672 GB/s bandwidth on the Pro 4000 vs 936 on a 3090 and 1,008 on a 3090 Ti. LLM inference is bandwidth-bound — the 6 year old card literally pushes tokens faster.
No NVLink. Can't pool VRAM. Same 24GB. The 3090 wins on every metric that matters for local LLM work except power draw.
Now look at the pricing landscape:
• RTX Pro 4000: ~$1,600 new — 672 GB/s, no NVLink
• RTX 4090: ~$3,400 used — 1,008 GB/s, no NVLink
• RTX 5090: ~$4000 — 1,792 GB/s, 32GB, no NVLink
• RTX 3090: ~$1,000 used — 936 GB/s, 24GB, NVLink
• RTX 3090 Ti: ~$1,200 used — 1,008 GB/s, 24GB, NVLink
The 3090 Ti matches the 4090's bandwidth at half the price AND has NVLink. Two 3090 Tis give you 48GB pooled for ~$2,400 — a 5090 gives you 32GB for ~$4,000 with no way to pool anything.
Google just dropped Gemma 4 under Apache 2.0, Anthropic just repriced cloud access, and every new open-source release pushes more people toward local inference. The 3090 is the last consumer GPU with 24GB + NVLink. Supply is fixed. NVIDIA isn't making more. Do the math on where these prices are headed.
CHEAP CHEAP CHEAP
Better.
Within a service and repair business AI has aided greatly in invoicing. Specifically in descriptions that breakdown technical jargon simply and comprehensively for customers to truly understand the scope of work being provided. Eventually I should be able to fully automate most computer tasks such as invoicing and more.
All local too, I’d say 90% does not require frontier models.
@loktar00 We’re going through an unknown by the majority “gold rush” if you will.
3090s are the gold and nvlinks are the shovels. Easier for resellers to buy up the supply and resell higher. However, there also other types of equipment that could be “shovels”
Both valid cards for different reasons.
Pro 4000 wins on power and density. 140W single slot is clean. Stack em all day. Native FP4/FP8 in hardware is a real edge for running compressed models fast.
3090 ti wins on bandwidth, pooled VRAM, and price. Two of them give you 48GB unified at 112.5 GB/s between cards for $2,400. Two Pro 4000s give you 48GB total but split 24GB islands talking over PCIe at 32 GB/s for $3,200.
Same total VRAM. 3.5x faster link. $800 less. And with 48GB pooled you dont need FP4/FP8 you have the room to run Q8 where it matters.
More VRAM = less compression = better answers. More bandwidth = faster
To each his own. Although for both instances I believe the 3090s bring more value to the majority looking for infrastructure independence. I own 4.
@loktar00 Anthropic billing change strengthens our thesis on a potential 3090 price boom.
Local inference demand only increases from here.
infrastructure independence. It just makes sense
I say we have a 2 week window
before this propagates further.
@MemoryReboot_ Did a deeper dive and yes I do agree on a ~1000$ floor
Also, hard to distinguish real and fake listings.
If we’re going purely from online listings on your common e stores
5090 floor ~ 3800
4090 floor ~ 3400
3090s floor ~ 1200
Lowest cost barrier to entry using 30B+ Gemma 4 models. 3090s
3090s get them while they’re cheap.
“It’s mostly about VRAM. The 3090 at 24GB enables the same 30B+ models as cards costing three times more, and newer cards are only 2.6-2.7x faster despite that price premium. The 3090 Ti’s only real edge over the base 3090 is ~8% more memory bandwidth (1,008 vs 936 GB/s) — same bandwidth as a 4090, actually. That’s why the Ti pricing catching up to the 4090 used to not make sense, but now it does since you’re buying bandwidth parity with the 4090 at a fraction of the cost.” - source trust me