NVIDIA GPU Clusters: H100, H200, B200, GB200 | Together AI (https://www.together.ai/gpu-clusters) citeturn128view3 [wordlim: 200] Crawled: today; Content type: text/html; Source: open({"ref_id":"https://www.together.ai/gpu-clusters","lineno":null}); Total lines: 540 L0: 【0†iframe†www.googletagmanager.com】 L1: L2: 【1†Image†cdn.prod.website-files.com】【2†Introducing Polymorphic Inference, starting with Together Compiler →】 L3: L4: 【1†Image†cdn.prod.website-files.com】【3†Introducing Together Post-Training: Better models, faster runs, lower costs →】 L5: L6: 【1†Image†cdn.prod.website-files.com】【4†Introducing Together Link: Run open models inside the harness you already use →】 L7: L8: 【1†Image†cdn.prod.website-files.com】【5†The Open Source AI Stack: models, inference, routers, harnesses & tools →】 L9: 【1†Image†cdn.prod.website-files.com】【6†GLM-5.3 vs. GLM 5.3 Flash on DeepSWE →】 L10: L11: 【1†Image†cdn.prod.website-files.com】【7†⚡ On-demand B200s now available on Together GPU Clusters →†api.together.ai】 L12: L13: * 【8†Inference】 L14: L15: * * * * * 【9†Dedicated Container Inference Inference for custom models】 L16: * 【2†New Polymorphic Inference Inference that gets better over time】 L17: L18: Explore Top Open-Source Models L19: L20: * * * * * 【10†Show all】 L21: L22: * 【8†Compute】 L23: L24: * 【8†Model Shaping】 L25: * 【8†Research】 L26: L27: * 【8†Developers】 L28: L29: * 【11†Customers】 L30: * 【12†Pricing】 L31: L32: * * * * * * 【9†Dedicated Container Inference Inference for custom models】 L33: * 【2†New Polymorphic Inference Inference that gets better over time】 L34: L35: Explore Top Open-Source Models L36: L37: * * * * * 【10†Show all】 L38: L39: * * * * 【13†Contact sales】 L40: L41: 【13†Contact sales】 L42: L43: 【14†Sign in†api.together.ai】 L44: L45: Together GPU Clusters L46: # Reliable self-serve, AI-ready GPU clusters at scale L47: L48: Go from zero to production in minutes. Bare-metal performance, InfiniBand networking, and managed orchestration — with flexible pricing for both on-demand and reserved capacity. L49: L50: 【7†Create cluster†api.together.ai】【15†Reserve capacity†api.together.ai】 L51: L52: 【16†Image†cdn.prod.website-files.com】 L53: ## Why Together GPU Clusters L54: L55: Infrastructure that keeps long-running jobs on track — with automated recovery, elastic scale, and zero DevOps overhead. L56: L57: Research & experimentation L58: L59: Spin up a cluster in minutes, test your hypothesis, shut it down when you're done. On-demand pricing means scratchpad experiments don't carry the cost of production workloads — so your team can move fast without burning budget. L60: L61: Distributed training L62: Reserve  from 8 to 4,000+ GPUs without rethinking your architecture. Our InfiniBand interconnect keeps gradient synchronization fast and communication overhead low — so your training runs finish faster, not just bigger. L63: L64: Inference & serving L65: L66: Go from trained model to production endpoint without switching providers. Low-latency networking, Kubernetes-native deployment, and flexible ingress configuration make it easy to serve at scale — and scale back down when traffic drops. L67: ## Everything you need to train at scale L68: L69: Managed infrastructure with built-in observability, orchestration flexibility, and research-grade performance. L70: L71: * * 【17†Image†cdn.prod.website-files.com】 L72: L73: Sustained performance L74: L75: 【18†Image†cdn.prod.website-files.com】 L76: L77: Multi-week stability L78: L79: 【18†Image†cdn.prod.website-files.com】 L80: L81: Reduced stragglers L82: L83: 【18†Image†cdn.prod.website-files.com】 L84: L85: Predictable latency L86: Maintain high utilization across multi-week training runs and model serving. Kernel, hardware, and storage acceleration reduce stragglers and keep latencies predictable. L87: L88: 【19†Explore the docs†docs.together.ai】 L89: L90: * 【20†Image†cdn.prod.website-files.com】 L91: L92: Resilient infrastructure L93: L94: 【18†Image†cdn.prod.website-files.com】 L95: L96: Acceptance testing L97: L98: 【18†Image†cdn.prod.website-files.com】 L99: L100: Automated remediation L101: L102: 【18†Image†cdn.prod.website-files.com】 L103: L104: Node repair L105: Scale compute and storage seamlessly from experiments to production over high-speed InfiniBand. Keep capacity online automatically using continuous health checks, automated remediation, and self-serve node repair. L106: L107: 【21†Explore the docs†docs.together.ai】 L108: L109: * 【22†Image†cdn.prod.website-files.com】 L110: L111: Observability & monitoring L112: L113: 【18†Image†cdn.prod.website-files.com】 L114: L115: Pre-built dashboards L116: L117: 【18†Image†cdn.prod.website-files.com】 L118: L119: Full-stack metrics L120: L121: 【18†Image†cdn.prod.website-files.com】 L122: L123: Real-time alerts L124: Monitor workloads instantly using pre-built Grafana dashboards and full-stack metrics across GPUs, storage, networking, and