Lately, it seems like every morning we wake up to a single dewdrop of information about AMD's upcoming RX Vega graphics cards. This morning's sprinkling comes from an entry in the online CompuBench database suggesting the existence of a Vega development board with higher core clocks and memory pools than previously reported. The database entry shows a card with a "CL_DEVICE_GLOBAL_MEM_SIZE" of 16978542592, suggesting a 16GB pool of memory, and a "CL_DEVICE_MAX_CLOCK_FREQUENCY" of 1600, which probably means a 1600 MHz core clock. Both of these figures represent big steps up from previous reports of Vega development cards clocked at 1200 MHz and with 8 GB of on-package memory.
Not too long ago, AMD promised it'd make GPU compute chips with as much as 25 TFLOPS of FP16 compute capability. Rumors suggest that the RX Vega will ship with 4096 stream processors, each capable of four FP16 operations per clock. Reaching 25 TFLOPS at the specified precision would require a GPU core operating with a minimum clock speed of 1500 MHz. The CompuBench entry suggests that at least one version of Vega silicon is capable of surpassing this requirement.
Even if those specifications turn out to be true, consumer-focused cards with those capabilities might not be forthcoming. By all accounts, AMD wants to tap the lucrative deep learning GPU market that Nvidia has dominated on the road to ever-increasing profits. The silicon package with 16 GB of HBM2 could be limited to GPU compute cards with stratospheric prices and no display outputs.