•  News
  • 50.25% (Neutral)
  • Member Topic Starter
AMD's next-generation Radeon is set to debut later this fall and rumors about the upcoming chip have begun to leak out into the wild. While unconfirmed, the specs are in line with our expectations and what AMD has previously said about the chip. Hawaii won't be a fundamentally new architecture, but a refined approach to the technology already baked into Graphics Core Next (GCN).

Specifically, the new chip is supposedly built to GCN 2.0 standards with support for DX 11.2 (this may already be present in current Radeon hardware), four raster engines (up from two), a maximum of 2,816 shader cores, a 900MHz clock speed, temperature-based Turbo modes rather than a current-based model, and a 430mm sq. die. So what's that mean compared to the Radeon 7970? First, we can assume that whatever tweaks went into GCN 2.0, they should drive performance per watt up by 10-20% over the current generation of hardware. That's a reasonable projection based on what we've seen AMD deliver from previous products and in-line with the kind of optimization you typically find in a second-generation architecture.

Pictures of the new Volcanic Islands GPU are hard to come by, so here's a regular volcano, courtesy of Wikipedia

One thing we've heard that 3DCenter doesn't mention is that AMD has stuck with the 64 cores per CU approach. This has implications for Hawaii's compute capability, where keeping a smaller number of cores per Compute Unit allows for more efficient resource utilization. At a theoretical max of 2,816 shader cores, the new chip is 22% larger than the Radeon 7970, which, combined with the performance increases baked into the second generation, should allow the new family to hit significantly higher performance rates. Memory controller efficiency is also reportedly improved, but the bus will remain at 384-bits. RAM speeds have not been disclosed. The GCN family has generally had strong bandwidth, so this should be no issue.

"Titan-izing" the GPU

If these specs on Hawaii are accurate, they suggest that the AMD took a page from Nvidia's book and opted to "Titanize" the graphics design. It's not hard to see why -- Nvidia's two-pronged solution paid off big time for the company in 2012 - 2013, with a strong GK104 launch in 2012 and the benchmark-breaking Titan shipping in early 2013. That card then trickled down to become the GTX 780, offering much of the same performance, but at a significantly reduced price. While NV doesn't move many of either card in either price bracket, the profit margins are likely to be considerable.

We should note that, without information on prices, there's no word on how aggressively AMD will position the chip, but this approach should work well for challenging Kepler and the next generation of NV hardware. Graphics problems tend to be parallel, Titan provided that boosting the number of cores had a significant impact on frame rates. Technologies like antialiasing are an easy "soak" for excess graphics capability -- if a chip packs more horsepower than your monitor can display, ramp up the detail settings and voila -- problem solved.

One reason we expect Hawaii will be a fairly modest departure from current GCN hardware is that it's in AMD's best interests to present a unified front for game developers. The more GCN 2.0 resembles the PS4 and Xbox One, the easier it will be to shift games to run on AMD's PC GPUs, and the better performance consumers can expect. That's not to say that AMD is going to slam on the brakes of GPU development to keep technology in lockstep with the Xbox One or PS4, but we suspect that the company built the GCN architecture according to design principles it intends to keep for future generations.

My two 670's are waiting to get the scooch down the bench to my old computer (pending reviews and benchmarks for the 9000 series). If the 7990 is anything to go by, the only game that I will not be able to play at max settings will be Batman. Pretty exciting stuff.

Joel H

I hope you aren't wedded to the "9000 series" name.



44 CUs is almost perfect gpu utilization when you figure the actual gaming performance difference (not counting compute-intensive perks) between 7970 and 7950 (or gk104 for all intents and purposes considering 256 sfus really make it 1792 ops anyway) per clock when accounting for their bandwidth requirements and excess...That is to say somewhere around 1877sp per 32 rops is exactly what they would want.

900mhz is perfect for a process that was made for a nominal power/performance/yield voltage of .85-9v and has attributes of approx 1ghz/1v.

fwiw, 900mhz and 2816sp would require 5940mhz on a 384-bit bus. 2560 at 1ghz would be 6ghz.

This sounds very un-ati-like.

I was very much expecting 40-42 (36 pro) units and 7ghz (6ghz pro) ram because of the clockspeeds 28nm is capable (to keep die size down) and available bw in the said power envelopes this gen, but this would be better for power consumption. Clockspeeds would be lower, but it should yield well at those clocks and consume less power considering the trade-off of around 10% clockspeed for 10% efficiency (that would be fairly linear with sufficient bw) and around 6mm-7mm2 per CU.


They can cal it 3CP0 as long as it works well.

Joel H

There is no "perfect" here. Power consumption is governed mostly by defect management in this day and age, combined with leakage characteristics. Memory bandwidth efficiency is more important than sheer bandwidth.