← LIVE TERMINALALL EDITIONSRIVALS SECTION

GPU Dispatch · 2026-W28

DISPATCH 2026-W28 · published 9/9/2026 UTC

Blackwell HGX B200 is shipping; H200 HBM3e remains the Hopper workhorse

UPDATED

NVIDIA's datacenter stack splits between shipping Blackwell (B200 on HGX, up to 1.4 TB aggregate HBM3e per 8-GPU board) and Hopper refresh H200 (141 GB HBM3e at 4.8 TB/s). Vera Rubin is publicly ramping but not yet a deployment baseline.

NVDA ecosystem — launch edition (2026-W28)

### What is shipping today

HGX B200 — NVIDIA lists HGX B200 as shipping now: eight Blackwell SXM GPUs on a single baseboard with 1.4 TB total GPU memory, fifth-generation NVLink (1.8 TB/s GPU-to-GPU, 14.4 TB/s total NVLink bandwidth), and FP4/FP8 tensor throughput up to 108 PFLOPS FP4 sparse at the 8-GPU system level per NVIDIA's public HGX specification table.

H200 — The first GPU with HBM3e: 141 GB at 4.8 TB/s memory bandwidth on the Hopper architecture. NVIDIA positions H200 for LLM inference and HPC memory-bandwidth workloads versus original H100 SKUs.

GB200 Grace Blackwell Superchip — Module-level design pairing Grace CPU with Blackwell GPU(s) for rack-scale AI factories; aggregate 384 GB HBM3e per superchip module in our catalog (system integrators ship these as NVL72/NVL36 configurations).

### Supply chain dependencies

Blackwell and Hopper HBM3e accelerators depend on advanced packaging at TSMC and HBM stacks from Samsung, SK Hynix, and Micron — see the live terminal supply-chain graph (F3) for foundry and memory vendor links.

### Forward visibility (sourced, not a deployment recommendation)

NVIDIA's public HGX pages also describe Vera Rubin as ramping into production with Rubin GPUs and sixth-generation NVLink. Treat Rubin as a forward roadmap item until system SKUs and software stacks are generally available in your region.

---

Technical research for informational purposes only. Not financial advice. No investment recommendations. Specs sourced from vendor documentation; verify before engineering or procurement decisions.

ANALYSIS

For inference-heavy LLM serving, H200's 141 GB HBM3e at 4.8 TB/s is often the practical upgrade path from H100 today. B200 matters when you need Blackwell FP4/MX formats and NVLink-5 scale-out in new HGX builds — compare on framework support and power/cooling, not peak FLOPS alone.

[NVIDIA HGX B200 platform specs][NVIDIA H200 GPU][NVIDIA Grace Blackwell Superchip]

AMD MI355X matches Blackwell on HBM3e density; Samsung supplies the stacks; Ascend serves China-domestic AI

UPDATED

AMD Instinct MI355X (288 GB HBM3E, 8 TB/s, CDNA4, 1400W TBP) is the clearest open-ecosystem alternative at similar memory tiers. Samsung HBM3E stacks are a shared supply-chain layer. Huawei Ascend remains the primary China-market accelerator under export-control constraints.

Rivals — launch edition (2026-W28)

### AMD Instinct (direct datacenter GPU competition)

MI355X (launched 2025-06-12 per AMD) — 288 GB HBM3E, 8 TB/s peak memory bandwidth, CDNA4 architecture, 1400W TBP, OAM form factor. AMD documents MXFP6/MXFP4 support and full-chip ECC. Platform brochures describe eight OAMs on UBB 2.0 with 2.3 TB aggregate HBM3E and Infinity Fabric mesh between GPUs.

MI325X — HBM3e refresh of the MI300 line at 256 GB; still relevant for clusters deployed before MI350/MI355 ramps.

Software stack — ROCm is the adoption variable versus CUDA; this is a technical integration question, not a benchmark horse race.

### Huawei Ascend (China-domestic AI silicon)

Ascend 910B/910C — Huawei's Ascend accelerators serve China-domestic training and inference clouds where NVIDIA datacenter GPUs face export restrictions. Public English-language spec pages are limited; our catalog lists 910B at 64 GB HBM2e with vendor documentation links — verify region-specific availability before any procurement planning.

### Samsung (HBM supply + logic roadmap)

Samsung competes on two axes:

1. HBM supplier — HBM3E stacks up to 36 GB (12H) with up to 9.2 Gbps per pin and up to 1,180 GB/s bandwidth per stack per Samsung Semiconductor's HBM product pages. NVIDIA and AMD both name Samsung (and Micron) as HBM3E suppliers on MI350-series materials.

2. Logic silicon — Mach-1 is an inference-focused accelerator roadmap item (LPDDR, not HBM datacenter class); distinct from GeForce/Instinct GPU competition but relevant to edge inference supply chain.

