Pinned Tweet
Why dual RTX 3090 is a perfect starter pack into local AI
- 48GB VRAM for ~$2000, just drop them into your pc
- Runs Qwen3.6-27B at Q6 with fat context
- CUDA + vLLM/llama.cpp work out of the box, zero headache
- Buy threadripper and scale your rig up to 8x3090
- Prices are structurally rising, supply demand shock
3090 is the people's card
I think we can all agree that Intel is no longer shit
If 2x DGX Sparks can hit 25-tps on Minimax, what is the theoretical limit for 4x Intel B70s?
I hit 93 tps, which is spot on with what ChatGPT suggested was the likely optimized outcome.
Intel is over 3x faster, and significantly cheaper.
I paid $5300 USD for the Intel build.
2
12
1,376
Small tip:
If you using tailscale and run agents on your server, set ACLs so the server can't reach your phone, only the phone reaches the server
If your agent ever gets prompt injected, it can't pivot to your other devices
2
3
223
Update on dual 3090 bottleneck saga
I was wrong about PCIe being the main problem. People in the replies pushed back, I tested more, here's the results:
1) Yes, my second GPU is stuck on PCIe 3.0 x1 because of cheap B760 board. But that's not the only issue
2) Three commenters pushed back saying memory bandwidth is the real bottleneck, not PCIe. One has PCIe 4.0 x8/x8 setup
3) Speed depends on what you generate:
- Short answers ~75 tok/s
- Long code files ~50 tok/s
Even though acceptance rate is higher on long output
TLDR: it's a stack of bottlenecks, PCIe x1 + memory bandwidth + MTP acceptance ceiling
No single upgrade fixes it all
New mobo will definitely help, RTX 6000 Pro helps more 😀
Tested TGP scaling on my dual 3090, 250-350W in 10W steps
Expected the classic playbook, power up > tok/s up, sweet spot somewhere in the middle
But got something different
Setup: Qwen3.6-27B AWQ via vLLM, MTP=4
Results:
250W → 75.14 tok/s
290W → 75.67 tok/s
350W → 75.61 tok/s
Zero change across 100W range 🤯
So I checked clocks under load:
150W → GPU clock 915-1095 MHz, 75 tok/s
350W → GPU clock 1935-1950 MHz, 75 tok/s
Almost 2x the clock speed, same throughput
This connects directly to my earlier PCIe test, second card sits on PCIe gen3 x4 through the B760 chipset, doing reduce at 13% of theoretical bandwidth
If you're running dual GPU on a consumer chipset with x16/x4 split, your power limit might be doing nothing, just down it to 150W, nothing will change
That was the last test on this mobo, time to upgrade my rig
I guess this kind of content is needed too, since a lot of people are just buying their first GPU and slapping it into their PC
Qwen 3.7 is coming, get ready your GPUs bros 😎
Qwen3.7 Preview By @Alibaba_Qwen lands on Arena for Text and Vision.
In Text Arena, Qwen3.7 Max Preview ranks #13 overall.
Alibaba is now the #6 lab in this arena.
- #7 Math
- #9 Expert
- #9 Software & IT
- #10 Coding
In Vision Arena: Qwen3.7 Plus Preview ranks #16 overall, making Alibaba the #5 lab.
Congrats to the @Alibaba_Qwen team on the latest progress!
2
33
3,149
Tested TGP scaling on my dual 3090, 250-350W in 10W steps
Expected the classic playbook, power up > tok/s up, sweet spot somewhere in the middle
But got something different
Setup: Qwen3.6-27B AWQ via vLLM, MTP=4
Results:
250W → 75.14 tok/s
290W → 75.67 tok/s
350W → 75.61 tok/s
Zero change across 100W range 🤯
So I checked clocks under load:
150W → GPU clock 915-1095 MHz, 75 tok/s
350W → GPU clock 1935-1950 MHz, 75 tok/s
Almost 2x the clock speed, same throughput
This connects directly to my earlier PCIe test, second card sits on PCIe gen3 x4 through the B760 chipset, doing reduce at 13% of theoretical bandwidth
If you're running dual GPU on a consumer chipset with x16/x4 split, your power limit might be doing nothing, just down it to 150W, nothing will change
That was the last test on this mobo, time to upgrade my rig
I guess this kind of content is needed too, since a lot of people are just buying their first GPU and slapping it into their PC
Got asked to test if PCIe is bottlenecking my dual 3090 setup
First card on PCIe gen4 x16 (direct CPU), second on gen3 x4 (through B760 chipset)
Ran nccl tests:
- PCIe gen3 x4 theoretical: 3.94 GB/s
- all_reduce bandwidth: 0.52 GB/s 😂
13% of theoretical lol
Second card routes through the chipset, this is why TP=2 gives me the same speed as single GPU
I'm ordering NVLink today
Don't listen to those retards saying card X isn't good enough for serious work, just start with whatever you can afford
I saw a dude running 27B on 3x1080ti for 30 tok/s, these GPUs are from the pharaoh era, and many other amazing examples
Just because you can't buy 6000 Pro today doesn't mean you won't be able to tomorrow
The main thing is just to start somewhere and get some experience, understand how it works, and figure out where to go from there
Even if you buy a 3090 and realize it's not for you, you can always flip it (and highly likely for a profit)
What's the point of buying a DGX Spark for $5k?
Imo to fully unlock the potential you need at least two, that's already $10k
Mac Studio M3 Ultra 256GB for $7.5k looks like the better play. It's faster and all in one box
The only reason is that it just looks cool af
What am I missing?
Hermes + GPT 5.5 is awesome
It's fast, smart, tool calling like a clockwork
This is such a breath of fresh air after 27B
But!
Things are evolving pretty fast, we’ll probably be running an equal model on 100-200GB in six months
3
13
780
10 year old Chinese kid running multiple agents on Mac Studio
We are definitely not trying hard enough lol
“未来的世界属于懂 Token 的人。” —— 出自一位10岁博主之口。
刚换了 Mac Studio 的他,不是为了打游戏,而是为了“养龙虾”(跑多个 AI Agent 协同工作)。他把复杂的 AI 产业链比作一个大蛋糕,从能源层到应用层,层层剖析。
别觉得小孩在过家家,他讲的“Token 是 AI 时代的硬通货”这个观点,可能比很多专家的报告都接近本质。
这届小孩的 AI 认知已经 Next Level 了,建议大人们反复观看,治治我们的“算力焦虑”。👇
2
5
705
"Local or cloud AI" is the wrong question
Cloud models are the smartest minds you can rent by the hour. Use them for what's worth renting a genius for
Local models are the staff you own. Not so great but never rate limited, 24/7 available, never reading your data. You don't need a janitor from Harvard
Using both is the ultimate cheat code
Rent the brains. Own the weights
Stop letting twitter convince you it's one or the other
Recommending a $9k 6000 Pro over a $1k 3090 to a local AI beginner is like recommending a Rolex to someone who just asked what time it is
Hard truth: yes, local hardware is expensive
DGX Spark, 3090, Mac Studio, all that. That's thousands of dollars, and few people limit themselves to just one thing
And a subscription is just $20/mo. $240 a year. Pocket change
And that's the problem
You're not paying $240. You're paying $240 forever
You will not stop using AI in a few years, it's forever with us
And the subscription price will only go up
Securing your own GPUs is a smart move
Hermes agent is the king, OpenClaw is top-2 and already irrelevant for several months
And after that — nothing, total emptiness
We see new models dropping almost daily from different labs, but we're not seeing the same with agents
But why?
1
150




















