We're probably not all that far away from the Radeon RX Vega launch, which AMD has slated for the first half of this year. Supplies of Fiji-powered Radeon graphics cards appear to be drying up at online retailers, and a new set of information turned up by the inveterate leakers at VideoCardz suggests that certain implementations of Vega silicon could be in testing now.
The site noticed that a new AMD graphics chip has shown up in the CompuBench database of OpenCL and CUDA performance numbers. This device, called AMD 687F:C1, apparently shares an identifier with earlier Vega products that AMD has shown in the past. CompuBench collects a wealth of information about the device being benchmarked, including a reported maximum clock frequency and a maximum number of OpenCL compute units, among other details.
In the case of this purported Vega silicon, CompuBench says the device reports 64 OpenCL compute units running at possible maximum clock frequencies of 1000 MHz or 1200 MHz. If a Vega NCU continues the GCN tradition of having 64 stream processors per unit, that would indicate a 4096-SP chip with 8.2 TFLOPS of raw computing performance at 1000 MHz, or 9.8 TFLOPS at 1200 MHz.
Those rough figures are slightly higher than the GTX 1070's 6.5 TFLOPS and the GTX 1080's 8.9 TFLOPS by our calculations. While TFLOPS aren't a perfect approximation of delivered performance by any means, they do mesh with my impressions of how an RX Vega performs in the forms AMD has demonstrated thus far.
If these figures are accurate, we're also curious what they mean about the Radeon Instinct MI25 accelerator's performance potential. I guessed that card could deliver 25 TFLOPS at a 1500 MHz clock speed—2x its single-precision throughput—using Vega's new packed math support. Assuming my guess is correct, perhaps Instinct cards are getting the best of Vega silicon to hit those high clocks at reasonable power levels. Whatever the case may be, we won't know until the Radeon RX Vega actually hits the streets.