#rdma
[ditching language, switching to RDMA mode]
bsky.app/profile/did:...
i've been saying, it's time to standardize a UTF-f32×N-sparse-quantized encoding. why push everything through the narrow, lossy bottleneck of an alphabet when neural activation tells the whole story!
We’re getting close to “that email could have been a single vector”
September 24, 2026 at 8:04 PM
But rdma tho
September 24, 2026 at 8:12 PM
I'm testing 1.5 TB of VRAM on a cluster of Mac Studios, courtesy of RDMA support in macOS 26.2, and Exo Labs' new version of Exo (released under the OSS Apache 2.0 license today).
December 18, 2025 at 2:03 PM
Okay, but even acknowledging that, you can see that the original Norman clearly has red/brown hair with dark stripes representing shadows caused by waves, whereas '90s Norman doesn't even have cornrows really, he has black hair with red stripes dyed into it

bsky.app/profile/rdma...
It's a haircut called "The Chester."

What was clearly meant to be regarded as dark, shiny hair was reinterpreted as "cornrows."
August 6, 2025 at 5:32 AM
With the last commit into DwarfStar now you can use DeepSeek v4.1 Flash in a single DGX Spark as well, with SSD streaming. Around 9 t/s generation. It works also dual-spark RDMA at ~22 t/s.
September 13, 2026 at 7:58 PM
if RDMA-over-ethernet wasn't cursed enough for you, we're now introducing RDMA-over-ELRS
May 9, 2026 at 5:31 PM
Sheesh!
1.5 TB of VRAM on Mac Studio - RDMA over Thunderbolt 5 | Jeff Geerling
www.jeffgeerling.com
December 19, 2025 at 7:50 PM
Just another example of something like this:

bsky.app/profile/rdma...
One of the most ridiculous things I've ever read which sadly sums up the mindsets of a lot of online "fans" these days was when I saw someone make the argument that Barry Allen was racist because he "was a cop" and just because we didn't see him racial profile people, doesn't mean he didn't do so.
May 11, 2025 at 6:30 PM
Big news for DwarfStar users: I got DeepSeek v4 Flash and GLM 5.2 working with Tensor Parallelism across 2 M5Max 128GB MacBooks via RDMA. It is especially interesting for GLM since otherwise, fully resident, can't fit a machine that money today can buy... Now it can.
July 7, 2026 at 3:32 PM
He was right.
August 15, 2026 at 5:55 AM
RDMA over avian carrier
May 9, 2026 at 5:33 PM
A month since USAID, they reflect on lessons beyond titles. At the RDMA farewell, they became the “Art Director,” turning limited resources into joy. Proof that creativity, planning, and teamwork endure—skills can travel further than roles. #USAIDForever #RDMA
September 19, 2025 at 3:20 AM
RDMAこんな使い方あるんだ。データセンター事業者くらいしか使わんのかと

developer.apple.com/jp/documenta...
TN3205: RDMA over Thunderboltによる低レイテンシ通信 | Apple Developer Documentation
Thunderbolt上のRDMAを使用して、Macコンピュータのクラスタ間で低遅延の通信を可能にする方法について詳しく説明します。
developer.apple.com
September 24, 2026 at 11:38 AM
数ヶ月前に社内の所属チーム勉強会で RDMA,Infiniband, RoCE の基礎を解説した資料を置いた。
身内向けでとっ散らかっていて汚いスライドなので後できれいに加筆・修正して上げ直す予定。
speakerdeck.com/markunet/roc...
HPCネットワーク基礎(RDMA/Infiniband/RoCE編)
speakerdeck.com
August 25, 2023 at 2:17 PM
I just pushed a new version of DwarfStar that is likely to break stuff, too many new things: 1. micro-batching. 2. CUDA multi device support. 3. GLM5.2 4. Metal tensor parallel execution (via RDMA). 4. DSpark speculative execution. 5. Many other things probably.
July 20, 2026 at 11:10 AM
RDMA girlfriend, it's like she's right there with you
October 29, 2025 at 10:23 PM
Nice, online redo to nvme over rdma SAN and commit time dropped to 0,2ms.
August 12, 2025 at 4:33 PM
"Unlocking High-Performance AI/ML in Kubernetes with DRANet and RDMA" opensource.googleblog.com/2025/07/unlo...
Unlocking High-Performance AI/ML in Kubernetes with DRANet and RDMA
Introducing DraNet, a native and declarative solution for high-performance networking in K8s, now in beta, simplifying RDMA and interface tuning
opensource.googleblog.com
July 15, 2025 at 7:42 PM
You are the best worst influence.

bsky.app/profile/rdma...
Blame Steve Shives for getting me addicted to this.
May 30, 2026 at 2:19 AM