### Export controls (factual context)

US rules continue to restrict advanced NVIDIA datacenter GPU exports to certain regions, which structurally benefits domestic alternatives (Ascend) and open-ecosystem Western suppliers (AMD) in affected markets — cite official policy documents for legal/compliance decisions, not this brief.

---

Technical research for informational purposes only. Not financial advice. No investment recommendations. Specs sourced from vendor documentation; verify before engineering or procurement decisions.

ANALYSIS

Rivalry in mid-2026 is still a memory-capacity + software-stack + regional-availability triangle. MI355X wins on paper HBM3E capacity versus a single B200 GPU; NVIDIA wins on deployed CUDA ecosystem and HGX integration depth — neither spec sheet tells you which fits your workload.

[AMD Instinct MI355X][AMD MI355X product brief (PDF)][Samsung HBM][Huawei Ascend 910B]

Model spotlights

B200-SXM-192GB

Flagship Blackwell datacenter GPU on shipping HGX B200 platforms — the reference SKU for new AI cluster builds in 2026.

  • 192 GB HBM3e per GPU in our catalog (vendor product page)
  • 5th-gen Tensor Cores with native FP4 support on Blackwell architecture
  • 5th-gen NVLink — HGX B200 lists 1.8 TB/s GPU-to-GPU NVLink bandwidth
  • HGX B200 system: 1.4 TB total memory across 8 GPUs, shipping per NVIDIA
ANALYSIS

Check system-level power and liquid-cooling requirements on HGX/MGX partner pages — the GPU TDP alone does not describe rack deployment.

GB200-SUPERCHIP

Module-level Grace + Blackwell integration for rack-scale NVL systems — where AI factories optimize for tokens/sec per rack, not per GPU socket.

  • 384 GB aggregate HBM3e per superchip module (catalog; 2× B200 + Grace CPU)
  • Grace CPU handles host-side data movement and preprocessing
  • Designed for NVL72/NVL36 rack configurations with NVLink fabric across modules
ANALYSIS

Procurement is at system/module level (DGX/NVL), not as a loose PCIe card — cooling and network fabric dominate TCO more than single-GPU FLOPS.

H200-SXM-141GB

First HBM3e GPU on Hopper — the pragmatic upgrade path from H100 for memory-bound LLM inference without a full Blackwell retool.

  • 141 GB HBM3e at 4.8 TB/s bandwidth (NVIDIA H200 spec table)
  • Up to 3,958 TFLOPS FP8 sparse tensor performance (H200 SXM)
  • Same 700W configurable TDP envelope as H100 SXM
  • 900 GB/s NVLink and MIG support up to 7 instances
ANALYSIS

NVIDIA publishes up to 2× LLM inference throughput versus H100 on Llama2-class models — validate on your model size and batching, not marketing slides alone.

MI355X-OAM-288GB

AMD's CDNA4 flagship at 288 GB HBM3E — the primary open-ecosystem alternative when CUDA lock-in is a constraint and memory capacity drives model fit.

  • 288 GB HBM3E at 8 TB/s peak bandwidth (AMD product page + brochure)
  • CDNA4 on TSMC 3nm/6nm; 1400W TBP; liquid-cooled platform SKUs common
  • MXFP6/MXFP4 matrix formats; SR-IOV up to 8 partitions
  • UBB 2.0 eight-GPU mesh with 2.3 TB aggregate HBM3E
ANALYSIS

ROCm framework compatibility for your model (PyTorch/JAX/custom kernels) is the gating item — run a pilot before assuming MI355X drop-in replaces B200.

SAMSUNG-HBM3E-12H

Not a compute GPU — but HBM3E stacks from Samsung (and rivals) gate Blackwell and MI350 memory bandwidth; supply constraints here ripple across every accelerator SKU.

  • Samsung HBM3E: up to 36 GB (12H stack), up to 9.2 Gbps per pin, up to 1,180 GB/s per stack
  • HBM4 roadmap public: up to 3,300 GB/s bandwidth, 36 GB capacity (Samsung HBM page)
  • AMD names Samsung Electronics as HBM3E supplier on MI350-series materials
ANALYSIS

Watch HBM stack qualification timelines when planning cluster ramps — accelerator launch dates and memory vendor ramps are coupled.

Comparison matrix

DimensionB200-SXM-192GBH200-SXM-141GBMI355X-OAM-288GBASCEND-910B
GPU memory (GB)19214128864
Memory bandwidth (GB/s)800048008000—
Memory typeHBM3eHBM3eHBM3EHBM2e
ArchitectureBlackwellHopperCDNA4Ascend
TDP / TBP (W)10007001400400
Form factorSXMSXMOAMOAM
ANALYSIS

MI355X leads on single-GPU HBM capacity; B200/H200 lead on CUDA ecosystem maturity and HGX fleet integration. Ascend 910B is a different deployment context (China-domestic, lower published memory). Compare on region, software stack, and model fit — not one dimension.

[NVIDIA HGX B200][NVIDIA H200][AMD MI355X][Huawei Ascend 910B]