Sabitlenmiş Tweet
MAKE ENGINEERS GREAT AGAIN
"A good engineer gets stale very fast if he doesn't keep his hands dirty."
- Wernher von Braun
English
StrongEngineer_
7.1K posts
StrongEngineer_
@hotschmoe
Christian • Father of 4 • Structural Engineer • e/acc • BTB Jungle Lurker • too many labels
Desert Southwest, USA Katılım Ekim 2021
555 Takip Edilen822 Takipçiler
@then_there_was There's plenty of people already getting payments directly from the government, it's just not us
UBI already exists for the bottom
English
I mean, we’re technically doing that now already. Maybe not always directly, but definitely indirectly.
First Squawk@FirstSquawk
ELON MUSK: AI WILL FORCE GOVERNMENTS TO SEND CASH DIRECTLY TO PEOPLE AS DEFLATION BECOMES THE BIGGEST THREA
English
first 3d printer we got for the office (leapfrog? running pronter or something) was close to $20,000 and it tooks us *weeks* and *dozens* of tries per house model to get printed, each model was another week of debugging to get something that was ugly and need insane post processing before we could use them in HoA design review.
now a $300 ender 3 spanks the pants off that thing
all in time my friends
Loktar 🇺🇸@loktar00
Running a huge model at home like GLM 5.2 Q2 where tk/s is low is a lot like 3D Printing large models. You wake up and you either have a gift waiting for you, or a big mess, it's all about the surprise! 😂 Wish GPUs were as cheap as 3d printers.
English
@_proteuss_ Totally usable in my opinion. Mature W4A8/W8A8 should be much better imo to activate int8 fastpaths
x.com/i/status/20685…
StrongEngineer_@hotschmoe
one intel b70 ($950), first day setup Qwen3.6-27B W4A16 (autoround), *no* MTP 128k context, kv cache fp16 1 session: 28.1 tok/s 2 concurrent sessions: 52.0 tok/s cumulative 4 concurrent sessions: 87.8 tok/s cumulative 64 concurrent sessions: 234.7 tok/s cumulative
English
@hotschmoe interesting - how do the models run on a non cuda card?
English
I almost bought a dgx spark at $2999, then I almost bought a rtx pro 5000 for $3400. Watched as prices continued to march.
Not getting left behind, I decided to throw my hat in the ring with Intel, got 2xB70s for $1900
Has barely been 2 days and I can't believe how much fun this is. Excited to contribute to Intel optimizations and provide useful AI to my family and friends
English
@PaulGugAI Starting with qwen, will expand out and log more!
x.com/i/status/20685…
StrongEngineer_@hotschmoe
one intel b70 ($950), first day setup Qwen3.6-27B W4A16 (autoround), *no* MTP 128k context, kv cache fp16 1 session: 28.1 tok/s 2 concurrent sessions: 52.0 tok/s cumulative 4 concurrent sessions: 87.8 tok/s cumulative 64 concurrent sessions: 234.7 tok/s cumulative
English
@hotschmoe Nice one mate. What models have you been serving on them?
English
Waking up in 6 hours to drive 4 hours. I could sleep but instead I'm on X, the everything app
English
@Xyoskeleton Exactly, the way I view Intel now is how I viewed rocm even a year ago
the bmg cards are underutilized now but they have so much potential to raise the ceiling and I'm hoping to contribute to that
English
@hotschmoe I think Intel is deff undervalued and I’m sure whatever they are working on next is gonna be really good too once they have maturity in this space. It’s how i feel about ROCm with AMD in 1-2 years it’ll be much more competitive
English
@jtlin It's gonna be a fight I'm sure but it's something I enjoy so excited to see where these bmg cards take us
English
@hotschmoe Interested in hearing about your experience with tensor parallel on multiple Intel ARCs and how it evolves. It seems like fantastic price/performance when it's mature.
English
i hate rust
i love zig
pi using qwen3.6-27b with rust runs circles around zig
hmmm
English
@hotschmoe That’s fantastic! I’ve followed it to see how everything unfolds. I’m particularly interested in observing the performance of those Intel cards.
English
@cheeez42 Qwen3.6-27b in w4a16 for now
Gonna try my hand at some custom kernel for custom w4a8 quants to hit int8 fastpaths on the b70 cards then move to mtp
I'll have two cards wired up, then possibly buy two more. My threadripper has slots for 4 dual slot cards so why not at this point
English
@hotschmoe Those intel cards have really been looking interesting. what model are you running?
English
water tastes better when I drink White Monster instead of water
English
with a single card i could hit 256k with kv cache at fp8 but i wouldnt do that myself
i have another card that i need to wire in this weekend, that would take my context well past model limit
128k is the *lowest* acceptable imo. if your paying attention to the agent (i love pi) then I dont hit the limit. but if i used pi with qwen3.6-27b the way I use claude code right now it would blow past the limits
i know pi does compaction and resume pretty well i've read but i havent actually left it alone long enough to do that, so maybe a non-issue?
(i still pay for claude and codex btw)
English
@hotschmoe @MiaAI_lab @NVIDIAAI Do you find 128k context limiting?
Perhaps I would optimize my usage/prompt it to compact messaging, or rely more on obsidian, if I was using local more, but I regularly hit higher context windows especially in heavy coding sessions.
English
I've tested how well Qwen3.6-27B NVFP4 runs on @NVIDIAAI DGX Spark
256k context, kv cache fp16, MTP --> 88gb VRAM
Results:
1 session: 18.62 tok/s
2 concurrent sessions: 38.43 tok/s cumulative
4 concurrent sessions: 73.73 tok/s cumulative
Near-linear scaling.
Full report 👇
English
@DeseretSaint I've been happily surprised by the performance, battery and lightweight-ness of my new zenbook a16
Got a snapdragon x2 elite extreme 48gb unified memory
No joke my best experience on windows in a decade
Definitely something to the unified memory systems
English
@hotschmoe I've already got a nicely spec'd AMD system with some beefy ram I built well before the whole pricing of ram/everything debacle really started, as I was into coding before AI was cool.
Fun to test with, but unified memory is looking really nice outside of the Apple ecosystem.
English
StrongEngineer_ retweetledi
