In addition to the new Intel Xeon 6900P With high-performance cores, Intel also wanted to offer data on its Gaudi 3 accelerator. As promised in the last HotChips, Intel has delivered more information about its next-generation accelerator. Intel has also taken the opportunity to announce the general availability of this Gaudi 3 accelerator for the next quarter.
Intel’s Gaudi 3 accelerator offers up to 1,835 TFLOPS Peak FP8, a memory of 128GB HBM2e and has a 600W TDP. All this in a size of double slot with 26.67 cm in length.
Intel has also wanted to offer performance data in comparison With NVIDIA’s top solution, the H100. As shown by Gaudi 3, it offers 1.09 times more performance in inference with LLaMa 3 8B and 1.8 times more performance per wattThese figures rise to 1.19x more in inference with Call 2 70B and 2 times more performance per watt.
This is possible thanks to its architecture with 8 MME units and 64 TCPwhich together with the 128 GB of HBM memory offer a bandwidth of 3.67 TB/s. It has 96MB SRAM per tile with a bandwidth of 19.2 TB/s. It is capable of managing up to 1200 GB/s bi-directional for networks and is mounted on a PCIe 5.0 interface. All this offers figures of 1,835 TFLOPS BF16 and FP8 for matrices and 28.7 TFLOPS BF16 for vectors. Its 14 decoders compatible with HEVC, H.264, JPEG and VP9 offer support for tasks of this type.
Along with this accelerator, Intel offers Gaudi Software Suitea set of tools that allow get the most out of it and maximum performance of this Gaudi 3 accelerator.
End of Article. Tell us something in the Comments!
Juan Antonio Soto
I am a Computer Engineer and my specialty is automation and robotics. My passion for hardware began at the age of 14 when I gutted my first computer: a 386 DX 40 with 4MB of RAM and 210MB of hard drive. I continue to give free rein to my passion in the technical articles I write on Geeknetic. I spend most of my free time playing video games, both contemporary and retro, on the more than 20 consoles I own, in addition to the PC.
Add Comment