AMD Fury X "Fiji" Voltage Scaling

[HumanSmoke]

But expecting aftermarket coolers and overclocking off the bat with a new arch is stupid,no one does it, it tests warranties and removes a respin opportunity ,now tell me nvidias different.

You mean like the GTX 980 Ti being launched on June 1st, and aftermarket cards with showing up immediately? In fact, the world's first 980 Ti review - from Zol.com - actually featured an AIB card with a nominal 11.5% overclock ( the Inno3D iChill X4 Ultra). Even the UK market had non-reference cards being paraded around on forums within days of launch including OC'ed waterblocked variants.
I'd say this differs quite markedly from the Fury X launch personally, but my perception is probably coloured by having a working memory.

 

[TheoneandonlyMrK]

New arch?? Not day one it had been done in a titan and stock coolers initially you reckon they would have let partners do their own 980ti cooling without fury x?.

Google ahoy

 

[HumanSmoke]

New arch?? Not day one it had been done in a titan and stock coolers initially you reckon they would have let partners do their own 980ti cooling without fury x?.
Google ahoy

Ah, so your excuse is that the Titan X is the "new arch" ? Well, OK, Titan X launched 18th March. Titan X overclocked cards launched the same day - including the Hydro Copper (which was available for pre-order on Titan X's launch day with orders being filled a month after initial launch....or a little less than the time between the Fury X launch and today) .....which again features a 11.5% nominal overclock out of the box, and Zotac's hybrid cooler version.

 

[TheoneandonlyMrK]

Aha so u fail at reading I said much earlier im not that bothered grow up.

 

[HumanSmoke]

Aha so u fail at reading I said much earlier im not that bothered grow up.

Awesome! You challenge people to find flaws in your argument, and when they comprehensively do so, you regress to the ad hominem attack.

Just as an aside, anyone who writes like you do probably shouldn't be criticizing the comprehension skills of others YKWIS?

 

[TheoneandonlyMrK]

Bitchy arnt ya lolz

 

[Aquinus]

No offense, but @HumanSmoke is actually defending his point well with links and well formed English (as he usually does.) Can you say the same, @theoneandonlymrk ? HS has evidence and logic on his side. He's not bitchy, he's calling you out which I fully support.

 

[TheoneandonlyMrK]

They still are not comparable Fiji is too different and was an at risk production run and imho amd chose wisely to hold something back for Christmas ,
also my post has gone in which I stated I wouldn't buy a fury x or 980 times because neither is good enough an upgrade to me.

 

[Aquinus]

also my post has gone in which I stated I wouldn't buy a fury x or 980 times because neither is good enough an upgrade to me.

For a reasonable price, sure. I think the point was that you can get more performance if you're willing to pay for it and that nVidia does regularly update their architecture and tend to do it for most of their performance level GPUs where AMD has been a little slow to introduce substantial changes to GCN.

I'm also concerned about HBM latency and if it's really being overcome with a wider data bus. Testing my own 390, I can say that from 3Ghz effective to 6Ghz that memory scales almost linearly with frequency. I also determined that for a 50% reduction in VRAM clocks, there was only a 33% decrease in performance IIRC (I need to re-install office, it broke with the Win 10 upgrade.) However, I think it shows how much time the GPU cores are spending waiting for data. In all seriousness, it makes me wonder if those 7 and 8Ghz effective GDDR5 DIMMs would continue to scale linearly and if memory clock increases performance synergistically with the core clock (assuming AMD's GPU IMC can handle those speeds.)

I would like to collect more information so I can plot it out before I come to any solid conclusion but, I think that unlike CPUs, GPUs have a much more simple and smaller cache and isn't as effective as CPUs at not hitting memory as often. Considering data-throughput of GPUs, it makes sense that there isn't a clearly defined upper or lower bound unlike CPUs unless dealing with very large sets of data (which still results in marginal gains most of the time.)

A secondary effect of decreasing memory clock is that the cores started running a little cooler and power draw on both the cores and the DRAM decreased (minimally for DRAM but it doesn't consume much to begin with.) Probably because they were spending more time waiting due to the lowered VRAM clocks, less logic was actively running.

