The RTX 3090 came out in September 2020. Back then NVIDIA sold it as a giant gaming and creator card with 24GB of GDDR6X memory. It was expensive, huge, hot, and honestly a little ridiculous for most gamers.
Six years later, that same 24GB number has become useful for a completely different reason.
Perplexity has now brought Portable Computer to Windows RTX PCs. It is the local version of Perplexity Computer, with the agent harness, orchestrator, local model, scheduled tasks, connectors and sandboxed execution running on your own machine. The hardware requirement for local inference is very clear: a supported NVIDIA RTX GPU with 24GB of VRAM or more.
That means an RTX 3090 qualifies.
Access without medium partner: RTX 3090 Run Perplexity’s Local AI Agent
An RTX 5080 with 16GB does not. An RTX 5070 Ti with 16GB does not. An RTX 5070 with 12GB does not. Even a 16GB RTX 5060 Ti misses the requirement. On laptops, the new RTX 5090 Laptop GPU reaches 24GB, but the RTX 5080 Laptop GPU stops at 16GB.
So basically, a six-year-old Ampere card can enter a local AI workload that several much newer Blackwell GPUs are blocked from before performance even matters.
That is a weird result, and it tells us something bigger about where PC hardware is going.
This is not really an old-GPU story
At first I wanted to frame this as another “RTX 3090 refuses to die” article. It is partly that. The card keeps finding new jobs because NVIDIA gave it 24GB when most gaming cards were nowhere near that number.
But the bigger story is what Perplexity chose as the minimum.
Portable Computer is not just local chat. Perplexity describes it as a local agent stack. It can plan multistep work, read local files, run recurring jobs, use connected apps, call tools and keep the main task on the PC. When a step needs web access or a stronger cloud model, the system can ask before sending that step outside the machine.
On supported Windows RTX PCs, the current local model listed by Perplexity is PPLX 27B. Local work does not consume Perplexity Computer credits, although Portable Computer itself is currently available to Pro and Max subscribers.
This is important because an agent needs more than a model file sitting in memory. It needs room for context, KV cache, runtime buffers and whatever the inference engine is doing around the model. The agent may also be handling long files or several steps in one run.
Perplexity does not publish a simple breakdown saying “X GB is weights, Y GB is cache, Z GB is overhead.” I spent more time than I expected checking this because NVIDIA’s launch post mentions a local model “such as Qwen 3.8 27B,” while Perplexity’s current Windows setup page lists PPLX 27B as the available local model and says Qwen 3.8 27B is not available on Windows RTX PCs right now. So I am sticking with Perplexity’s current product page here.
The 24GB requirement is the useful number.
The RTX 3090 was accidentally built for this moment
NVIDIA did not design the RTX 3090 for Perplexity Portable Computer. Obviously.
The card launched on September 24, 2020, with 10,496 CUDA cores and 24GB of GDDR6X. NVIDIA pitched it at extreme gaming, 8K, creators and researchers who needed larger datasets. The memory was one of the reasons the card existed in the first place.
That choice aged better than anyone expected.
Local AI has a habit of making old high-memory GPUs useful again. A card can lose badly in gaming performance to a newer generation and still stay interesting because one number did not shrink: VRAM.
This is different from normal gaming upgrades. In a game, a newer GPU with much higher compute and better architecture can usually compensate for a lot. In local LLM inference, there is a hard moment where the model and runtime either fit in GPU memory or they do not.
Yes, you can offload to system RAM. You can quantize harder. You can choose a smaller model. You can move part of the workload elsewhere. But once that starts happening, the comparison changes.
Capacity becomes a gate before speed becomes a benchmark.
The RTX 3090 happens to sit right on Perplexity’s gate.
New does not mean enough VRAM
This is the uncomfortable part for anyone who just bought a newer card.
NVIDIA’s current desktop lineup has an RTX 5080 with 16GB, RTX 5070 Ti with 16GB, RTX 5070 with 12GB, and RTX 5060 Ti configurations with 16GB or 8GB. These are newer GPUs with newer Tensor Cores, better gaming features and newer architecture.
For this specific Perplexity local-agent requirement, none of those memory sizes reaches 24GB.
The desktop RTX 5090 does. It has 32GB.
That means the product line has a strange gap. If local AI becomes part of why someone buys a GPU, the jump from 16GB to 32GB becomes much more important than it looked on a gaming spec sheet.
And laptops are even harsher. The RTX 5090 Laptop GPU has 24GB, while the RTX 5080 Laptop GPU has 16GB. So the local Perplexity requirement effectively pushes laptop buyers all the way to the top memory tier if they want this feature on NVIDIA mobile hardware.
