Jason A. Hoffman

Jason A. Hoffman, PhD | March 21, 2026

---

Claims such as “20% of GPUs fail” often combine measurements from different stages. The percentage may refer to wafer yield, packaging attrition, burn-in screening, system test, hardware replacement in service, or disruptions attributed to a GPU during a training run. Those quantities do not share a denominator or time window.

The estimates used below place burn-in loss around 3–8% and total manufacturing attrition around 15–25% in mature production. A separate Meta observation can be annualized to roughly 9% in-service failure under its assumptions. These figures describe different populations and should not be collapsed into “20% of GPUs fail during burn-in.”

A manufacturing-yield calculation provides one useful decomposition.

---

Four Manufacturing Stages

A datacenter GPU passes through four sequential manufacturing stages before it ships. Losses are multiplicative: a module must survive every stage.

Stage	What happens	Typical loss
Die yield	Wafer fabrication (TSMC N4/N5)	5–15%
Packaging yield	CoWoS-L assembly (chiplets + interposer + HBM)	5–15%
Burn-in	Thermal/voltage stress screening	2–8%
System-level test	Full module functional validation	1–3%

The “typical loss” ranges above are industry-wide figures for complex multi-chip modules, drawn from semiconductor industry reporting and test house disclosures. They are not NVIDIA-specific published numbers. Published per-stage yield data for Blackwell does not exist. NVIDIA acknowledged shipping “low-yielding Blackwell material” in its Q3 FY2025 earnings commentary, and SemiAnalysis has reported on CTE mismatch issues during the CoWoS-L ramp, but neither source gives a clean per-stage breakdown.

What we can do is show what the math looks like at different points in the range:

• Optimistic (mature production): 90% × 90% × 95% × 98% = 75.4% yield (24.6% total attrition)
• Pessimistic (early ramp): 85% × 82% × 92% × 96% = 61.4% yield (38.6% total attrition)

These figures illustrate the structure; they are not measurements. Moderate losses at each stage can compound into total attrition of 20–40%. A quoted 20% may therefore describe attrition across a sequence, while burn-in supplies only one part. Without a source and denominator, however, the percentage cannot be assigned to a stage confidently.

---

Blackwell Adds Manufacturing Challenges

Three factors compound relative to prior generations:

Power density. Blackwell GPUs draw 700–1,000W per device. Burn-in chambers must deliver and dissipate this power at elevated junction temperatures (typically 125°C+). KYEC, which handles over 90% of NVIDIA’s AI chip testing, had to upgrade their burn-in ovens from 600W to 1 kW capacity just to handle Blackwell.

Dual-die architecture. The B200 uses two compute dies connected via NVLink-C2C on a CoWoS-L interposer. A failure in either die, the silicon bridge, or the interposer kills the entire module. The failure surface area is larger than a monolithic chip.

HBM stacking complexity. Each B200 module includes 8 stacks of HBM3e memory, with 8–12 DRAM dies per stack. A single defective die can kill an entire stack, and a failed stack can render the module non-functional.

According to SemiAnalysis reporting (not NVIDIA disclosure), a CTE mismatch between the GPU chiplets, silicon bridges, interposer, and substrate caused warping and failures under thermal cycling at 1,000W during the initial Blackwell ramp. NVIDIA addressed this by redesigning the top routing metal layers and adjusting the bump geometry. The B200 chip yield is now estimated at 90–95%, though not yet at TSMC’s internal targets.

---

The Other Number: In-Service Failure

Then there’s the separate question of what happens after the GPU ships and goes into production:

Metric	Source	Value
Annualized GPU failure rate	Meta (16K H100 cluster, 54-day window)	~9%
Cluster MTTF (16K GPUs)	Meta	1.8 hours
Cluster MTTF (131K GPUs)	Meta	14 minutes
GPU share of unforeseen disruptions	Meta (Llama 3 405B training)	30.1%
HBM memory share of disruptions	Meta (Llama 3 405B training)	17.2%

The Meta numbers come from the Llama 3 technical report, Table 5: a 54-day snapshot of 16,384 H100 GPUs with 419 unexpected disruptions.

The approximately 9% figure annualizes a short observation window and should not be treated as a measured lifetime failure probability. Burn-in screens for early defects before shipment, while the Meta report records disruptions during operation. The rates answer different reliability questions.

A claim that “30% fail” may also confuse the share of disruptions attributed to GPU hardware with the share of GPUs that failed. One is a fraction of events; the other would require a cohort of devices and a time interval. Manufacturing attrition and an annualized in-service rate should not be added because they apply to different populations and stages.

---

How Yield Affects Cost

Manufacturing attrition raises the expected cost of a shippable module. If C is the pre-yield manufacturing cost assigned to one attempted module and Y is total yield, the expected cost per shippable module is C / Y. The incremental cost associated with attrition is C × (1 / Y - 1).

That formula should not be applied to an OEM sale price or cloud rental price as though either were manufacturing cost. Those prices also contain memory, boards and systems, integration, support, channel terms, margin, scarcity, and product-mix effects. Public B200 price estimates cannot reveal the dollar cost of failed modules without a defensible estimate of C and the scope included in Y.

Better yield lowers manufacturing cost per shippable unit, all else equal. The division of that gain among supplier margin, system price, and customer price depends on contracts, demand, competition, and product mix.

---

The Short Version

When someone says “20% of GPUs fail,” ask them: which stage?

• Manufacturing attrition describes attempted units lost before shipment.
• Burn-in loss is one component of manufacturing attrition.
• In-service hardware failure needs a shipped-device cohort and observation window.
• Disruption share describes events attributed to a cause, not the share of devices that failed.

Require a stage, cohort, denominator, and time window before using any GPU “failure rate.”