Slashdot Mirror
AMD Radeon R9 Nano: 6 Inches Of High-Priced, High-Performance Graphics
Vigile writes: Back when AMD announced it would be producing an even smaller graphics card than the Fury X, but based on the same full-sized Fiji GPU, many people wondered just how they would be able to pull it off. Using 4096 stream processors, a 4096-bit memory bus with 4GB of HBM (high bandwidth memory) and a clock speed rated "up to" 1000 MHz, the new AMD Radeon R9 Nano looked to be an impressive card. Today, PC Perspective has a review of the R9 Nano and though there are some quirks, including pronounced coil whine and a hefty $650 price tag, it offers nearly the same performance as the non-X Radeon R9 Fury card at 100 watts lower TDP! It is able to do that by dynamically adjusting the clock speed from ~830 MHz to 1000 MHz depending on the workload, always maintaining a peak power draw of just 175 watts. All of this is packed into a 6 inch PCB — smaller than any other enthusiast class GPU to date, making it a perfect pairing for SFF cases that demand smaller components. The R9 Nano is expensive though with the same asking price as AMD's own R9 Fury X and the GeForce GTX 980 Ti. Readers have also submitted links to reviews at Hot Hardware and Tom's Hardware.
Most SFF cases I've seen don't have two slots for the video card. At best you're going to be covering up your only other expansion slot with this card. Most SFF power supplies aren't set up to feed 175W cards either. I can see this being useful in a specialist setting, like a gamer STB, but as a drop in replacement for people with SFF machines it's going to be problematic.
I read the internet for the articles.
The 28nm is frustrating, yes. But AMD and nVidia seem to both be suffering equally from a bad case of "not being Intel, lord of the fabs". It's interesting to compare their strategies for making the most of their situation.
nVidia: Pushes closed gameworks devtools with so much tessellation that triangles become smaller than pixels. Hampers performance on nVidia cards, severely hampers performance on AMD cards. Gameworks contracts stipulate that nVidia has executive control over the final shipped code.
AMD: Bets big on hardware capable of asynchronous computing going back to GCN 1.0 (2012). Creates proof of concept API (Mantle) used in a handful of games to show it is faster than DirectX 11. Microsoft later assimilates all the key ideas of Mantle into the DirectX 12 specification.
The result: AMD users have to spend 5 minutes googling and turning down tessellation settings to below placebo-levels to get acceptable frame rates on some current titles like Witcher 3. nVidia users will have to buy new cards to run DirectX 12 titles natively; their current cards can emulate async compute successfully but cannot harness any of the performance benefits.
The difference: While they have both been stuck on 28nm for years, AMD spent their time and money on making structural hardware and API design improvements that would pay dividends later (not just in async compute; AMD has now shipped a product with HBM, another promising technology). nVidia spent their time and money on rent-seeking profit schemes that can best be described as hurting their customers less than the competition.
Even though AMD and nVidia are basically tied for performance and value now on current titles, AMD cards have an interesting feature set. It reminds me of the GeForce 5800 vs. the Radeon 9700, where the major advantages of team red didn't become apparent until a year or so after release when DirectX 9 games ran slower and looked worse on nVidia's cards. I'm definitely interested in how this will play out with DirectX 12 games set to release in under a year (like the new Deus Ex).