I do not think Perplexity sat down and decided to make the 3090 fashionable again. They picked a hardware floor that gives their local agent enough room to work.
The funny part is what that floor exposes about the rest of the market.
A faster GPU can still fail before the race starts
Imagine two PCs.
One has an RTX 3090. The other has a much newer RTX 5080.
For gaming, rendering and plenty of AI tasks that fit in memory, I would expect the 5080 to be the more attractive card. It has newer architecture and much higher memory bandwidth. I am not trying to pretend a 3090 became magically faster because Perplexity released an app.
But Portable Computer does not begin with “which GPU has more throughput?”
It begins with “do you have at least 24GB of VRAM?”
The 3090 says yes.
The 5080 says no.
That is the difference between performance and eligibility. We normally mix those two ideas together because modern GPUs tend to improve in several ways at once. Local AI is breaking that assumption.
A newer card can be faster and still not be eligible for a workload because the memory capacity went the wrong direction for that use case.
That is kind of brutal.
27B models explain why 16GB is getting awkward
A 27B parameter model is not absurdly large by 2026 standards. We are already talking about 70B, 100B and much larger MoE models on local workstations.
Still, 27B is big enough that memory planning matters.
If you do very rough math, 27 billion parameters at exactly four bits would be around 13.5GB for raw weights. But that number is only a napkin calculation. Real model packages include quantization metadata and other data structures, and inference needs memory for context, caches and runtime work. Perplexity may also be targeting a setup that has enough headroom for agent use instead of one that barely boots the model.
So I would not look at 13.5GB and conclude that Perplexity is wasting 10GB for no reason. We do not have their exact VRAM budget published.
What we do have is the supported configuration they are willing to ship: 24GB minimum.
This is the same reason a machine that technically loads a model is not always a good machine to use every day. You can squeeze things into memory with aggressive settings and then discover context size is tiny, prompt processing is awful, or the first tool-heavy agent run pushes the setup over the edge.
I have done enough local model testing to know that “it loaded” is sometimes the least useful benchmark in the room.
The RTX 3090 has another advantage: the software already knows it
There is also a boring reason the old card keeps showing up.
It is NVIDIA.
Portable Computer’s PC requirement is specifically built around supported NVIDIA RTX GPUs. That means a new AMD system with a huge unified-memory pool does not automatically qualify, even if it has far more than 24GB available to its GPU.
This is where local AI gets annoying. Hardware capacity is only one part of the decision. Software support can override the entire spec sheet.
A 128GB or 192GB AMD mini PC may be better for another local model workload because it can keep a huge model in one shared pool. For Perplexity Portable Computer today, the official Windows path is NVIDIA RTX.
So the 3090 gets credit twice. It has enough memory, and it sits inside the software ecosystem Perplexity chose to support.
CUDA has been doing this for years. The ecosystem itself becomes part of the hardware spec.
Local agents are making memory more important than local chat did
A simple chatbot can be surprisingly forgiving.
Load a model. Send a short prompt. Generate a few hundred tokens. Done.
Agent workloads get messy fast. A task can involve a local repository, several PDF files, tool output, a browser step, a connector call, a long conversation history and scheduled work that runs later without you sitting there watching it.
Perplexity’s own examples include reviewing pull requests, analyzing tax documents and finding where a startup’s activation funnel is dropping users. The whole point of Portable Computer is that sensitive material can stay on the device for the local parts of the job.
That means context becomes part of the product, not an optional benchmark setting someone turns up for fun.
And once context gets bigger, memory pressure stops being theoretical.
The agent also needs enough room to stay stable during long jobs. A setup that is fine for a five-minute conversation may be a bad setup for a task that runs across many files for an hour.
This is probably why the 24GB floor feels conservative compared with the smallest possible way someone could force a 27B model to run.
I would rather Perplexity ask for too much VRAM than pretend an 8GB card is a local-agent workstation because the model managed to produce one sentence.
Privacy is the reason people will care about this
The local part is not just a performance trick.
Perplexity says tasks can stay on the Windows or Linux machine, and the local model work does not use Computer credits. Its connectors include things like Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub. When the agent wants a cloud model or external web access, the user can approve that step.
For code, tax documents, internal company files and personal data, that split makes sense.
I would still read the permissions carefully. “Local-first” does not mean every connected service magically becomes local. If the agent posts to Slack, searches the web or escalates a reasoning step to a cloud model, something is leaving the PC because that is the job you asked it to do.
