Intel Flex GPUs take aim at the Nvidia A10: up to 5x faster in 8-bit decode and HEVC transcode applications
The Intel Flex 170 is a 150 watt affair with 32 Ray Tracing units and 32 Xe Cores and a full height PCIe design. It features the Xe HPG architecture with 2 media engines and a single slot design. The Flex 170 GPU is shipping now to customers. The Intel Flex 140 GPU is a 75 watt variant with 16 Ray Tracing units and 16 Xe Cores and a half height PCIe. It also features the Xe HPG architecture with 4 media engines and and a single slot design. The Xe Media Engine can handle up to 8k60 in 12-bit HDR decode and up to 8k10-bit HDR encode. VP9, AVC, HEVC and AV1 are all supported with a complete media processing and delivery software stack. Now let’s talk about the benchmarks. Intel is claiming up to 5x performance in some workloads when comparing against the NVIDIA A10 GPU. In AVC 8-bit decode workloads, the Intel Data Center GPU Flex 140 achieves 168 streams versus 37 on the NVIDIA A10. Similarly, HEVC, AV1 and VP9 streams are 208, 218 and 228 versus the A10’s 81, 49 and 66. Coming over to transcode performance, the Intel Data Center GPU Flex 140, achieves 8 streams versus 1 for 4K60 performance quality H.265 HEVC transcoding and 36 streams versus 7 on the 1080p60 performance quality preset. Keep in mind the benchmarks use the the Flex 140, which should have twice the encode/decode performance of the Flex 170 thanks to double the Media Engines so you can cut the streams in half for the Flex 170 GPU (this performance should be up to 2.5x NVIDIA A10). Intel is also claiming a 30% distribution cost savings vs the x264 medium. Coming over to the Cloud Gaming benchmarks, Intel is reaffirming broad support for the entire cloud gaming software stack and running VDI seamlessly across the CPU and GPU. Here Intel has not compared its cards to the NVIDIA A10, which likely means the latter would win in this workload. However, the cards still achieve a very impressive number of streams in various cloud-oriented games. The Flex 170 GPU can support up to 23 streams of the Asphalt 9: Legends cloud game. Up next we have a lot of inference benchmarks. The most interesting of these benchmarks is one where the workload is combining HEVC transcoding and Resnet50. Presumably, this is a standard real world pipeline and shows Intel’s Flex GPU 170 pulling ahead of nvidia by 35%. Most of the other benchmarks do not have a direct comparison with NVIDIA but do show case a lot of AI Inference workloads which should help provide reference base lines to prospective clients. Finally Intel is announcing 15+ system design wins including Lenovo, Cisco, Dell, HP and Supermicro. The full slide deck can be seen below:
