· Research & technology· 日本語で読む
2389.ai Publishes Experimental Report on NVIDIA DGX Spark
2389.ai published an experimental report on NVIDIA DGX Spark. The publication date could not be confirmed from the materials provided. Rather than delving into detailed evaluation results, the report raises questions about verification methods and differences from large-scale computing infrastructure.
What happened
2389.ai published an experimental report on NVIDIA DGX Spark. The publication date could not be confirmed from the materials provided.
Analysis
The key to reading this report is not to treat the results of an individual experiment as equivalent to large-scale operational performance. An experimental article about DGX Spark and an announcement about infrastructure connecting multiple computing resources show different things. The former covers testing under specific conditions, while the latter describes an equipment configuration. Grouping both under “NVIDIA computing power” obscures differences in scale and measurement conditions.
In related coverage, we reported on Lambda using NVIDIA CPO switches to connect GB300 NVL72 racks equipped with more than 10,000 GPUs Lambda connects GB300 NVL72 racks with NVIDIA CPO switches; more than 10,000 GPUs. No performance figures were provided in that report. Rather than treating the experimental report and this infrastructure information as successive stages of the same product evaluation, they should be compared as materials with different subjects of verification and types of evidence. The experiment’s results do not, by themselves, justify extending conclusions to rack-scale processing performance or operational track records.
The competitive implications are also not simply about which product performs better, but about the conditions under which users can evaluate computing resources. For large-scale infrastructure connections, evaluation involves not only the number of GPUs but also the network configuration and actual performance figures. To use an experimental article about an individual product as a basis for judgment, it is necessary to check whether the target model, software environment, measurement method and comparison conditions are specified. When these conditions differ, the extent to which results can be applied to other equipment or operations also changes.
Access is another issue. The U.S. Bureau of Industry and Security’s interim final rule covers remote access to advanced AI semiconductors by companies headquartered in China and elsewhere U.S. BIS regulates remote access to advanced AI semiconductors. No direct connection between this rule and the DGX Spark experimental article could be confirmed, but evaluating computing resources requires separately considering not only equipment specifications but also the form of use and the rules that apply to users. Future verification should examine reproducibility under consistent conditions, operational status and costs. See the original source for details.
Sources
Translated from the Japanese original: /news/2389-ai-nvidia-dgx-spark-no-jikken-repooto-wo-keisai-9e1bd4. The Japanese version is authoritative if they differ.