forkbomb.eth (@forkbombETH)
작성자가 NVIDIA의 DGX Spark를 곧 소개(시연)할 예정이라고 예고한 트윗입니다. 곧 관련 콘텐츠가 공개될 것이라 알리며 @cline, @OpenAI, @deepseek_ai를 태그해 업계 파트너 및 관심 주체들과의 연관성을 암시합니다. 제품 데모나 성능/응용 사례 공개 가능성이 높아 하드웨어·AI 인프라 관련 관심 포인트입니다.
#DGX
Đang chuẩn bị mua 2× DGX Spark, lo ngại kết nối chỉ 1 cáp 200 Gbps gây băng thông giới hạn so với bộ nhớ thống nhất ~275 Gbps. Thêm cáp thứ hai (dual‑link) có thể thu hẹp khoảng cách. Cáp khuyên dùng: QSFP56 200G (0.5 m) hay QSFP112? Người dùng muốn cổng Ethernet Mellanox để nối thẳng ZFS 7450 Pro. #DGX #AI #InfiniBand #Networking #CôngNghệ #CôngNghệAI
https://www.reddit.com/r/LocalLLaMA/comments/1qquinu/bottlenecked_dgx_spark_by_network/
Bucking
Today at work I got to enjoy another round of “Operation: Donkey Punch”, a term I picked up from my hacking friends. It’s where, after fighting with a device to make a change through standard means and failing repeatedly because the standard means are absurdly dumb or limited by policy or design, you tear it apart, make the change while it’s not looking, and then put it back together.
In this example, my department got a pair of Dell DGX devices — basically a set of ARM cores and a high-end NVIDIA GPU with lots of memory. They run a custom version of Ubuntu and are sold as desktop AI accelerators. They’re about the size of a Mac Mini. Team wants to try running models locally.
Thing is, the Setup Wizard expects you, the customer who just shelled out $5000 each, to be in a SOHO office environment, and it makes bad assumptions that don’t work in the Enterprise. My office has a MITM web filter that decrypts web traffic, sniffs it, and encrypts it with our own CA certificate. Every device that needs web access (which is mostly HTTPS these days) must have this cert installed or the device will trust nothing.
Despite being on the lab wired network, the Setup Wizard kept giving me the prompt to select a WiFi AP; that’s because even with correctly configured Ethernet, when it tries to call home to see if it’s truly on the Internet, it fails the cert trust and falls back to demanding WiFi. We can’t use WiFi in the lab; security policy.
There’s no widget to add a cert. Can’t even login on ssh or an alternate terminal. Completely locked out until Setup Wizard finishes. I tried every way to make it work.
Frustrated, I decided it was time to void the warranty. I opened the case, removed the storage, attached it to my workstation, copied the cert file to the right folder, simulated what update-ca-certificate does to “install” the cert, reinstalled the storage into the DGX, and powered it up. Restarted the Setup Wizard, and at the point where it would’ve asked for WiFi, it went directly to downloading system updates and finishing.
Fist up. Big ol’ punch to the head. Take it, bitch.
#Dell #DGX #DonkeyPunch #hacking #NVIDIAVượt ngoài hỗ trợ của NVIDIA, một lập trình viên đã cụm hóa 3 DGX Sparks bằng cách tự viết plugin NCCL với 1500 dòng code C, đạt tốc độ suy luận phân tán trên 8 GB/s.
#NVIDIA #DGX #HPC #AI #Tech #Programming
#CôngNhệ #TríTuệNhânTạo #LậpTrình
Bisschen late to the party, aber dennoch: Ich freue mich auf die erste echte #OnPrem LLM-Inference mit unserem neuen #NVIDIA #DGX #Spark.
Loreto Parisi (@loretoparisi)
‘4x DGX Spark on @exolabs ?’라는 짧은 트윗으로, Exolabs에서 DGX Spark 장비를 4대 배치하거나 도입을 검토 중임을 시사, AI 인프라 확장·하드웨어 업데이트와 관련된 의미가 있음.
Dell's version of the DGX Spark fixes pain points
https://www.jeffgeerling.com/blog/2025/dells-version-dgx-spark-fixes-pain-points
#HackerNews #Dell #DGX #Spark #AI #Innovation #Technology #PainPoints
Người dùng đã thử nghiệm Spark với mô hình Nemotron3 Nano 30B, đạt tốc độ xử lý batch ấn tượng ~1300 token/giây với 200 yêu cầu đồng thời. Hiệu suất này rất hứa hẹn so với thế hệ trước và B200. Bạn nghĩ sao về việc so sánh với cấu hình 4x 3090?
