A lot of information about the AMD Radeon 380X has come up, and it looks like this might be a flagship model in its Radeon lineup. The existence of AMD Radeon 380X (Codenamed Fiji) and its HBA feature was highlighted via two LinkedIN profile from two of the AMD employees. One of them termed the upcoming card as ‘the largest King of the Hill line of products’.
Through another source who works in AMD, it was known that they are working on ‘the world’s first 300W 2.5D discrete GPU SoC using stacked die High Bandwidth Memory and silicon interposer’. This came from Linglan Zhang, one of the technical staff at AMD who posted in his LinkedIN profile. What is interesting is that AMD is implementing an HBM- DRAM.
For those who are not familiar with this term, High Bandwidth Memory are vertically stacked memory chips which have two functions- significantly higher bandwidth and ability to stack memory vertically helps to save space on the PCB. HBA also allows lower power consumption in comparison with a traditional memory implementations.
The vertically-stacked memory used in real-life is something I am looking forward to. Each HBA stack has 4x DRAM modules using 2x 128-bit channels per die. Unlike the traditional DDR implementation, each channel is independent, giving the freedom to each to run at optimal frequency and latency. As of now, current HBA package provides 1GB-32GB capacity per stack, along with 128GB/s- 356GB/s memory bandwidth.
Furthermore, Ilana Shternshain (ASIC physical design engineer in AMD) accidently mentioned AMD 380X in her job profile. It read: “taping out state of the art products like Intel Pentium processor with MMX technology and AMD R9 290X and 380X GPUs.”
It should be noted that the existing GPU’s TDP- AMD Radeon R9 290X- is 290w. Not really sure how this is someone would be bragging about, considering Nvidia GeForce GTX 980- a single GPU core based solution- carries a TDP of 165w. Even if the performance is justified, in today’s day of age a lot of end-users would want to see performance and efficiency while maintaining a cooler temperature on load.
Other than the information retrieved from LinkedIN profiles, it is speculated the unknown (for now) AMD R9 GPU will use 4096 stream processors with 64 compute units, 256 texture units that will use 4GB memory. This memory interface is said to provide a jaw-dropping memory bandwidth of 640GB/s.