0xSero (@0xSero)
Lambda의 5,000달러 컴퓨트 크레딧, Nvidia의 H100 8장 클라우드 제공, TNG Technology의 B200 2주 사용 제공 등으로 72시간 만에 10만 달러 이상 가치의 자원을 확보했다고 언급한다. AI 개발자를 위한 대규모 컴퓨트 지원 사례로, 최신 GPU 자원 확보가 큰 가치가 있음을 보여준다.
In 72 hours I got over 100k of value 1. Lambda gave me 5000$ credits in compute 2. Nvidia offered me 8x H100s on the cloud (20$/h) idk for how long but assuming 2 weeks that'd be 5000$~ 3. TNG technology offered me 2 weeks of B200s which is something like 12000$ in compute
Snowflake's Arctic Long Sequence Training: How to Train LLMs on 15 Million Tokens Without Selling a Kidney
#ALST #Snowflake #LongContextTraining #DeepSpeed #HuggingFace #SequenceParallelism #LLMTraining #H100 #Llama8B #Qwen3 #GPUMemoryOptimization
Snowflake AI Research just open-sourced Arctic Long Sequence Training (ALST), a framework that pushes LLM training from a measly 32K tokens to over 15 million — a 469x improvement — using standard Hugging Face models and H100 GPUs. Here's what it means for you.
Andrej Karpathy (@karpathy)
nanochat이 단일 8x H100 노드에서 GPT-2 역량 모델을 약 2시간 만에 학습시켰다고 발표했습니다(한 달 전 약 3시간에서 단축). fp8 지원과 여러 튜닝, 그리고 데이터셋을 FineWeb-edu에서 변경한 것이 주요 개선 포인트로, 실시간 인터랙티브 학습에 한층 근접했다는 기술적 진전입니다.
nanochat now trains GPT-2 capability model in just 2 hours on a single 8XH100 node (down from ~3 hours 1 month ago). Getting a lot closer to ~interactive! A bunch of tuning and features (fp8) went in but the biggest difference was a switch of the dataset from FineWeb-edu to
Mark Gadala-Maria (@markgadala)
신규 비디오 생성 모델 'HELIOS' 공개: 14B 규모의 autoregressive diffusion 모델로, 단일 텍스트 프롬프트로 최대 60초의 일관된 비디오를 생성한다고 발표됨. 성능은 NVIDIA H100 한 장에서 초당 19.5프레임으로 실시간급 처리에 가까워 같은 규모 모델로는 최초 사례로 보임.
🚨 BREAKING: NEW VIDEO MODEL "HELIOS" GENERATES 1 FULL MINUTE OF VIDEO FROM A SINGLE PROMPT >MODEL: 14B autoregressive diffusion model — first of its size to hit real-time >OUTPUT: Up to 60 seconds of coherent video from a single text prompt >SPEED: 19.5 FPS on one NVIDIA H100
As the AI arms race accelerates, the 18-month hardware refresh cycle has transformed GPUs from simple components into high-value infrastructure assets. This article explores why selling hundreds of units—like NVIDIA’s H100 or A100—requires a shift from "peer-to-peer" thinking to "Enterprise ITAD" strategy.
#DataCenter #ITAD #GPU #EnterpriseTech #NVIDIA #TechStrategy #BuySellRam #CircularEconomy #AI #H100 #Blackwell #GPU #TechNews #EnterpriseAI #AssetRecovery
Hello Fediverse! We are ServerMO, providing high-performance infrastructure for 21 years. 🚀
What we offer:
🔹 Elite GPUs: NVIDIA H100 & A100.
🔹 Bare Metal: AMD EPYC & Intel Xeon.
🔹 Network: 1-100Gbps Unmetered + BGP Support.
🔹 Global: 250+ locations across 6 continents.
🔹 Security: Free 250Gbps+ DDoS Protection.
Instant setup in 24-48hrs. Stop renting, start owning!