But keeping the main model and sensitive file analysis on hardware you own is a real change from sending every prompt and document to a remote model by default.
This is where a six-year-old 24GB GPU suddenly has a job it never had at launch.
The subscription part is easy to miss
Portable Computer is local, but it is not a free open-source agent you download and forget about.
Perplexity currently says the feature is available to Pro and Max subscribers.
Local model work does not consume per-token Computer credits, which can matter if you run a lot of repetitive tasks. But the product still lives inside Perplexity’s subscription and software stack.
That distinction matters because “local AI” is starting to mean several different things.
There is fully open local software where you choose the model, runtime and UI yourself. There are polished local-first products like this where the company manages most of the setup. Then there are hybrid products that run some steps locally and move harder work to the cloud after approval.
Portable Computer is very much in the second and third bucket.
That is probably why it can give normal users a cleaner setup than stitching together Ollama, an agent framework, a sandbox, connectors and model routing by hand.
You are trading some control for convenience.
Fair trade for many people, I think.
The 3090 is not suddenly the best GPU to buy
This needs saying because hardware articles can get silly very quickly.
An RTX 3090 is old. It is a large card. It uses a lot of power. It lacks newer Blackwell features. Used cards can have an unknown history. Some have spent years gaming, rendering or mining. And if your workload fits comfortably on a newer GPU with less memory, the newer card may be the better machine in almost every practical way.
I would not buy a 3090 just because one Perplexity requirement happens to line up with its 24GB memory.
But if I already owned one?
Different story.
A card that might have looked ready for retirement can now run a local agent product that officially rejects several newer GPUs on memory capacity alone. That is useful life extension you can actually point to.
And for people shopping used hardware specifically for local AI, it is another reminder that VRAM can age better than GPU generation.
The 3090 is not winning because Ampere became new again.
It is winning because NVIDIA gave it enough memory six years ago.
This should worry people buying 16GB cards for “AI”
PC companies use the word AI on almost everything now.
AI laptop. AI desktop. AI GPU. AI PC.
Then a real local agent arrives and says 24GB minimum.
That mismatch is going to become harder to ignore.
A 16GB GPU is not useless for local AI. Far from it. There are many good models and workloads that fit inside 16GB, and smaller models keep getting better. If your main use is gaming plus some local chat or image work, 16GB can still be completely reasonable.
But the word AI-ready is doing too much work in marketing.
Ready for which model?
Ready for what context size?
Ready for image generation, local agents, coding models or multimodal work?
Ready today, or ready for the next two years?
Perplexity has accidentally given buyers a very simple test. One new local-agent product already draws its line at 24GB.
That does not make 24GB the universal minimum for AI.
It does make 16GB look less future-proof than the sticker on the box suggests.
The next GPU war may be about memory tiers, not generations
The RTX 3090 versus newer cards comparison looks funny because we are used to reading GPU names like a ladder.
30 series, then 40 series, then 50 series. Newer means better.
For local AI, the ladder is becoming two ladders.
One is compute speed.
The other is memory capacity.
Sometimes they move together. Sometimes they do not.
The RTX 5090 is the easy example because it has 32GB and huge throughput. It qualifies and it is fast. The awkward part is everything below it. Once the next tier drops to 16GB, a six-year-old 3090 suddenly has a capability the newer card does not have for this product.
That is why I keep coming back to memory when looking at local AI hardware.
Not because bandwidth and compute stopped mattering. They matter a lot.
Memory decides which race you are allowed to enter.
The six-year-old card did not get smarter. The workload changed.
This is probably the part I like most about the story.
The RTX 3090 did not receive a secret update. Its VRAM did not grow. Ampere did not become Blackwell overnight.
The software changed around it.
In 2020, 24GB was sold around 8K gaming, creators and big datasets. In 2026, the same memory capacity clears the entry requirement for a private local AI agent that can work through files, tools and recurring tasks on the machine.
Meanwhile, several newer GPUs were designed with less memory because their target market and price point made that trade reasonable.
Now local AI is assigning value to the old decision again.
I would not call the RTX 3090 future-proof. That phrase usually becomes embarrassing if you wait long enough.
But it has aged in a way most gaming hardware does not.
Perplexity Portable Computer needs 24GB.
The RTX 3090 has 24GB.
Six years later, that very simple match is enough to make a card from 2020 relevant in one of the newest PC AI workloads of 2026.
And if local agents keep asking for more memory, we may look back at all those 12GB and 16GB “AI GPUs” and wonder why the old monster was the one that got the important part right.