#AI #HieuNang #XuLyBatch #DGX #Spark #Nemotron3 #GPU #Performance #BatchProcessing
https://www.reddit.com/r/LocalLLaMA/comments/1ptp8lq/dgx_spark_and_batch_processing/
Just in case anyone out there is interested, the #dgx_spark does about 12min for the top 2bil passwords on an MD5 crypt hash. Sure that's not what it's meant for but come on...
#hashcat
#hashcat7
#dgxspark
#dgxsparkgb10
#dgx
Để mở rộng quy mô trên 2 DGX Sparks trong một cluster, bạn có thể tận dụng dual 100Gbit QSFP28 links và cấu hình ROCE v2 với layer 3 links. Điều chỉnh NCCL variables để sử dụng ROCE v2 và kết nối cả hai cổng CX7 với switch để tăng băng thông.
#LocalLLaMA #NVIDIA #DGX #clusters #AI #vietnam
https://www.reddit.com/r/LocalLLaMA/comments/1oieip0/theoretically_scaling_beyond_2_dgx_sparks_in_a/
Nvidia DGX Spark: When Benchmark Numbers Meet Production Reality
https://publish.obsidian.md/aixplore/Practical+Applications/dgx-lab-benchmarks-vs-reality-day-4
#HackerNews #Nvidia #DGX #Spark #Benchmark #Production #Reality #AI #Technology #HPC
NVIDIA DGX Spark vs Mac Studio vs RTX-4080: Ollama Performance Comparison:
https://www.glukhov.org/post/2025/10/dgx-spark-vs-mac-studio-vs-rtx4080/
#ollama #nvidia #dgx #dgxspark #performance #devops #mac #macstudio #llm #gpt
I don't see any reason to own an NVIDIA #DGXSpark except for one:
Already being part of the NVIDIA ecosystem.
It's far too expensive. A #MacBook M4 Max is a serial contender, and #StrixHalo perform almost the same.
Consumers are not the target, but businesses that need to prototype their models before hitting the Cloud with 1:1 NVIDIA software stack.
https://www.youtube.com/watch?v=Pww8rIzr1pg
#NVIDIA #LLM #LMStudio #RTX #DGX #Apple #MTX #CUDA #AI #ArtificialIntelligence #Linux
黃仁勳親手送給馬斯克 全球最小 AI 超級電腦 DGX Spark 冀推動 AI 普及
NVIDIA 行政總裁黃仁勳於 2025 年 10 月 13 日，親自將首批全球最小人工智能超級電腦 DGX […]
#人工智能 #AI電腦 #DGX Spark #NVIDIA
https://unwire.hk/2025/10/15/nvidia-dgx-spark-ai/ai/?utm_source=rss&utm_medium=rss&utm_campaign=nvidia-dgx-spark-ai
Nvidia DGX Spark: great hardware, early days for the ecosystem
https://simonwillison.net/2025/Oct/14/nvidia-dgx-spark/
#HackerNews #Nvidia #DGX #Spark #hardware #ecosystem #AI #computing #technology
NVIDIA DGX Spark: A Box of AI Chocolate
Looks pretty, although the fact that it cannot beat a 5090 in inference should tip you off that it’s designed to hold a lot of RAM, not necessarily be on the bleeding edge of perfo(...)
🌗 NVIDIA DGX Spark 深度評測：本地 AI 推理的新標竿
➤ 簡約設計蘊藏強悍 AI 運算力，為本地模型開發與研究開創全新可能
✤ https://lmsys.org/blog/2025-10-13-nvidia-dgx-spark/
本文深入評測 NVIDIA DGX Spark，這是一款將超級運算級效能濃縮於桌上型工作站的創新型 AI 系統。評測者讚賞其精緻的工業設計與強大的 GB10 Grace Blackwell 核心，特別是整合了 128GB 統一記憶體，這使得大型模型無需頻繁進行 CPU 與 GPU 間的數據傳輸，極大簡化了處理流程，尤其適合原型開發與實驗。儘管其記憶體頻寬相較於獨立 GPU 系統有所限制，導致原始效能略遜一籌，但 DGX Spark 在運行較小模型及結合批次處理時展現出優異的表現。評測者亦探討了 SGLang 框架在 DGX Spark 上的應用，包含預處理與解碼分離（PD）及專家並行（EP）技術，並實測了其與 Ollama 框架的效能，同時驗證了預測
#AI #硬體評測 #NVIDIA #DGX Spark #本地推理 #SGLang
Client Info
Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst