Compare time to the same training target
Read the article’s two 16-GPU submissions as complete-system results for one matched LoRA task.
AI-assisted learning visualization · schematic article explanation
MLPerf Training times a system until it reaches a defined quality target. These two article entries use the same LoRA task, target and GPU count; unrelated pretraining tasks cannot be averaged into this comparison.
What is the difference between B200 and GB200?
B200 is a Blackwell GPU; GB200 combines a Grace CPU with two Blackwell GPUs. The GB200 NVL72 places 36 Grace CPUs and 72 GPUs in one liquid-cooled NVLink domain. HGX B200 uses a different host and system configuration. Compare complete systems for your workload, rather than treating the names as two interchangeable GPU configurations.
Is GB200 faster or more energy-efficient than B200? The comparable training submissions below show a faster GB200 result for one workload. They do not establish a universal speedup or a performance-per-watt ranking; those require comparable workload and power measurements.
For inference planning, start with the GPU buying guide. For communication bottlenecks, read NVLink and scale-up vs scale-out.
The comparison
MLPerf Training measures the time a complete system needs to reach a specified quality target on a defined workload. It does not isolate the GPU from its host, cooling, network, software stack or tuning. The MLPerf Training v6.0 results include B200 and GB200 submissions, but a cross-submission difference is not a controlled measurement of GPU-only speed or energy efficiency.
The previous version of this article divided training time by GPU count and called the result “efficiency.” That calculation was invalid: adding GPUs reduces both the numerator and increases the denominator, making a larger cluster appear better even when its scaling is poor. It also pooled Llama 2 70B LoRA fine-tuning with Llama 3.1 405B pretraining. Their times cannot be averaged into one hardware ranking.
A comparable published workload
NVIDIA's MLPerf Training v6.0 results table lists the following closed-division submissions for Llama 2 70B LoRA fine-tuning on SCROLLS GovReport. Both target an evaluation loss of 0.925 and use 16 GPUs:
| Submission | System | Reported time to target |
|---|---|---|
| 6.0-0071 | 16 × GB200, HPE GB200 NVL72, NVIDIA NeMo | 5.3 min |
| 6.0-0069 | 16 × B200, HPE ProLiant Compute XD685, NVIDIA NeMo | 6.3 min |
For these two submissions, the GB200 system finished 1.0 minute sooner. The ratio of the reported times is 6.3 ÷ 5.3 ≈ 1.19. This is a system-level observation; the entries have different server platforms and can differ in configuration and optimization. It does not establish a general 19% GPU advantage, and it contains no measured energy or cost data.
The MLCommons results dashboard is the place to check submission details and any later result corrections. The numbers above are rounded as published on NVIDIA's results page.
What the hardware specifications say
The GB200 NVL72 product page describes a liquid-cooled rack with 36 Grace CPUs and 72 Blackwell GPUs in one NVLink domain. NVIDIA states 130 TB/s aggregate NVLink bandwidth for that domain. A GB200 Grace Blackwell superchip combines one Grace CPU with two Blackwell GPUs. These are design specifications, not measured application speedups.
An HGX B200 system uses B200 GPUs in a different host and interconnect configuration. The larger NVLink domain of NVL72 can matter when a workload requires frequent communication across more than one server. Whether it improves a specific training job depends on model partitioning, batch size, network, software and the comparison system.
How to compare deployments
For a decision between B200 and GB200, run the same model, data, quality target and software revision on both candidate systems. Report wall-clock time together with GPU count. For scaling, compare speedup against the increase in GPU count on that same workload. For energy and cost, measure system power and total run cost, including cooling and infrastructure, rather than inferring them from training time or product specifications.
The published v6.0 entries above show that one 16-GPU GB200 submission finished this LoRA task sooner than one 16-GPU B200 submission. They do not support the former article's “2× efficiency” claim, a performance-per-watt ranking, or a general purchasing recommendation.