AMD Radeon R9 300 Series MegaThread: FAQ and Resources

Page 55 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar


What about memory? +100 MHz? Also I want to see if I can get a stable clock on the core at +100 MHz or more if I can, but I may have to tweak the voltages to get that. any ideas?
 


I thought fury x's memory was locked from the factory? Though i did hear about some after market programs. I'd be wary about ocing hbm, word on the street is you won't get that high. Maybe just the core?
 


Could also be pseudo demand. Hold back supplies increase demand etc, or it outsold what they predicted? I mean i'm assuming that's the major market as opposed to higher end gear. And with nvidia's .5 gb l2 cache debacle, the cards may have fallen favorably in amd's hands. Wonder for how long.
 
AMD's stock just had its largest single-day increase in value since going public in 1979!

http://www.profitconfidential.com/stock/amd-stock-is-advanced-micro-devices-inc-a-dollar-6-stock/
 


must be ps4k and nintendo nx news. got to give it to amd for locking up the entire "console" market. hope it keeps things rolling for them.
 


Ahh that thought has totally passed for me! It must be. And it also renews my excitement in the Nintendo NX.
 
Just my opinion but i think that make amd stock jump so much was those deal with China. That and their positive outlook for next quarter. Though i don't know if those deal with china will be good in the long term. because this is china that we talk about. Qualcomm also have somekind of licensing issues with China. If i'm not mistaken last year Qualcomm accuse china to pay less than what they should and suspect that china goverment did help their local company to hide the real amount of money that need to be paid to qualcomm. So i wonder if AMD will end up in similar positon in the future.
 
It had mostly to do with improved quarterly financials coming in a preliminary statement that also disclosed (as renz469 mentioned) a promising joint venture with the Chinese firm Tianjin Haiguang Advanced Technology Investment Co, which will be to produce SoC products for large servers in china. I guess we'll have a clue if China comes up with another super computer 🙂

Below are the SEC filings in case any one is interested. They're the best source over any news reporting on them.

Current Report on Form 8-K (essentially a cover sheet): https://www.sec.gov/Archives/edgar/data/2488/000000248816000118/a20160421q116earnings8-k.htm

Official News Release: https://www.sec.gov/Archives/edgar/data/2488/000000248816000118/q116991.htm

First Quarter Results: https://www.sec.gov/Archives/edgar/data/2488/000000248816000118/q116992.htm
 
Not sure about supercomputer but in this area intel also very aggressive right now with their xeon Phi push. The fastest super computer to date still is china Tian He that use intel xeon phi. They did not win the power efficiency stuff but if i remember correctly it was also have quite a gap from no. 2 super computer (titan) that using nvidia K20X. heard that intel might win the contract because they sell xeon phi so cheap. I once saw xeon phi deal (mention in phoronix) as low as $400+ where as nvidia tesla and amd firepro still costs thousands. and if you look at top500 list (the latest list was november 2015) the number of super computer using amd firepro is almost non existent. Amd launch S9150 (which is easily superior than nvidia GK110/210 and intel phi) back in 2014 and some people predict that the top ten of top500 list in 2015 will be dominated by amd accelerator but in the end it did not happen. It seems to me that most HPC client are waiting for Pascal and intel KNL. only few that cannot wait for those two go with AMD S9510 recently.
 


That's because of software environment. In the HPC field AMD's firepro is best hardware wise but it requires Opencl programming. AMD wants to deal with this with its Open Compute(Boltamann) initiative for CUDA and Zen apu for X86. I'm talking about the HBM2 apu which will be used for HPC not consumer apu, of course.
 
In theory they might have the best hardware. But in the end it will depend if they can utilize all that raw power into real world performance. This is one of the problem with GPU. this is the fastest supercomputer using AMD accelerator in top500 list (rank 162 as per november 2015 list):

http://www.top500.org/site/50458

Look at it's Rpeak (theorical performance) vs its Rmax (real world performance).
 