#ServerMo #DedicatedServers #BareMetal #GPU #H100 #A100 #SysAdmin #DevOps #AI
NVIDIA GPU Cluster Liquidation: Maximize ROI and Asset Recovery
The shift to Blackwell is accelerating the depreciation of NVIDIA A100, H100, and H200 clusters. What were recently frontier training assets are now facing mid-life value cliffs due to performance-per-watt gaps, power density limits, and liquid-cooling requirements.
This turns GPU cluster liquidation into a capital strategy, not just decommissioning. Timing the secondary market, preserving service records to capture refurbished premiums, and enforcing IEEE 2883 data sanitization are key to maximizing ROI and funding next-generation deployments.
In compressed AI refresh cycles, asset recovery speed directly impacts infrastructure competitiveness.
#GPU #AIInfrastructure #DataCenter #AssetRecovery #H100 #A100 #H200 #Blackwell #ITAD #AIHardware #GraphicsCard #VideoCard #HPC #tech
Liquidating NVIDIA A100 H100 H200Learn how to liquidate NVIDIA A100, H100, and H200 GPU clusters to maximize resale value, ensure secure data sanitization, and fund next-generation upgrades efficiently. clusters: maximize resale value, manage depreciation, ensure data sanitization, and fund Blackwell GPU upgrades efficiently.
Python Trending (@pythontrending)
InferenceX라는 오픈소스 연속 추론(continuous inference) 벤치마킹 프로젝트에서 Qwen3.5, DeepSeek, GPTOSS 등 모델을 대상으로 GB200 NVL72, MI355X, B200, GB300 NVL72, H100 등 다양한 추론 하드웨어를 비교하는 벤치마크를 소개하며, 곧 TPUv6e/v7 및 Trainium2/3 지원 예정임을 알립니다.
Kosseila (CloudDude) (@CloudDude_)
DeepSeek 관련 업데이트로 보이는 게시물: DeepSeek v3.2는 643GB 규모의 가중치로, 16개 GPU(2× H100 노드)로 운영되며 KubeRay로 자동화되어 배포되었다고 공유합니다(텐서 병렬 TP=8, 파이프라인 병렬 PP=2). 대규모 모델 운영 스펙을 공개하는 내용입니다.
Как мы готовили Kubernetes под ML-нагрузки: пошаговый гайд (и что пошло не так)
Привет! Я Дмитрий, инженер и руководитель направления MLOps в Совкомбанке. Специализируюсь на разработке и эксплуатации ML-платформ на базе Kubernetes и GPU. С 2010 года в ИТ: строю инфраструктуру для машинного обучения, внедряю Kubeflow и GPU-оператор, настраиваю MIG на H100 в корпоративных средах с повышенными требованиями к безопасности и надежности. В последние годы фокусируюсь на оптимизации ML-пайплайнов, повышении утилизации GPU (включая MIG-профили) и интеграции MLOps-практик в процессы продуктовых команд. В 2022 году в некоторых командах разработки уже существовали проекты с применением ИИ, но как отдельные компоненты, не хватало единой платформы управления. По мере роста количества и сложности бизнес-задач возникла необходимость в создании ML-платформы как сервиса с едиными стандартами авторизации. Мы изучили доступные инструменты, попытались объединить их в одном Kubernetes-кластере, столкнулись с рядом ограничений — и в итоге пришли к архитектуре на базе Kubeflow и GPU-оператора. В статье рассказываем, какие сложности были в ходе проекта, как выстроили работу с Kubeflow, настраивали H100 с MIG-разделением и что важно учесть, если вы планируете строить ML-платформу на bare-metal-GPU в корпоративной среде.
https://habr.com/ru/companies/sovcombank_technologies/articles/994534/
#MLOps #DevOps #Kubernetes #Kubeflow #GPU #NVIDIA #H100 #MIG #baremetal #GPUоператор
sayzard
TechLİfe
BuySellRam.com
ServerMO
ALEXBSR
Habr