AMD Polaris Lineup: RX 460/470/480


The Philosophy

Why the make the RX 480: 16% of PC users invest in high end GPUs ($300+), so 84% use less than $300 GPUs – RX 480 hits that target market well, and aims to gain market share with lower end GPUs that have a substantial amount of power. Also, with lower TDP that means more PCs able to use this card, so little restrictions on user base.
Notebook gamers compromise heavily as you either have to choose performance, temperatures, size or weight – not all of them. With 5 million sales current total, there is room for a lot more laptop gaming, and AMD plans to provide that space.
VR is quite hard to reach price wise, with an Oculus setting you bad £400+, and the PC you’d need coming in at £700+. Only 1% of all PCs are capable of VR, although this is a slightly bloated figure as this included all workstation/office/desktop/ultrabook/non-gaming oriented devices, but still makes a good point. Game devs are currently developing for smaller market, as 80% of game devs missing revenue targets for VR games, which is a huge problem in terms of companies not being able to make money on this platform, meaning there is less inscentive to make VR games.

AMD’s Challenge 1: How do we make the next 250 million gamers PC gamers?.

-> 95% of gamers run at 1080p (or lower)
-> 84& in $100-300 GPU market.

Challenge 2: How do we bring real VR to the first 100 million consumers?

->1.43 bn PCs worldwide, only 13 million PCs qualify.
-> Palmer Lukey “People want VR before they can afford it”
-> Reasons to not VR: “Cost”, “Headset bulky”, “VR not developed yet, no situations to use”, “Not enough difference between standard and VR”

Challenge 3: How can we get the best and widest range of content to PC gamers as fast as possible?

-> 87% of all VR systems are AMD based (inc consoles)
-> 5 TFLOP chip = current console power + 2 TFLOP chip chip = last gen console
-> 5 TFLOP chip is what developers are developing for for the next 2-3 years.

Radeon RX series

-> Console-class GPU performance for every notebook, console-like life cycle, 3+ year life cycle.
-> Increabible VR experiences for 100 million conusmers
-> Empowerment to deliver great content on Day 1 to both consoles and PCs.

Polaris 10:

-> 36 CUs
-> 256 bit memory GDDR5
-> DP1.4 HDR
-> 4th Gen GCN

Polaris 11:

-> 14 CUs
-> 128bit GDDR5
-> DP 1.4 HDR
-> 4th Gen GCN
14 CUs TBC
7GBps memory
75W TDhP (no power connector)
End of July
4/8GB GDDR5 256Bit
150W TDP
‘VR Premium’ (HTC Vive/Oculus)
Hitman/Quantum Break/Forza @ 1440P does well.
29th June

7GB/s mem chips
110W TDP
Mid July

New Overclocking GPU tool which will be built into a soon-to-be-released Crimson Driver, and has a simple desgin with new features like: Dynamic GPU Frequency curve, other curves, power limit control.
->HDMI 2.0b Freesync supported 4K 60Hz
-> Improved gemoetry processing
-> Full Async compute
RX 480 Core:
-> 14nm
-> 4 ‘True’ Async compute engines
-> 2 HWS
-> 36 CUs
-> 4 Geometry units
-> 32 Pixel Outputs
-> 144 Texture Units
-> 576 32b Load/store units
-> 2Mb L2 cache
-> 256 bit GDDR5
-> Crossfire = DX9/10/11
-> RX 480 supports up to 4 way
-> DX 12 can support as many GPUs as dev wants
-> 14nm
-> 4 Async compute engines
-> 2 HWS
-> 16 CUs
-> 2 Geometry units
-> 16 Pixel Outputs
-> 64 Texture Units
-> 576/2 32b Load/store units
-> 1Mb L2 cache
-> 128 bit GDDR5
-> Crossfire = DX9/10/11 2 GPUs

1.95x performance available thanks to explicit multi-GPU support, but main demo on stage running at 1.83x scaling.
GTX 1080 was rendering less snow thanks to driver, apparently to aid performance.

Enhanced Geometry Engines
Allow for Primitive triange discarding to happen much faster, as it’s now 4-8 discard per clock cycle (4 right now, possibly 8 soon). This is great for VR as game devs can’t use bump mapping (an old-school cheat way to make a flat surface look like it is bumpy), have to use tesselation (on the fly maths to add complexity to flat objects).

15% IPC improvements thanks to GCN 4.0

Willing to open source code and GPUOpen platform to make development better/easier