[BitsAndChips]390X ready for launch - AMD ironing out drivers - Computex launch

[dacostafilipe]

This is all just wishful thinking.

I'm not making stuff up, it's listed as OEM on AMD's page + slides.

This looks like HD 8000 all over again with an incoming renaming.

 

[RussianSensation]

Do you not know that different sites use different methods for power measurements. Different games and different apps.

There is nothing unusual and different results are the norm. Every game stress your GPU in different ways. Heck, the same game will stress your GPU depending on the location and what is going on

The problem is TPU's Max chart is not a game/app measurement. Furmark is just a power virus. Pretty much most professional websites and most members on our boards have understood that Furmark is a waste of time to measure GPU's power usage since it's a synthetic test. IMO, the only power measurements that should count are real world programs such as games, rendering, distributed computing projects, etc. -- essentially programs and applications that GPU user use to get results that benefit them directly. Since Furmark provides no direct benefit to any user, it's essentially a worthless bench since no real world program can load all parts of the GPU and memory simultaneously like a power virus can. Furmark is even worthless today for GPU OC stress testing because AMD/NV incorporate GPU throttling based on the drivers. That means you would be better off running Crysis 3, Metro LL, GTA V, Far Cry 4, etc. to stability test your overclock. Furmark did serve a purpose in the past when driver throttling was not in place and it could quickly find an unstable overclock. Today, it's no longer reliable for this either.

Using Furmark to measure power usage is akin to using 3dMark to extrapolate real world gaming performance in next gen game engines (which of course 3DMark cannot do since no game AAA game is based on "3DMark game engine"). Both of these programs are useless for their intended functions, which is why we need real world power usage and real world game testing data for any type of conclusions but a lot of PC gamers won't move on with the times and acknowledged that what mattered in the past is no longer relevant today.

Ditto with full Tonga R9 380X vs R9 280X.
But no it's not a slam dunk and it's not a Maxwell killer.
Welcome back to the slow pace of GPU / CPU innovation.

Your statement is hard to grasp because it leaves no space for R9 290/290X unless you suddenly think R9 290 = 390 and R9 290X = 390X. So what is Fiji exactly, R9 395X2 dual Tonga XT? or is Fiji silicon all made up and it's nothing but Hawaii XT with HBM?

Furthermore, there is no confirmation at all what the specs are for retail R9 370/370X/380/380X/390/390X cards but you are so confident that the retail 380X is full Tonga based on what? Everything besides R9 390/390X/395X2 could be re-brands for the retail market or they may be not be. We don't have hard data to prove either case. We don't even know for certain if the retail 300 series will be called 300 or 400 series.

AMD’s OEM 200 series and 200 series lineup have also existed before in parallel and included some cards that weren’t part of the official retail lineup such as Radeon R9 255 and R9 275. Conversely, the OEM lineup also didn’t feature any R9 280 series cards that the retail sector had, which resulted in a big gap between Pitcairn and Hawaii for OEMs. This simply confirms that the OEM lineup will exist alongside the non-OEM discrete lineup and right now drawing direct conclusions that OEM SKUs = all retail SKUs or vice versa is premature.

GTX 760 retail: 32 ROPs, 256-bit bus, 2 GB of 6.0 GHz GDDR5
GTX 760 OEM: 24 ROPs, 192-bit bus, 1.5 GB of 5.6 GHz GDDR5

And then later they released yet another GTX 760 OEM to bring it up to retail specs (with 3GB of RAM, no less).

Just like 860M was both a Kepler and a Maxwell GPU, but then 960M is a re-badge of the Maxwell 860M. The OEM market is often completely different from what we get in the retail sector.

Either way, if we go back to the original rumours of Fiji XT being a 4096 SP, 256 TMU, 4096-bit videocard clocked at 1Ghz+, then the gap between that card and a full Tonga XT or Hawaii XT 290X would be way too massive for AMD to have 380X just a full Tonga and then 390X is a 4000 shaders monster that's 75-100% faster. GPU line-ups do not work like that.

