LinkedInยทThursday, 6 August 2026ยท20d ago
๐ก๐ฉ๐๐๐๐ is amazing, but their biggest ๐ฎ๐ฑ๐๐ฎ๐ป๐๐ฎ๐ด๐ฒ isn't just hardware, ๐ถ๐'๐ ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐๐จ๐๐. It's the reasonโฆ
Xavier Geerinck
CTO @ Scrydon | Sovereign Data & AI Layer | Agentic & Ontologies | ๐ก๏ธ
๐ก๐ฉ๐๐๐๐ is amazing, but their biggest ๐ฎ๐ฑ๐๐ฎ๐ป๐๐ฎ๐ด๐ฒ isn't just hardware, ๐ถ๐'๐ ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐๐จ๐๐. It's the reason people are "๐น๐ผ๐ฐ๐ธ๐ฒ๐ฑ ๐ถ๐ป", what if there was a way to get around this? - Enter "๐ง๐๐ฅ๐
"
๐ง๐๐ฅ๐
(๐๐ฑ๐ข๐ค๐ฉ๐ฆ ๐๐๐โ๐ด ๐ฐ๐ฑ๐ฆ๐ฏ-๐ด๐ฐ๐ถ๐ณ๐ค๐ฆ, ๐ฎ๐ถ๐ญ๐ต๐ช-๐ท๐ฆ๐ฏ๐ฅ๐ฐ๐ณ ๐ค๐ฐ๐ฎ๐ฑ๐ช๐ญ๐ฆ๐ณ ๐ด๐ต๐ข๐ค๐ฌ) is built to break that lock-in and make custom ML kernels target hardware beyond CUDA cleanly.
โข Combined thread-level control + Triton-style tile ops
โข Native performance on Blackwell (B200)
โข Works across vendors without massive compiler bloat
If you like digging into high-performance ML compilers, this oneโs huge.
๐ Blog: https://lnkd.in/eNjg3uTs
๐ฅ GPU Mode Talk: https://lnkd.in/efxMPN92
๐ฟ Benchmark Leaderboard: https://lnkd.in/ew2CX5j4
๐ฌ 1
View on LinkedIn Cross-referenced
Related on the wire
AI this, AI that, but now finally proven in production! It resulted in -60% Memory for us and -50% CPU! How? I've been reading andโฆ
What if you could OWN complex AI problem solving on your own premises / airgapped / hybrid cloud environment? That's what we askedโฆ
A ๐๐ฎ๐ป๐ฑ๐ฏ๐ผ๐
๐๐ต๐ฎ๐ ๐ผ๐ป๐น๐ ๐ต๐ผ๐น๐ฑ๐ ๐๐ต๐ฒ๐ป ๐๐ต๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น ๐ฏ๐ฒ๐ต๐ฎ๐๐ฒ๐ isnโt a sandboxโitโs a request. We watched frontierโฆ
Big move in the AI agent space: ๐จ๐ฆ ๐ฃ๐ฎ๐๐ฒ๐ป๐ ๐ญ๐ฎ,๐ฒ๐ณ๐ฌ,๐ฌ๐ฐ๐ฑ ๐๐ญ ๐๐ฎ๐ ๐ผ๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฎ๐น๐น๐ ๐ฝ๐๐ฏ๐น๐ถ๐๐ต๐ฒ๐ฑ. The patentโฆ
Is Agentic AI in production just a fancy game of roulette? ๐ฒ Right now, everyone is eager to deploy ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐๐ ๐๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐๐โฆ
A crucial reality check from the FLI report! ๐ ๐ถ๐๐๐ฟ๐ฎ๐น'๐ ๐ฟ๐ฎ๐ป๐ธ๐ถ๐ป๐ด ๐๐ต๐ผ๐๐ ๐๐ต๐ฎ๐ ๐ฟ๐ฒ๐ด๐๐น๐ฎ๐๐ถ๐ผ๐ป ๐ฎ๐น๐ผ๐ป๐ฒโฆ