AMD and nVidia might realize that wider but slower memory will require better data caching to optimize performance because GPUs aren't doing anything while they're waiting for data. It might get a bunch at once and can start processing away but how quickly will it be done with that data before it needs more and has to start waiting again? It's just another question that we (and maybe even AMD,) doesn't have the answers to. I do think that memory clocks play a much larger role in GPU functioning that many believe and I do think it's related to both width and frequency unlike modern days CPUs which have very sophisticated caching which mitigates unneeded reads/write to system memory.

 

[TheoneandonlyMrK]

Amd and Nvidia have tools for monitoring and logging all aspects of a gpus performance.
It's these that they and devs optimize performance with, I'm sure a download SDK is available.

Awesome! You challenge people to find flaws in your argument, and when they comprehensively do so, you regress to the ad hominem attack.

Just as an aside, anyone who writes like you do probably shouldn't be criticizing the comprehension skills of others YKWIS?

Just as an aside Is every conversation an argument to you.
I no longer get involved with Your debates for that reason, arguing ,,because I can't be bothered.

You also tend to jump on the edges of any of my comments while insinuating I'm a fanatic in some way yet inevitably dodging my main point.

Like this ,my point is I wouldn't buy now but if I did I would likely get a proper block on it as a 120 rad and partial block limits performance in both max and noise terms as do traditional coolers.



And that's for a 980ti or fury x weva's

And that Amd played it fair and smart its not a terrible card for the money and its brought about volume production of something or other sooner than other companies would like yay progress.

 

[HumanSmoke]

Just as an aside Is every conversation an argument to you.

argument (ˈɑːɡjʊmənt) : a process of deductive or inductive reasoning that purports to show its conclusion to be true. As in:

Awesome! You challenge people to find flaws in your argument, and when they comprehensively do so, you regress to the ad hominem attack.

I suppose I could have used "hypothesis" or "theory" in place of "argument", but that would have imbued your original statement with much more gravitas than it deserved.

I no longer get involved with Your debates for that reason, arguing ,,because I can't be bothered.

No, I suspect you just like flinging half-assed facts around, but don't like having them shown up for what they are.

Like this ,my point is I wouldn't buy now but if I did I would likely get a proper block on it as a 120 rad and partial block limits performance in both max and noise terms as do traditional coolers.

No. Your original post was divided into two parts. The first part was your personal opinion on purchasing hardware. Your usage and reasoning are particular to your needs, and as such, I wouldn't presume to know your requirements better than you do yourself.
The second part of your post made two definitive statement - namely that allowing aftermarket coolers and overclocking on launch day is stupid, and that no one does it, and challenged your readers to show how Nvidia's launches were any different. I strongly disagree with the first point you made, demonstrated the second to be false, and conclusively answered your question....all without recourse to name calling and hurling insults -which is your reaction to having facts presented to you and your question answered.

Not at all just unrealistic foot stamping ,go to it ,me personally I think id get a proper Waterblock oor just hold on cos none if them appeal to me I'll hold with my 7970 for now.
But expecting aftermarket coolers and overclocking off the bat with a new arch is stupid,no one does it, it tests warranties and removes a respin opportunity ,now tell me nvidias different.

I don't judge personal requirements of people I don't know. I will however, debate fact.

And that Amd played it fair and smart its not a terrible card for the money and its brought about volume production of something or other sooner than other companies would like yay progress.

No question. It isn't to my taste ( I prefer to choose my own watercooling components, the restrictions AMD have placed on the cards are too restrictive for my liking, and the pricing sucks - the package weight also makes it uneconomic to source from overseas), but it no doubt fills the requirements of many others, and has lifted AMD's profile- which is a prime requirement for a healthier and more competitive market for the consumer.

 

[Vayra86]

For a reasonable price, sure. I think the point was that you can get more performance if you're willing to pay for it and that nVidia does regularly update their architecture and tend to do it for most of their performance level GPUs where AMD has been a little slow to introduce substantial changes to GCN.

