Big Battlemage: Two new Arc graphics cards with ample memory for AI
Intel launches Big Battlemage: The BMG-G31 chip is set to offer more compute power and double the graphics memory for (AI) workstations at around $1000 USD.
(Image: Intel)
Intel is launching two new Arc Pro graphics cards for workstations: the Arc Pro B70 and the lower-spec Arc Pro B65. Both feature 32 GB of graphics memory, making them suitable for memory-intensive AI acceleration tasks. The B65 will be exclusively available through Intel's graphics card partners such as Asrock, Sparkle, Gunnir, Maxsun, and Lenovo. Intel plans to sell its own branded Arc Pro B70 cards for around $1000 USD, which translates to approximately 1050 to 1100 Euros in Germany after currency conversion and taxes.
According to Intel, the Arc Pro B70 is between four and 69 percent faster than the already available Arc Pro B60 with its smaller GPU, based on various tests from the SPECviewperf15 suite. No performance figures have been released for the new Pro B65 yet, but its BMG-G31 graphics chip is configured with the same compute power as the Pro B60 and shares the same upper limit for electrical power consumption. The faster and larger memory or the twice as fast system interface could have a positive impact here, as the large Battlemage chip, unlike its predecessors with the BMG-G21 chip, has a fully equipped PCIe 5.0 interface with 16 lane pairs. It also uses 24 GB of RAM instead of 32 GB.
Otherwise, the technical specifications are similar to those of the smaller siblings: four DisplayPorts 2.1, XMX units for matrix calculations in AI applications, and extensive video functions to offload the processor, including support for codecs like H.264, H.265 (HEVC), AV1, and VP9.
While the Arc Pro B65 operates with a Thermal Design Power (TDP) of 200 watts, Intel has configured its own Arc Pro B70 model with 230 watts. Partners have more flexibility, with a range of 160 to 290 watts, allowing for different card sizes beyond the full-height, dual-slot form factor.
Videos by heise
Intel claims: "Faster than Nvidia RTX Pro 4000"
In some pre-presented benchmarks, Intel claims the Arc Pro B70 delivered noticeably higher performance than Nvidia's RTX Pro 4000. However, Nvidia's cards have a lower TDP of only 140 watts compared to 230 watts and also feature only 24 GB of graphics memory each.
For instance, the larger memory of the Arc Pro B70 allows for up to 2.2 times larger context windows per GPU in the Large Language Model Llama 3.1 8B. The memory is exhausted only at around 93,000 tokens. In the Ministral Instruct 2410 8B benchmark, Intel states that an Arc Pro B70 outperforms an RTX Pro 4000 by up to 83 percent in throughput (tokens/s) with simultaneous access from multiple users, and is sometimes six times faster in terms of response time (Time to 1st Token).
When used in a quad configuration, Intel also positions the Arc Pro B70 ahead, for example with Mistral-Small and 24 billion parameters. However, in all these comparisons, Intel uses the bFloat-16 format, which is becoming less common for inferencing use due to its significant memory footprint. Quantization to 8, 6, or even 4 bits is often used.
Nevertheless, there are also results with lower precision: Intel also sees its Arc Pro B70 significantly ahead in Qwen3 32B with FP8 or DS-R1-Distill-Qwen3 with INT4 quantization. More precise details will only be revealed by independent tests.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(csp)