Kubernetes. Gain complete system visibility without writing custom instrumentation. L125: L126: 【23†Explore the docs†docs.together.ai】 L127: L128: * 【24†Image†cdn.prod.website-files.com】 L129: L130: Self-serve developer experience L131: L132: 【18†Image†cdn.prod.website-files.com】 L133: L134: CUDA version choice L135: L136: 【18†Image†cdn.prod.website-files.com】 L137: L138: Project-level RBAC L139: L140: 【18†Image†cdn.prod.website-files.com】 L141: L142: Multi-tool access L143: Provision clusters instantly with pre-configured tooling alongside selectable drivers and CUDA versions. Manage cross-team access via project-level RBAC using the CLI, SDK, API, Terraform, or the web console. L144: L145: 【25†Explore the docs†docs.together.ai】 L146: L147: 【26†Image†cdn.prod.website-files.com】 L148: L149: 【27†Image†cdn.prod.website-files.com】 L150: L151: 【28†Image†cdn.prod.website-files.com】 L152: L153: 【29†Image†cdn.prod.website-files.com】 L154: ## Frontier research-powered training performance L155: L156: The Together Kernel Collection, built by our Chief Scientist Tri Dao (creator of FlashAttention), delivers improved training and inference performance. L157: L158: * 【30†Image†cdn.prod.website-files.com】【31†Image†cdn.prod.website-files.com】 L159: L160: Together Kernel Collection L161: L162: * 【32†Image†cdn.prod.website-files.com】【33†Image†cdn.prod.website-files.com】 L163: L164: ThunderKittens L165: L166: * AI Training Performance: NVIDIA Hopper to Blackwell, with TKC L167: 【34†Image†cdn.prod.website-files.com】 L168: L169: 【35†Image†cdn.prod.website-files.com】 L170: L171: TKC vs SOTA Approaches L172: L173: 90% faster training L174: L175: Training a 70B parameter Llama-architecture model (BF16) with an optimized TorchTitan + Together Kernel Collection (TKC) reached 15,264 tokens/second/GPU on NVIDIA HGX B200, up from 8,080 tokens/second on NVIDIA HGX H100—a 90% jump in training speed. L176: L177: 【36†learn more】 L178: L179: * FP8 GEMM Performance (M x N x K) L180: L181: * ThunderKittens B200 L182: L183: * cuBLAS H100 L184: L185: * cuBLASB200 L186: 【37†Image†cdn.prod.website-files.com】 L187: L188: 【38†Image†cdn.prod.website-files.com】 L189: L190: ThunderKittens vs cuBLAS L191: L192: ~2× faster L193: L194: ThunderKittens’ FP8 kernel for NVIDIA HGX B200 matches NVIDIA cuBLAS GEMM performance while delivering ~2× speedup over H100 FP8 GEMMs, leveraging Blackwell’s Tensor Core–accelerated matrix operations. L195: L196: 【39†learn more】 L197: ## Flexible pricing models L198: L199: Both options are fully self-serve. Choose based on your capacity requirements and commitment level. L200: L201: * On-Demand L202: L203: Standard hourly rate L204: L205: Commitment L206: L207: None—pay hourly, terminate anytime L208: L209: Best for L210: L211: Starting with on-demand for flexibility L212: L213: Capacity L214: L215: Based on real-time availability L216: L217: Scale L218: L219: Up to 256 GPUs L220: L221: 【7†Create now†api.together.ai】 L222: L223: * Reserved L224: L225: Lower hourly rate L226: L227: Commitment L228: L229: Up to 6 months, pay upfront L230: L231: Best for L232: L233: Guaranteed access with better economics L234: L235: Capacity L236: Locked in for your duration L237: L238: Scale L239: L240: Up to 4,000+ GPUs L241: L242: 【40†Reserve capacity】 L243: ## Browse other NVIDIA GPUs L244: L245: Self-serve GPUs with transparent billing. L246: L247: H100 L248: L249: Hardware L250: L251: NVIDIA H100 (80 GB) L252: L253: On-demand L254: L255: $3.99/hr per GPU L256: L257: Reserved L258: L259: Starting at $3.45/hr per GPU L260: L261: Scale L262: L263: 8 to 4,000 GPUs L264: L265: 【7†Create cluster†api.together.ai】 L266: L267: H200 L268: L269: Hardware L270: L271: NVIDIA H200 (140 GB) L272: L273: On-demand L274: L275: $5.99/hr per GPU L276: L277: Reserved L278: L279: Starting at $4.15/hr per GPU L280: L281: Scale L282: L283: 8 to 4,000 GPUs L284: L285: 【7†Create cluster†api.together.ai】 L286: L287: HGX B200 L288: L289: Hardware L290: L291: NVIDIA HGX B200 (180 GB) L292: L293: On-demand L294: L295: $9.29/hr per GPU L296: L297: Reserved L298: Starting at $7.99/hr per GPU L299: L300: Scale L301: L302: 8 to 4,000 GPUs L303: L304: 【7†Create cluster†api.together.ai】 L305: L306: HGX B300 L307: L308: Hardware L309: L310: NVIDIA HGX B300 (270 GB) L311: L312: On-demand L313: L314: $11.49/hr per GPU L315: L316: Reserved L317: L318: Starting at $9.89/hr per GPU L319: L320: Scale L321: L322: 8 to 4,000 GPUs L323: L324: 【7†Create cluster†api.together.ai】 L325: L326: GB200 NVL72 L327: L328: Hardware L329: L330: NVIDIA GB200 NVL72 (186 GB) L331: L332: Reserved L333: L334: Contact us for pricing L335: L336: Scale L337: L338: 72 to 4,000 GPUs L339: L340: 【13†Contact sales】 L341: L342: GB300 NVL72 L343: L344: Hardware L345: L346: NVIDIA GB300 NVL72 (288 GB) L347: L348: Reserved L349: L350: Contact us for pricing L351: L352: Scale L353: L354: 72 to 4,000 GPUs