources. I don't use local models for coding since commercial ones are so much stronger, but if I had to go back to local models for coding too they would be more useful than anything commercially available as recently as 4 years ago.
Likes it
The initial benchmarks I saw didn’t really have the DGX spark (GB10) doing much better in generation throughput than the AMD Strix Halo. Prefill, the GB10 does pretty well, much better than the Strix. Memory bandwidth is 273 Gb/s. Which is nowhere near a GPU’s. It’s a 4K machine. Personally, I’d rather have two GPUs and run a quantize model. I have two 32GB AMD r9700 cards, cost $2600. Quantized models get me 120K ish of context window and TPS is about 60% of what I see with the same model on my 4090 (which only has enough vram to load weights and about 6K context). Sure I can’t run a 100B+ model but neither can a single GB10 unless no context window is what you are going for. So you buy a second 4K machine?
Doesn’t like it
I have trained transformers on a 4090 (not language models). Here’s a few notes. You can try out pretty much all GPUs on a cloud provider these days. Do it. VRAM is important for maxing out your batch size. It might make your training go faster, but other hardware matters too. How much having more VRAM speeds things up also depends on your training code. If your next batch isn’t ready by the time one is finished training, fix that first. Coil whine is noticeable on my machine. I can hear when the model is training/next batch is loading. Don’t bother with the founder’s edition.
Doesn’t like it
I have to say, the experience of trying to sign up for Nvidia Enterprise so you can try the "NIM" packaged version of this model, is just icky and and awful now that I've gotten used to actually free and open models and software. It feels much nicer and more free to be able to clone llama.cpp and wget a .gguf model file from huggingface without any registration at all. Especially since it has now been several hours since I signed up for the Nvidia account and it still says on the website "Your License Should be Active Momentarily | We're setting up your credentials to download NIMs." I really don't get Nvidia's thinking with this. They basically have a hardware monopoly. I shelled out the $4,000 or so to buy two of their 4090 GPUs. Why are they still insisting on torturing me with jumping through these awful hoops? They should just be glad that they're winning and embrace freedom.
Doesn’t like it
Who is "they"? RTX 4090s are terrible for this task. Off the top of my head: - VRAM (obviously). Isn't that where the racks come in? Not really. Nvidia famously removed something as basic as NVLink between two cards from the 3090 to the 4090. When it comes to bandwidth between cards (crucial) even 16 lanes of PCIe 4 isn't fast enough. When you start talking about "racks" unless you're running on server grade CPUs (contributing to cost vs power vs density vs perf) you're not going to have nearly enough PCIe lanes to get very far. Even P2P over PCIe requires a hack geohot developed[0] and needless to say that's umm, less than confidence inspiring for what you would lay out ($$$) in terms of hardware, space, cooling, and power. The lack of ECC is a real issue as well. - Form factor. Remember PCIe lanes, etc? The RTX 4090 is a ~three slot beast when using air cooling and needless to say rigging up something like the dual slot water cooled 4090s I have at scale is another challenge altogether... How are people going to wire this up? What do the enclosures/racks/etc look like? This isn't like crypto mining where cheap 1x PCIe risers can be used without dramatically limiting performance to the point of useless. - Performance. As grandparent comment noted 4090s are not designed for this workload. In typical usage for training I see them as 10-20% faster than an RTX 3090 at much higher cost. Compared to my H100 with SXM it's ridiculously slow. - Market segmentation. Nvidia really know…
Doesn’t like it
Good question, I think the answer is even at thousands a window device battery can't hit 18 hour specs. Can someone name a windows device even at 2k+ that acts like an M chip? In fact the pricier windows usually mean GPU and those have worse battery then cheap windows(my 4090 is an hour or so off charge)
Doesn’t like it
I have a MacBook M2 and a PC with a 4090 ("just" one of them) - the VRAM barrier is usually what gets me with the 4090 when I try to run local LLMs (not train them). For a lot of things, my MacBook is fast enough, and with more RAM, I can run bigger models easily. And, it's portable and sips battery. The marketing hype is overblown, but for many (most? almost all?) people, the MacBook is a much more useful choice.