We need to go back to the basics of AMD's product branding.

$370 HD6970 was 32% faster than a $240 HD6870 (that gen's mid-range) at 1600P.

$500 HD7970Ghz was 33% faster than a $380 HD7950 at 1600P.

$550 R9 290X was 33% faster than a $299 R9 280X (that gen's mid-range) at 1600P.

There is no 75-100% gap anywhere here. There is just no way 380X is a full Tonga XT with only 2048 SPs but 390X is a 4096 SP monster. It's not logical and AMD has never branded products that way.

The second part is AMD almost always aims to have its next gen mid-range to be as fast as last gen flagship, or extremely close to it. This is another strong reason why a 380X being a Tonga XT is a theory that hardly makes sense in the context of 380X vs. 390X naming scheme.

HD5770 (new gen mid-range) ~ HD4870/4890 (last gen flagship)
HD6870 (new gen mid-range) ~ HD5870 (last gen flagship)
R9 280X (new gen mid-range) ~ HD7970/7970Ghz (last gen flagship)
R9 380X is a new gen mid-range ==> should be at least as fast as R9 290X, something a 2048 SP Tonga chip would have no chance of accomplishing.

Right now R9 290X is a whopping 64% faster than an R9 285. That means there is no way for a Tonga XT to catch up to an R9 290X even if it's the fully unlocked version.

Some of the theories in this thread are so irrational, it feels like the closer we get to release date, people are going to conclude that R9 2xx series is faster than R9 3xx series....it's getting this bad.

 

[arandomguy]

There were concerns that the GTX 980 would be slower than the GTX 780ti leading up to launch.

 

[RussianSensation]

There were concerns that the GTX 980 would be slower than the GTX 780ti leading up to launch.

Maybe from people who don't pay attention to NV's GPU positioning or learn history of NV's GPU development. NV's real next gen mid-range has always been at least as fast as last gen flagship, no exceptions since GeForce 2. By this very definition 980 had to be at least as fast as a 780Ti, no questions asked. Most of us just weren't sure how much faster it would be: 5%-35%? The only reason this would change if 980 was GM206 and 980Ti was GM204, etc. As long as you follow the internal code-names (i.e., GM204 is a mid-range next gen) and if 980 = GM204, it would at least tie a 780Ti. By the same definition GP204 (next gen mid-range Pascal) would be at least as fast as the Titan X. NV follows this rule to the letter and has not deviated from this in 15+ years. You almost don't need to know the specs of Pascal GM204 make a very well educated guess that it should at least match the Titan X unless NV begins a new chapter in GPU building/branding.

If you follow AMD, it's actually very similar in that the next gen mid-range tends to be almost as fast or faster than the last gen flagship. I already provided 3 examples in the post above. By that definition, 380X ~ R9 290X and 390X would be 30-33% faster, roughly. But if the rumours are pointing to R9 390X being a card 50-55% faster than the R9 290X, that means by this definition 380X has to be faster than the R9 290X if AMD follows the same GPU replacement strategy it has used for the last 2 decades (much like NV does).

 

[arandomguy]

You missed the point.

Speculation is rampant, wild and often misrepresented (disguised) as facts leading up to these launches. History shows that. It isn't specific to this launch.

 

[RussianSensation]

You missed the point.

Speculation is rampant, wild and often misrepresented (disguised) as facts leading up to these launches. History shows that. It isn't specific to this launch.

