Outerport (@outerport) header
Outerport (@outerport) profile picture

Autonomous design for industrial controls (YC S24)

Joined July 2024
Photo by Outerport (@outerport)Photo by Outerport (@outerport)Photo by Outerport (@outerport)Photo by Outerport (@outerport)Photo by Outerport (@outerport)Photo by Outerport (@outerport)
スケルトンゲームボーイ的な筐体の中で、プリント基板上で動くプラントシミュレータのおもちゃを弊社システムで一気通貫でやらせてみた こちらは実際に製作してみて弊社の記念品として用意する予定です・・・
3
30
2,444
弊社ホームページでちょっとしたプラント制御シミュレーションを実際に触って遊べるようになりました😎🏭
3
12
1
96
8,624
TOPPAN様との協業をご紹介いただきました! 世界最強のHuman-in-the-loopのデータ構造化基盤を作っていきます💪
3
5
3
83
22,336
ダイキン工業様との協業を発表しました!設計でのAI活用において、必要なノウハウが詰まっている図形・図表を構造化するOuterportの技術を本格導入していただきました。 CG/CV/AIの技術を総動員して、新たなハードウェア技術の設計・開発・製造がもっと速くできる世界を目指して精進いたします🛠️🏭⚙️
13
19
4
203
45,421
ダイキン工業、AIによる図表・図面の「構造化」プラットフォームとしてOuterportを導入し、AIエージェント活用を推進 prtimes.jp/main/html/rd/p/00…
3
15
2,169
Parsing CAD drawings is one thing but parsing architectural plans takes the craziness up a notch
6
18
3
182
15,359
Yeah document parsing is cool, but what about CAD drawing parsing... (but with documents too 🥺)
34
45
4
601
94,511
Outerport retweeted
My little prediction of 2025: As AI scaling is plateauing (steady amazing progress like microprocessors vs jumps), we'll begin to see more of a focus on products. Instead of a multi-modal architecture, there will be a focus on a multi-model architectures: how do we jointly train and integrate the AI models to work together to maximize perf while updating the models quickly. I would expect to see every major industry get targeted (already happening): AI graphic designer, AI programmer, AI lawyer, AI doctor, AI chip designer, AI FP&A, etc. The AI labs will have to cut burn and finally ask themselves: who is this for? why do they want it?
26
29
13
516
78,572
.@outerport is building the infrastructure base for multi-model architectures; figuring out the technical details like hot-swapping, time multiplexing, live-deployments, kv cache management, and more 😉
My little prediction of 2025: As AI scaling is plateauing (steady amazing progress like microprocessors vs jumps), we'll begin to see more of a focus on products. Instead of a multi-modal architecture, there will be a focus on a multi-model architectures: how do we jointly train and integrate the AI models to work together to maximize perf while updating the models quickly. I would expect to see every major industry get targeted (already happening): AI graphic designer, AI programmer, AI lawyer, AI doctor, AI chip designer, AI FP&A, etc. The AI labs will have to cut burn and finally ask themselves: who is this for? why do they want it?
2
6
2,742
How do you hide latency for loading models from disk? Most implementations do disk -> CPU RAM, and then a CPU RAM -> GPU RAM sequentially. The latency of loading into the GPU is the sum of the two steps. In practice, these 2 steps can actually be interlaced because a model weight is not a blob of memory and is actually a collection of tensors. We can hide the latency of the model load by parallelizing disk->RAM and the RAM->GPU by processing these two tasks in 2 different thread pools that communicate to each other with a channel queue that sends over tensors. One pool of threads will load memory pages and mark named tensors as complete when all pages have loaded, and push the tensor onto the queue. Another thread will take in completed tensors and start the transfer into GPU for that tensor. In a naive implementation this would actually be very slow since each tensor can be small and opening up a CUDA context and kernel for each named tensor would create large amounts of overhead. So, we also need to do buffering so that there is a sufficiently large set of tensors that are marked complete before a batched write to GPU happens. (We experimentally determined 64MB pages are pretty good 🙂) As a result the bottleneck ends up being almost entirely the "disk to RAM" step- because the RAM to GPU is comparatively fast and can be parallelized.
1
2
20
2,386
We think that this is exciting because it will unlock more creativity by allowing for more complicated AI workflows where multiple models need to be loaded in sequence. We have already seen creative technologists do amazing things with AI workflows using tools like ComfyUI.
1
3
305
Check out our blog post for more details! outerport.com/blog/fast-flux…
2
249
Replying to @outerport
0 cold start when? 🚀
1
2
309
Outerport retweeted
👇
This quoted post is unavailable.
1
1
467
One of the interesting graphs is the comparison against CUDA virtual memory- implicit memory management through the device driver always loses in comparison to explicit memory management systems like FlexGen, InfiniGen.
221
弊社(@Outerport)がニューヨークでタイムズスクエアデビューしました!
2
2
1
80
6,865
Outerport retweeted
2️⃣ Developer tools and automation are hot 26% of the AI startups are building scaffolding for the next tech boom, making it faster & easier to build better AI applications. @keetapi: APIs for any website @outerport: VMWare for GPUs @mem0ai: Memory layer for AI agents +48 more
1
1
4
949