AMD has released what is possibly the most awaited graphics chip of 2015, Fiji. It has 8.9 billion transistors, which is more than any other GPU. The key feature of the Fiji chip is the use of High-Bandwidth Memory (HBM) instead of the GDDR5 that has been used in high-end GPUs for many years. GDDR5 memory is extremely fast, but it also uses a lot of power. The increasing bandwidth needs of faster and faster GPUs will result in too much power being spent on bandwidth if current trends are maintained. GDDR5 chips are also rather large and take up a lot of space in a graphics card. AMD has decided to solve the problem using HBM. HBM chips are stacked on top of each other and run at a comparatively low frequency (500 MHz / 1 Gbps instead of GDDR5's 1.75 GHz / 7 Gbps). The GPU and memory stacks are placed on top of a interposer, which is a piece of silicon with wires that connect chips to other chips. The interposer results in the memory being much closer to the GPU, reducing power, and also allows for much wider memory buses than what would be possible otherwise. Thus, the memory bus is 4096 bits in Fiji's implementation (4 stacks, 1024 bits per stack) and so the total memory bandwidth goes up compared to GDDR5. The result is that Fiji has 512 GB/s of memory bandwidth, which is more than the R9 290X (320 GB/s) and the GTX TITAN X (336 GB/s), and uses a lot less power too. These benefits don't come without costs though. The HBM and interposer are expensive to make, and the first iteration of HBM only supports a maximum of 1 GB per stack. Thus, Fiji will have "only" 4 GB of memory, which was large last year but not so much now that newer cards have 6-12 GB of memory. Even AMD's own R9 390 and R9 390X, which were also announced today, have 8 GB of memory. The first graphics cards based on Fiji will be under a separate name, Fury. That name may be familiar to some of the elderly folks here. Superficially, Fury seems to be AMD's version of NVIDIA's TITAN, representing parts that not only use a large chip, but are also much more powerful than existing parts. However, the Fiji chip is going to be used in a wider variety of parts than Big Kepler or Large Maxwell (I say "Large" instead of "Big" because that Maxwell part has slow double-precision) in the consumer area. The R9 Fury X is the most powerful single-GPU part, and like the R9 295X2, it is water cooled. The price is $649. I expect it to be about twice as fast as the HD 7970 GHz Edition, which would put it slightly above the 980 Ti and TITAN X. I think it's impressive how far AMD has come since the HD 7970 without any process shrinks. The R9 Fury (no suffix) is an air-cooled part, and while we don't have details on its configuration, I expect it to use a partially disabled Fiji. The price is $549. The R9 Nano is really small, only 6" in length. My graphing calculators are longer than that. This card will be available in the summer. And last but presumably the most powerful, a currently unnamed dual-GPU card will be present in AMD's Project Quantum, an upcoming small form factor case. As for the rest of the lineup, few details were given besides the names and the memory capacities, but the rumor mill expects only minor updates. Sources: AnandTech, AnandTech.