Hot Chips Intel says its datacenter-focused Flex-series GPUs, codenamed Arctic Sound, are finally ready, and computer makers are expected to begin shipping systems over the next few months.
While the cards share a similar Xe core architecture to Intel’s AI and HPC-focused Ponte Vecchio GPUs, detailed at length at this week’s Hot Chips virtual event, the chipmaker’s Flex-series GPUs target a very different demographic: media and game streaming applications where stream density outweighs computational power.
The cards themselves are closer to Intel’s recently announced Arc discrete GPU family, but have been tuned for use in datacenter environments. The cards feature up to four Xe media engines and 32 Xe cores and ray tracing units, while AI acceleration is achieved using the card’s XMX matrix math processors.
The Flex-series GPUs are available in two form factors: the 75W Flex 140 with 12GB of memory and the 150W Flex 170 with 16GBs. Both feature a PCIe 4.0 interface and are passively cooled single slot GPUs. However, the lower TDP offering is half height, which enables it to be deployed in high-density systems with up to 10 cards.
According to Intel, the Flex 140 offers five times the media transcode throughput and twice the decoding performance of Nvidia’s A10 GPUs, and can support up to 36 concurrent 1080p 60fps streams or eight 4K 60fps streams. And, at least for streaming applications, this scales linearly, with a 10-card host capable of delivering 360 simultaneous streams at that resolution or 80 4K streams using the popular HEVC H.265 format, it’s claimed.
Intel’s big feature with this generation is native support for AV1 encoding, which the chipmaker says its the first to deliver in a datacenter form factor.
The royalty-free codec was developed under the Alliance for Open Media by several large streaming media companies including Amazon, Netflix, and Google to name a few. The format promises to deliver a 30 percent bandwidth savings compared to the already space-efficient HEVC format.
The chipmaker claims this stream density significantly reduces the operating cost for media streaming and Android game streaming workloads, as well as applicability for streaming AI and the metaverse.
Speaking of cloud gaming, Intel claims a single Flex 170 can deliver 68 game streams at 720p 30fps, and is validated on nearly 90 popular Android titles at launch.
However, for AI and HPC applications, Intel’s Flex GPUs come up short. According to Intel, the Flex 140 delivers peak performance of 8 teraflops in FP32 calculations, putting it on par with Nvidia’s $450 MSRP A2000. Intel’s Flex 170, meanwhile, offers roughly twice the performance at 16 teraflops of FP32, roughly half the performance claimed by Nvidia’s A10 at 31.2 teraflops.
The performance figures here suggest the Flex 140 is simply a dual die version of Intel’s A50 workstation GPU announced earlier this month. That card boasted peak performance of 4.8 teraflops of FP32 in a half-height, dual-slot, actively-cooled form factor.
Intel’s Flex-series GPUs will be available from more than 15 OEM partners, including Cisco, Dell Tech, Hewlett Packard Enterprise, Supermicro, Inspur, and H3C in the coming months. ®