I'm also concerned about HBM latency and if it's really being overcome with a wider data bus. Testing my own 390, I can say that from 3Ghz effective to 6Ghz that memory scales almost linearly with frequency. I also determined that for a 50% reduction in VRAM clocks, there was only a 33% decrease in performance IIRC (I need to re-install office, it broke with the Win 10 upgrade.) However, I think it shows how much time the GPU cores are spending waiting for data. In all seriousness, it makes me wonder if those 7 and 8Ghz effective GDDR5 DIMMs would continue to scale linearly and if memory clock increases performance synergistically with the core clock (assuming AMD's GPU IMC can handle those speeds.)

I would like to collect more information so I can plot it out before I come to any solid conclusion but, I think that unlike CPUs, GPUs have a much more simple and smaller cache and isn't as effective as CPUs at not hitting memory as often. Considering data-throughput of GPUs, it makes sense that there isn't a clearly defined upper or lower bound unlike CPUs unless dealing with very large sets of data (which still results in marginal gains most of the time.)

A secondary effect of decreasing memory clock is that the cores started running a little cooler and power draw on both the cores and the DRAM decreased (minimally for DRAM but it doesn't consume much to begin with.) Probably because they were spending more time waiting due to the lowered VRAM clocks, less logic was actively running.

AMD and nVidia might realize that wider but slower memory will require better data caching to optimize performance because GPUs aren't doing anything while they're waiting for data. It might get a bunch at once and can start processing away but how quickly will it be done with that data before it needs more and has to start waiting again? It's just another question that we (and maybe even AMD,) doesn't have the answers to. I do think that memory clocks play a much larger role in GPU functioning that many believe and I do think it's related to both width and frequency unlike modern days CPUs which have very sophisticated caching which mitigates unneeded reads/write to system memory.

The problems with early Maxwell (970) showed how scheduling in the driver can manipulate the way memory is used. We also know this to be true; the driver determines much of how data is fed to the memory. But I do think that comparing GDDR5 to HBM is not apples/apples. GDDR5 scaling linearly could be explained by a number of different reasons, the most straightforward one being that in most implementations, GDDR5 is coupled with a well balanced GPU. The vast majority of cards especially above mid-range is such a well-balanced solution. The fact that your performance scales linear with clock % is simply because VRAM is a limiting factor, and that is because it is inside a product that is well engineered and well balanced. Reducing core clocks would achieve a similar result.

If anything could be deducted from the Fury X results with respect to memory then it is probably that the higher latency will crush the top-end of framerates for games that feed off VRAM a lot. On the other hand, the wider bus will allow a larger framebuffer to get processed faster, explaining the strong showing of Fury X @ 4K. This principle would also explain why Maxwell is capable of achieving higher 1080p framerates overall and keeps scaling as overclocks go up, while Fury can't gain much from overclocking across all resolutions.

 

[the54thvoid]

Not at all just unrealistic foot stamping ,go to it ,me personally I think id get a proper Waterblock oor just hold on cos none if them appeal to me I'll hold with my 7970 for now.
But expecting aftermarket coolers and overclocking off the bat with a new arch is stupid,no one does it(1), it tests warranties(2) and removes a respin opportunity(3) ,now tell me nvidias different(4).

1) AMD haven't done it for a while. They screwed themselves with the 290X by NOT allowing AIB's to add better coolers because they thought their own jet blower was sufficient (or they pushed to release faster than AIB's knew, either way - AMD made a boo boo). Nvidia allowed GM204 cards to be released at launch with aftermarket coolers. GM204 is not GM 200, besides, GM200 came out after. As for not expecting overclocking off the bat cos it's stupid - no, your position is simply not supported by enthusiasts and product designers.
2) Any voltage addition voids almost any warranty period. We all know that but most enthusiasts that like their tech overclock. You still RMA cards as the over volting is normally software based (Precision X, Afterburner etc). Even well established cards state over volting void warranties - hell - my Kingpin says it. So it's not a valid argument to state that "expecting aftermarket coolers and overclocking...tests warranties" as it ALWAYS voids warranties.
3) Doesn't affect respin opportunity at all. Not even sure what you mean. Tweaking the process and delivering a more efficient version of the same arch has nothing to do with what overclocking' is expected at release. All a respin is designed to do (in my own understanding) is deliver the same chip with higher efficiencies and better margins for the company.
4) Nvidia are not different by much but they did allow the 980 (as @HumanSmoke stated) to be released at launch with custom coolers and overclocks. However, Nvidia do lock down a lot of overclocking but only by means of voltage and power limits. The 980 range however was (and is) an overclocking demon. My best guess is that AMD pushed Fiji to it's limits with it's 1050 Mhz clocks and it doesn't have any headroom. That plus the process is immature (involving HBM and interposers) so yes - they don't want things breaking.

