1. Can NameTag 3 be trained on a regular CPU or is a GPU required?

Training NameTag3 on a CPU is possible, but it will take significantly longer than on a GPU. The total time depends on the size of your dataset and the CPU’s performance.

For reference, the NameTag 3 Multilingual 250203 Model was trained on an NVIDIA Quadro P5000 GPU with 21 corpora and a batch size of 8 in about 9 hours.

Typical CPU slowdowns compared to this GPU run:

• Fast 16-core CPU: ~10× slower, ~3 days
• 8-core desktop CPU: ~6 days
• Laptop CPU or weaker: up to ~40× slower, ~15 days

As a rough guideline, CPU training is often around 30× slower than GPU training.

2. What commercial licensing options are available for NameTag 3?

Please refer to the Commercial page for details about the available commercial licensing options.

3. What token throughput can the NameTag 3 webservice achieve?

With pre-tokenized input, i.e., tokenization was not performed by the server, using --input=vertical (see Formats), and measured on localhost to exclude network latency. GPU memory consumption as reported by nvidia-smi during processing.

GPU/CPU	Batch	Threads	GPU Mem (GB)	Multilingual	Czech CNEC2.0
				Versions: 250203, 260521, CL 260717	Version: 240830
				Params: 355M	Params: 126M
				Entities: flat	Entities: nested
				Context: document	Context: sentence
				Decoding: non-autoregressive	Decoding: autoregressive
NVIDIA A40	8	1	10	Throughput: 11.1k tok/s	Throughput: 868 tok/s
	16	1	10	Throughput: 11.4k tok/s	Throughput: 1404 tok/s
	32	1	10	Throughput: 11.6k tok/s	Throughput: 2216 tok/s
	64	1	12	Throughput: 12.2k tok/s	Throughput: 3592 tok/s
NVIDIA A100	8	1	10	Throughput: 20.7k tok/s	Throughput: 874 tok/s
	16	1	10	Throughput: 22.0k tok/s	Throughput: 1429 tok/s
	32	1	10	Throughput: 23.0k tok/s	Throughput: 2285 tok/s
	64	1	12	Throughput: 23.6k tok/s	Throughput: 3769 tok/s
CPU (AMD EPYC 7313)	64	32	-	Throughput: 314 tok/s	Throughput: 511 tok/s

With tokenization included in processing, i.e., the server tokenized the input text, using --input=untokenized (see Formats), and measured on localhost to exclude network latency. GPU memory consumption as reported by nvidia-smi during processing.

GPU/CPU	Batch	Threads	GPU Mem (GB)	Multilingual	Czech CNEC2.0
				Versions: 250203, 260521, CL 260717	Version: 240830
				Params: 355M	Params: 126M
				Entities: flat	Entities: nested
				Context: document	Context: sentence
				Decoding: non-autoregressive	Decoding: autoregressive
NVIDIA A40	8	1	10	Throughput: 7.0k tok/s	Throughput: 860 tok/s
	16	1	10	Throughput: 7.1k tok/s	Throughput: 1372 tok/s
	32	1	10	Throughput: 7.2k tok/s	Throughput: 2088 tok/s
	64	1	12	Throughput: 7.4k tok/s	Throughput: 3011 tok/s
NVIDIA A100	8	1	10	Throughput: 7.3k tok/s	Throughput: 847 tok/s
	16	1	10	Throughput: 7.5k tok/s	Throughput: 1350 tok/s
	32	1	10	Throughput: 7.6k tok/s	Throughput: 2006 tok/s
	64	1	12	Throughput: 7.6k tok/s	Throughput: 2907 tok/s
CPU (AMD EPYC 7313)	64	32	-	Throughput: 300 tok/s	Throughput: 425 tok/s