Intel recently released a new generation of professional graphics cards, the Ruixuan Pro B70/B65, which used the second-generation large GPU that was originally planned to create the Ruixuan B770 gaming graphics card. Unfortunately, the latter has been cancelled. Unlike game cards, reviews of this type of professional cards are very rare, and we are preparing them. Here’s a look at Germany’s HardwareLuxx experience.
Ruixuan Pro B70 uses TSMC's 5nm process, with a core area of 368 square millimeters, integrated 27.7 billion transistors, and is equipped with 32 Xe2 cores, a maximum frequency of 2.8GHz, 256-bit 32GB GDDR6 video memory, a bandwidth of 608GB/s, and four DP output interfaces.
Computing performance FP8 367 TOPS, FP32 22.9 TFlops, and supports PCIe 5.0 x16, the power consumption of the whole card is 230W, the adjustable range is 160-290W, and a single 8-pin power supply.
Ruixuan Pro B65 is reduced to 20 cores and 2.4GHz frequency. The video memory remains unchanged, the computing power is reduced by almost half, FP8 197 TOPS, and the entire card power consumption is 200W.
For this test, four B70s were luxuriously arranged, all of which are public versions, equipped with 16-core Xeon W5-3435X and eight 16GB DDR5-4800. The system has both Windows 11 and Ubuntu.
In single card testing,The performance of workstation applications is still good, basically on the same level as the AMD Radeon Pro W7800, and sometimes it can be significantly ahead, or even tied with the more advanced Radeon Pro W7900., such as Bldender Classroom.
In the AI test, AIDA64 double-precision computing power, GeekBench AI quantification, MLPerf Llama generation speed and first word time, Procyon image and text generation, etc. all performed well, especially the first word time far exceeded AMD and NVIDIA.
There is no problem with dual-card or four-card recognition, but the efficiency is worrying. Almost no workstation or AI test can use four cards well, and Blender cannot even use two cards well.
Ubuntu LMStudio is the only one that can reflect the advantages of multiple cards, but it cannot achieve linear improvement. The core reason isMultiple cards are not directly connected, but bypass the PCIe channel.
In addition, dual-card 64GB video memory is not enough to handle some large 120B models. Four-card 128GB is more practical.
Power consumption is gradually increasing.Single card standby is 26W, AI inference is about 180W, dual cards are close to 370W, and four cards are more than 720W.
However, considering the official nominal power consumption of the entire card is 230W, the four cards should theoretically exceed 900W, which is not bad.
