MTS@Geometric - AI auto-optimizing GPU kernels Co-creator with @julian_gollop of Chip 'n Clawz Prev Silent Hill 2, Brink, Heavenly Sword, Phoenix Point, Apple
Sofia, Bulgaria
Joined March 2007
- Tweets9.9K
- Following2K
- Followers1.6K
- Likes5.7K
Pinned Tweet
I’ve been building in this space here:
github.com/DeanoC
Current project:
github.com/DeanoC/Spiderweb - My distributed hosted OS for any AI agent
github.com/DeanoC/SpiderMonk… - Custom agent specialized to Spiderweb
(demos / videos coming shortly)
AI agentic have made forks so much easier, its basically trivial to keep your fork upto date with the upstream, which was always the biggest concern.
nextpnr/yosys don't allow AI PRs so I just keep a fork until they change there mind. Its totally there choice and doesn't affect me in anyway now!
Deano Calver retweeted
Archrebel Tactics Demo Out Now!
#Archrebel #IndieGame #LaserSquad #RebelStar #XCOM #IronMan #MicroProse @UBadler
Deano Calver retweeted
Back in the late 90's we used to send our changed source files to the "Keeper of the Build/Merge Guy" via email, and we shipped a commercial game for Mattel that way.
Deano Calver retweeted
DeepSeek V4.1 Flash 1017 came out yesterday and it already runs on Lucebox. Day zero support for high quality quants.
huggingface.co/Lucebox/DeepS…
Our DS4.1 1017 quantizations are already available on day-0, high quality and they can be used on 128GB strix/gorgon halos with only 6GB offloading.
Each Lucebox user will have optimized quants in the machine from tomorrow 🚀
huggingface.co/Lucebox/DeepS…
Even my token usage @bot thinks the new X subscription is bad for me! (SuperGrok Max with free Cursor Ultra)
The new plans, called The X Subscription (launched Oct 12), put Cursor, Grok chat and Grok Bot into one shared weekly pool. Claude, GPT and other third-party models stop being included and get billed at API price from prepaid credit. Your legacy Cursor Ultra and SuperGrok Heavy carry on unchanged unless you switch. Switching is permanent, and no deadline has been published. If you swapped both for X Ultra at $200, you'd very likely end up with less total capacity. That's my inference, since the pool sizes aren't public. The only published figure is "at least 2x" the Cursor-model coding usage of old Ultra, and that now has to cover Grok and Bot as well.
Deano Calver retweeted
omg this sh*t is ableist as hell
Just wait until nobody cares to send patches to your stupid mailing lists. In a few years, it will cease to be a way to prove a person's competence. Everyone will have their own kernel fork instead. You guys can just entertain yourselves.
Deano Calver retweeted
1/ Today, we're excited to launch PFlash V2.
If you ask questions over long documents, retrieved passages or a whole codebase on one GPU, reading the prompt can be most of the wait before the first token.
PFlash sits in front of the model and decides which parts of a long prompt the model actually reads. On six public long-context sets it answers 92 of 124 questions against full prefill's 94, while the model reads a median 27% of the prompt, and on one GPU the first token arrives 3.7× sooner on average.
Deano Calver retweeted
From tabletop to desktop! The complete 1.0 digital adaptation of FROSTHAVEN is OUT NOW on PC ❄️
Now you can get the full experience of the acclaimed board game from Isaac Childres and Cephalofair Games, brought to life by Snapshot Games on Steam & Epic Games Store.
Get it today: playfrosthaven.com/
Deano Calver retweeted
Smaller, Smarter, Scalable.
Meet #MaixStation, From 27B to 552B, From 3" SoM to 10" miniRack Cluster, Beats DGX Spark.
🎟️Reserve now: sipeed.com/maixstation
Deano Calver retweeted
If you enjoyed Hob's Barrow please consider posting a very quick little review on Steam, they really help to keep us afloat! 🙏💜 Incredibly much appreciated.
New project news soon!
store.steampowered.com/app/1…
Deano Calver retweeted
PETAFAB: Teaser trailer is out!
youtube.com/watch?v=NDLACyme…
Music by Ari Pulkkinen (Resogun, Nex Machina, Super Stardust HD).
Video editing by Timo Vasara.
DE10-nano FPGA (Mister), Atari ST BIG Demo using my yosys/nextpnr/mistral fork (used AI so can't be upstreamed at the moment) from scratch (except the 68K though has been modified).
CPU/Video/Audio/IO/memory cycle accurate
ALT Atari ST The B.I.G. Demo menu perfect visuals on FPGA via OSS toolchain
Deano Calver retweeted
Over three decades and counting... and still one of the best of its kind.
UFO: Enemy Unknown, also known as X-COM: UFO Defense in the US, was one of the best games of the 90s. MicroProse published this sci-fi strategy classic in 1994, and it became an absolute cult hit. The music and sound effects still give me nerd chills - this game simply oozes greatness.
You’re in charge of the defense and research organization Extraterrestrial Combat (X-COM). You start with basic tech and weapons but gradually learn more about the alien invasion and their technology.
Virtually all reviews gave it high 90% ratings, and Computer Gaming World called it the #1 sleeper hit of all time. It flew off the shelves with massive sales and is widely regarded as a cornerstone of strategy gaming.
Some games feel truly timeless. This is one of them.
I got Codex/Grok to produce a from scratch RISC-V to test my OSS NextPnr/Mistral work.
fes_rv32_cpu
RV32I + Zicsr + Zicntr, multicycle, barrel shifter, 3 interrupts
COMB 1,887 FF 442 M10K 2 Routed fmax 79.7 MHz
Not a bad start IMHO (20-27 MIPS), bit behind PicoRV32 so far
I wonder if Beowolf cluster should make a come back to distribute GPUs across limited PCIe slots per box.
Couple of fast network ports (100Gb/s). Memory distributed as well...
Deano Calver retweeted
Now we support Deepseek V4.1 Flash! Almost 50 tok/s for coding with speculative decoding.
R9700 makes the Gorgon up to 3x faster 🏎️
One of the main pieces of feedback we got on the Lucebox Zero 495 is that it's not clear why it matters to have both unified memory and a discrete GPU in the same machine.
It matters a lot. In the article below, we show how the R9700 gets you 3x the speed on a big model like DeepSeek V4.1 Flash.
The Lucebox Zero 495 runs the model across three tiers of memory: the 32 GB on the Radeon AI PRO R9700, the unified memory of the Ryzen AI MAX+ PRO 495, and the SSD.
Each expert lives wherever its usage justifies it. Compared with the same chip without the R9700, a long prompt gets its answer about 3x sooner, and code generation reaches up to 49 tok/s.
Amazing work from: @davideciffa @DeanoC @GeometricAGI @luceboxai
Fantastic results here for @luceboxai lucebox.com/blog/deepseek-v4…
My work at @GeometricAGI has been leading the quantizer that is able to take a model and given the target model and machine optimize and condition the weights for it.
Our evolutionary kernel optimizer than produces kernel code designed for those layers.