Totally agree with you. CUDA is made for Nvidia gpus. It gets the best performance out of its hardware. And opencl is for all. You need to code it for specific hardware to get CUDA like performance which isn't done as it defeats the purpose of 1 program fits all. I don't think this a problem AMD can solve soon. That's why I think only Zen can save AMD.
 
With opencl you can also tweak the code down to architecture specific. but you will sacrifice portablity doing so. Even for CUDA optimization for different nvidia architecture will be different. But even if the code is not portable the idea is for everyone to use one API. So there is no need to learn specific API for each hardware. Though for professional stuff they want everything to be done as fast as possible. As long as you can get the job done they will overlook stuff like proprietary vs open.
 
Reading the data, if accurate, it looks like a decent step up from the 7970/280X and 285/380X cards, but a shy away from reaching the 390/390X.

Now, if they can do that with less than half the power, it would be cool. Plus, they don't mention anything else, like clocks or actual chip.

Salt recommended, indeed.

Cheers!
 


Whoa! They predict a Terabyte-per-second of memory bandwidth! That's pretty insane.
 


There is no prediction that is the spec for HBM2:

http://www.anandtech.com/show/9969/jedec-publishes-hbm2-specification

HBM2 will go from 512GB/s to 1TB/s worth of memory bandwidth and will probably be in the replacement for Fury X and the Titan of whatever for nVidias Pascal. However neither are expected till 2017.
 


. . . Even more insane. I wonder whether they've done any testing on this kind of stuff to see whether swapping that much info actually can happen. I'm not aware of an oscilloscope or similar device that can take measurements that fast. It's probably a theoretical number based on a calculation, and we've all seen how those play out over the years. It's the kind of thing that's given rise to the growing attention to distinctions between theoretical max performance and real-world performance. It's just that we have no way to test real-world performance of the memory because it isn't the bottleneck, and because we don't have a few million dollars lying around to get our hand on equipment that can reliably measure these things.
 
1TB/s is nothing in reality. Lets consider that fiber optics have been able to push beyond that for years:

http://www.extremetech.com/extreme/192929-255tbps-worlds-fastest-network-could-carry-all-the-internet-traffic-single-fiber

That is roughly 32TB/s and that is not even the top limit. Every time they think they hit the limit to fiber they find another.

In the end will be latency. 1TB/s of memory bandwidth between the VRAM and GPU will allow the latency between them to be cut down and thus improve performance as there will be less time for the GPU to wait on what it needs to throw out.

It all has to be seen though. I am sure it will just sit there until the rest of the PC catches up. Considering nVLink and Intels 3DXpoint I don't think we will have to wait long for the rest of the PC to catch up to HBM in bandwidth, especially since HMC should also be coming out soon.
 


That fiber = :ouch:

Xpoint = can't wait!
 
Here's an actual review (with an 'r', not a preview) of the Radeon Pro Duo.
http://www.pcper.com/reviews/Graphics-Cards/AMD-Radeon-Pro-Duo-Review

It includes some nice modern abstract expressionism art, along the lines of a Jackson Pollock:

GTAV_3840x2160_PLOT.png
 


AMD has always had frame pacing issues. Was the only thing that really annoyed me when I had my HD7970Ghz that they took so long to sort it out. Their bane is still crossfire though and you can tell from this.

Then again the Radeon Pro Duo is a professional GPU, not for the consumers. Or so they say.
 
As usual with AMD, they will take at least 6 months to get drivers up to snuff with most games.

I mean, they have been slowly, but steadily improving the Fury performance. Just like they did with the 7970 and all the cards before it. I'd even say the 295X is still not close to it's top performance.

AMD should really increase their software engies amount. I am inclined to think they have little engies and too many managers, as big companies tend to have.

Cheers!
 
You must log in or register to reply here.

TRENDING THREADS

Latest posts

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom