Identify the B200 and GB200 NVL72 in NVIDIA's Blackwell lineup.

Flashcard·Easy·4.0 · 0·~30s·Asked atBrowserbase LangChain Wipro·Relevant atNVIDIA

Attempt it

Your answer

The **B200** is NVIDIA's Blackwell-generation datacenter GPU, the successor to the Hopper H100 and H200. Released in 2024 and shipping in volume through 2025-2026, the B200 has 192 GB of HBM3e memory at roughly 8 TB/s of bandwidth, and its tensor cores support a wider range of low-precision formats than Hopper: native FP8 (carried over from Hopper) plus native FP4. The **GB200 NVL72** is a rack-scale system, not a single GPU. It packages 72 B200 GPUs into one rack, all connected by an NVLink switch fabric so any GPU can address any other GPU's HBM at NVLink bandwidth (much faster than networking between separate servers). The 72 GPUs together expose more than 13 TB of HBM3e on a single NVLink domain. Why this matters for 2026 inference. **The B200's FP4 path** halves bytes per weight again over FP8, which compounds quantization wins for both bandwidth and compute. **The GB200's shared fabric** makes it practical to serve models whose weights and KV cache exceed any single GPU's memory: large mixture of experts models like Llama 4 Maverick, very long context reasoning workloads, and multi-stage agent pipelines where tight all-to-all bandwidth would otherwise bottleneck. In 2026 the GB200 NVL72 is the de facto frontier inference platform. Anthropic, OpenAI, Google, and DeepSeek all use it (or comparable racks) for their highest-tier serving, and open-source stacks like vLLM, SGLang, and TensorRT-LLM have explicit support for the GB200 fabric model.

The **B200** is NVIDIA's Blackwell-generation datacenter GPU, the direct successor to H100 and H200. It ships with **192 GB of HBM3e memory** at roughly 8 TB/s of memory bandwidth, and its tensor cores support **native FP4 and FP8** in addition to BF16 and INT8. FP4 is the Blackwell-specific addition, doubling the low-precision compute path over Hopper's FP8 ceiling. The **GB200 NVL72** is a rack-scale system. It bundles 72 B200 GPUs (paired with 36 Grace CPUs) into a single chassis and links every GPU into one NVLink switch fabric. The fabric exposes a flat memory-addressable domain: any GPU can read any other GPU's HBM at NVLink bandwidth (roughly 1.8 TB/s per GPU on Blackwell's NVLink 5). Total HBM in the domain is over 13 TB at multi-PB/s aggregate bandwidth. Why this matters in 2026: **Single-GPU wins.** B200's 192 GB HBM3e is large enough to fit most frontier dense models in FP8 without tensor parallelism, and the FP4 path doubles compute for prefill-heavy workloads. FP4 KV cache on B200 cuts decode-time bandwidth pressure further compared to FP8 on Hopper. **Rack-scale wins.** Large mixture of experts models (Llama 4 Maverick, DeepSeek V4, Qwen 3 family at the largest tiers) split experts across many GPUs and rely on fast all-to-all communication between experts every token. On separate servers connected by Ethernet or InfiniBand, that all-to-all is the bottleneck. On NVL72's shared NVLink, it runs at HBM-like speeds. The same logic applies to extremely long context windows where KV cache exceeds any single GPU's memory and needs to be partitioned across the fabric. In production, the GB200 NVL72 has become the reference platform for frontier serving in 2026. The major labs use it (or their hyperscaler equivalents) for their top-tier model deployments, and the open-source serving stacks support its fabric model explicitly.

TL;DR

B200 is NVIDIA's Blackwell-generation single GPU: 192 GB HBM3e and native FP4 plus FP8 tensor cores.

Memory aid

Easy to grasp

Think about scaling up a kitchen. A bigger oven (the B200) holds more food, cooks at lower temperatures the older oven could not handle, and finishes faster. But the kitchen still has space limits. A rack-scale system like GB200 NVL72 is more like converting a whole floor into one shared kitchen with 72 connected ovens and one giant pantry everyone can reach. Now you can prepare a banquet that no single oven could have managed alone, because every cook has direct access to every ingredient. For language model serving, the bigger oven is helpful, but the connected-kitchen layout is what makes serving a trillion-parameter mixture of experts or a multi-rack reasoning workload practical without sending data over slow connections.

Key concepts

Inference hardware basics NVIDIA B200 NVLink fabric

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

B200 and GB200 NVL72 are two distinct hardware concepts that often get conflated. The B200 is a single GPU, the Blackwell-generation successor to H200. The GB200 NVL72 is a rack-scale system that packages 72 B200s into one tightly-linked NVLink domain. Both matter for 2026 frontier inference, but they matter for different reasons and they unlock different workloads.

This question separates candidates who can name 2026 frontier hardware from candidates who can plan capacity for it. The discriminators are precise: Blackwell generation, FP4 added over Hopper's FP8, 192 GB HBM3e at the GPU level, 72-way shared NVLink fabric at the rack level. A serving engineer who knows those four facts can talk about MoE inference, long-context decode, and capacity economics with confidence.

This deep dive walks the B200's specs and what FP4 changes operationally, the NVL72's fabric model and what shared addressability unlocks, the workloads that genuinely need each, and the cases where the cost is wasted on smaller workloads that would run on Hopper.

B200 as a single GPU

B200 is the flagship implementation of the Blackwell architecture, the generation after Hopper. The form factor is SXM5 with a dual-die package: two GPU dies sharing the same coherent memory space, presented to software as one device. Total HBM3e is 192 GB at roughly 8 TB/s of memory bandwidth, a step up from H100's 80 GB at 3 TB/s and H200's 141 GB at 4.8 TB/s.

The headline tensor-core feature is FP4. Blackwell adds an FP4 path on the tensor cores that doubles compute throughput over FP8 at the same bandwidth. The transformer engine on Blackwell schedules precision dynamically: a workload can run weights in FP4 and activations in FP8 with hardware-managed scale tracking, so the slower path activates only where the dynamic range requires it.

FP4 is what makes Blackwell qualitatively different from Hopper for inference. FP4 weights are a quarter of BF16 weight memory and half of FP8 weight memory. FP4 KV cache, where the quality budget allows, halves decode-time bandwidth again versus FP8 KV cache. The compute ceiling on FP4 matmul is roughly double FP8, which lifts the roofline ridge point and gives prefill workloads more compute headroom before they saturate. The operational result is that a B200 serves more frontier traffic per dollar and per watt than an H100, by a margin that compounds at scale.

GB200 NVL72 as a rack-scale fabric

Workloads NVL72 unlocks

When NVL72 is the wrong fit

The 2026 cost framing

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

Anthropic's frontier serving fleet for Claude Opus 4.7 uses GB200 NVL72-class hardware (or hyperscaler equivalents) for the largest reasoning workloads.
OpenAI's GPT-5.5 serving relies on NVL72-class fabric for routing the largest mixture of experts experts across many GPUs.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow does FP4 on B200 change the roofline picture compared to FP8 on H100?

FP4 doubles peak compute over FP8 at roughly the same HBM bandwidth, so the ridge-point intensity doubles. A workload that just hit the compute roof on FP8 becomes bandwidth-bound again on FP4 unless it also cuts bytes moved with FP4 KV cache or further weight compression.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Treating GB200 NVL72 as just 72 separate B200s. The point is the shared NVLink fabric that lets all 72 GPUs address each other's HBM at NVLink speeds, making the rack behave like one giant GPU.

60 second bullets to scan on the way to the call.

B200's architecture generation and release year
B200's HBM3e capacity and approximate bandwidth

Primary sources. Browse if you want the original framing.