But taking all those points - Lisa Su did state (or her colleagues did) that it was an overclockers dream. The head of AMD no less said it. But it's not an overclockers dream at all. Hell - the 7970 was (is) awesome for it (I got my old Powercolor to 1300Mhz). To me it seems AMD got a bit lost after Tahiti (79xx) and Hawaii wasn't a great overclocker and Fiji just plain sucks at overclocking. Yes, Hawaii was good (but got eclipsed by full Kepler), and Fiji is also a great card (but get's thumped by Maxwell for the same price).

And that's the sting. I can buy an Nvidia card that is as cheap or cheaper than AMD's top line and the Nvidia card is faster. Since when was Nvidia cheaper than AMD? Think about it - it's been a while.

And one last thing (Columbo style) your 7970 is slow compared to what's out these days. I know because I had two.

 

[terroralpha]

silent air cooler , OC addict , dead silent pc , high end gpu , = odd combo what you doing that requires that ridiculous mix of mission impossible bar filling forums with amd hate.

Ive had significantly better then that water cooling for years, speed controlled fans and pumps all with profiles to be quite or loud and cool but both at the same time then oc = NONSENSE.. get real.

i want a silent 5ghz quad gpu pc that can do vr like feckin real while using a tea cups worth of power and giving off no heat too but its not gonna happen this year.

the gigabyte GTX 980 Ti was on back order, so I got an MSI GTX 980 Ti LE "lite edition" instead. came out to only $610. arrived today. it's exactly like the MSI 980 Ti gaming, but with no factory OC. i spent the last hour tuning it. the fastest stable OC i could get with no artifacts, tearing, stuttering, etc, was on 285MHz on the core and 400MHz on the memory. and this isn't even a very good sample, most GTX 980 Ti cards can do 500MHz extra on the memory. i got a fire strike score of 17336. i did a search and no Fury X that FutureMark has on record comes even close to my score. check out the link below

http://www.3dmark.com/compare/fs/5629552/fs/5476446/fs/5586616

and the best part? the PC is super silent. i can't hear it at all from 5 feet away. so yes, everything I said is possible.

 

[wege12]

Any word on when this voltage control software for the Fury X might be released to the public? It seems a little suspicious that it hasn't been released yet after advertising to the world that its fully functional...

 

[nashathedog]

Any word on when this voltage control software for the Fury X might be released to the public? It seems a little suspicious that it hasn't been released yet after advertising to the world that its fully functional...

This is a copy paste "Unwinder has said it was due to be released in 4.2.0 along with the Lightning Ti. Not sure if his plan has changed, it was only last week he said it though."

Maybe he can deny or confirm for us?

(Just realised that is on Afterburner this is about Sapphire's)

 

[wege12]

(Just realised that is on Afterburner this is about Sapphire's)

Well, if they both have the same functionality for the Fury X, I'll take what I can get

 

[Figus]

That's for the full system, not the card itself. Keep that in mind.

Still unbelievable... i got a 4670k@4.2ghz with 5 hdd (3 ssd and 2 standard) a fury and in benchmark looping my sistem will never goes over 330W (Usually is under 290W in gaming)

 

[nashathedog]

So we finally got voltage control on Trix,

Any good experiences?

I can run Firestrike with a core clock of 1100 on my Sapphire Fury Tri-x on stock voltage
http://www.3dmark.com/fs/5681651

I made several attempts to run it at 1128 with various voltages and it failed every time.

I gave up in the end, From what I saw it's a waste of time.

My memory won't even increase from 500 to 501 without blue screening

 

[Yuriadc]

Wizzard if you could post the stock dpm 7 vddc in your sample i could compare to mine. Mine vddc 1.68 and asic 71.7