Radeon R9 Fury X
Fury X is the first solution using a new type of High Bandwidth Memory. HBM memory outperforms GDDR5 memory in every way, offering higher bandwidth with lower power consumption and a significantly smaller physical size to help accommodate smaller form factors.
The Fiji GPU, which powers the new model of the video card, uses the old proven 28 nm process technology from TSMC, as the “transitional” 20 nm process did not live up to expectations, not providing any significant improvements compared to 28 nm.
Fiji has become the largest ASIC released by AMD. The die area is 596 mm2, which is 36% larger than the area of Hawaii (438 mm2) and almost equal to the area of NVIDIA GM200 (601 mm2).
AMD has never released internal names for the interim versions of GCN, but the combination of features places Fiji in a category informally called GCN 1.2. First introduced in the Tahiti GPU (Radeon HD 7970), GCN is dual-purpose: 3D graphics rendering and, to an equal extent, general-purpose computing (GP-GPU). The latter feature enables the presence of hardware schedulers in the core, as well as the potential ability of all GCN-based ASICs to perform double-precision (FP64) operations at 1/2 the speed of FP32.
Both features consume a significant portion of the processor's transistor budget, which makes it quite impressive that AMD has managed to increase chip capacity to 8,9 billion transistors without deviating from the principles laid down in the early days of GCN. FP64 performance in Fiji-based consumer products is limited to 1/16th of FP32, but AMD will likely release FirePro accelerators that fully unlock the capabilities of the new ASIC.
From GCN 1.1 products, Fiji inherits flexible control of supply frequency and voltage GPU, an XDMA block that provides synchronization of several GPU in CrossFre via PCIe bus, and TrueAudio audio DSP.
Version 1.2 brought performance optimizations for geometric processors to GCN, as well as a new loseless color compression format, which saves memory bus bandwidth. In addition, it became possible to encode and decode video in H.264 format with a resolution Ultra HD
Based on the configuration of the computing units, Fiji is easier to describe as a double Tahiti. Fiji and Hawaii are in a 16:11 ratio (~45% increase) in the number of shader ALUs (stream processors) and texture units.
This unusual ratio in the configuration of shader ALUs and texture units in Hawaii and Fiji looks suspicious. And for good reason. The Fiji block diagram shows that the number of the largest building blocks GPU – Shader Engine – still equals four, but the volume of each has increased. Thus, the chip’s front-end, which performs triangle rasterization, becomes its bottleneck.
HBM is a joint project between AMD and SK Hynix that aims to solve both problems through two technologies: a) packaging RAM chips into a three-dimensional “stack”; b) connecting them to GPU via a silicon substrate (interposer).
As for multilayer packaging of microcircuits, conductors that allow connecting two silicon chips located one above the other are no longer something surprising. Hynix's innovation lies in conductors passing through several layers of silicon - TSVs (Through-silicon vias).
Silicon interposer, surprisingly, is a less sophisticated technology. This “gasket” is manufactured using standard photolithographic equipment, but this does not require advanced technologies - a standard of 65 nm is sufficient. Silicon itself here only serves as a substrate in which copper connections are laid. Then, multilayer HBM and ASIC Fiji chips are installed on microscopic metal balls (microbumps) formed where the connections exit to the surface.
The HBM implementation for Fiji-based products includes four 1GB "stacks" connected to GPU 4096-bit bus. The chips operate at 500 MHz with DDR technology (1 Gbit/s per line). Thus, the resulting interface bandwidth reaches 512 Gbit/s.
Specifications Radeon Radeon R9 Fury X
|
||||||
Chip
|
||||||
Frequencies
|
||||||
Memory
|
||||||
Interface and TDP
|
The Fury X comes with a fully functional Fiji chip, and has an official TDP of 275W. The upper limit of the clock speed GPU is 1050 MHz. Since the HBM memory chips are mounted on a common substrate with GPU, all that is on the outside of the printed circuit board is the power supply and I/O ports. That's why the Radeon R9 Fury X has surprisingly modest dimensions for a video card of this class, even more modest because AMD used a compact water cooling system instead of an air cooler.