Right, but ignoring the last 15-20 years of NV/AMD/ATI GPU development to make an exception is not logical unless some posters on these forums have strong data to believe AMD changed its strategy. You need to ignore the specs and focus on the trends then. 5770 at least tied 4870, 6870 came very close to the 5870, 7870/7950 beat 6970, R9 280X beat 7970 and is a barely slower than 7970Ghz. By that account 380X would be in 290X territory which rules out 2048 SP Tonga automatically. 380X could be Tonga if AMD introduces newer names like 390/390X (Hawaii enhanced) and 395/395X (Fiji). Otherwise, the performance delta between a 380X and 390X would be too enormous to make any sense of it. The alternative scenario is 390X is nowhere near a 4096 SP/256 TMU product and is instead barely faster than a 290X. But in that case, the other end of the speculation doesn't add up. Who in the world would use 512-640GB/sec HBM for a card barely 5-10% faster than an R9 290X? That makes no sense either.

All I know is the forum has never been so divided on any GPU launch. With this launch we have people on the forums literally convinced that R9 390X is nothing but R9 290X with maybe 5% more performance and HBM throw in for marketing reasons only with little benefit to power usage, and on the other end people expecting 50-55% faster performance at 4K and massive improvements in perf/watt (think 40-50% better than R9 290X). If R9 390X is anywhere close to 50% faster than R9 290X at 4K in a 290W power envelope, and if 380X can trade blows with a 980 at the 200-210W power level, there will be a lot of members on here eating crow.

 

[Erenhardt]

Even a refreshed Hawaii will leave a large gap between 390x and 380x, nevermind tonga.
I expect Fiji to be in more than 2 SKUs.

Possibly stretching Fiji through 390x (8/4GB)-390 (4GB)-380x(3GB)-380(3GB) and refreshing 290X as 370X could make w most of their R300 series lineup.

Ideally they should have 2 GPUs. One on the very top end of desktop hierarchy - Fiji. One for the mobile top end (mainstream sweat spot after failing on the binning for mobile). (Their mobile refresh announcement lacks any high end parts).

But, there are rumors about unsold amd inventory stuck on the market. Maybe amd wants to clear that before 14nm GPUs come out next year (?), because after that, those old parts will not find a new home.

So, to summarize, anything is possible.
Patience is a virtue.

 

[ocre]

RS, I never said we should go by futuremark or TPU, or toms.

Just saying, There is good reasons why different sites have different results.

I don't believe futuremark is very useful. Gpus throttle just because the exe is detected. Its not a good measure for anything other than what a card uses in futuremark.

A lot of times people will say this site is wrong or that site can't be trusted because they don't like the results, I think that comes from a lack of understanding. Its not only the games or apps that vary fro. Review site to review site, the locations and action of the benchmark run can make one sites results wildly different than another, even if they benchmark the exact same game and settings.

 

[jj109]

Even a refreshed Hawaii will leave a large gap between 390x and 380x, nevermind tonga.
I expect Fiji to be in more than 2 SKUs.

Possibly stretching Fiji through 390x (8/4GB)-390 (4GB)-380x(3GB)-380(3GB) and refreshing 290X as 370X could make w most of their R300 series lineup.

Ideally they should have 2 GPUs. One on the very top end of desktop hierarchy - Fiji. One for the mobile top end (mainstream sweat spot after failing on the binning for mobile). (Their mobile refresh announcement lacks any high end parts).

But, there are rumors about unsold amd inventory stuck on the market. Maybe amd wants to clear that before 14nm GPUs come out next year (?), because after that, those old parts will not find a new home.

So, to summarize, anything is possible.
Patience is a virtue.

I would be surprised if Fiji weren't put into at least two SKUs. HBM yields are probably low so there must be some interposer assemblies with only 2/4, 3/4 working stacks. 2048-bit and 3072-bit is still good enough for high-end, but the question is can the Fiji memory controller handle that kind of configuration?

It'll be really confusing if they had OEM R9 380 be 4GB Tonga and non-OEM R9 380 be Fiji, but Nvidia has done it in mobile with Kepler/Maxwell so anything can happen in the name of binning.

 

[Qwertilot]

I'm not saying what will happen but you can definitely see how they could end up with Tonga as the 380/X and a massively faster 390/X. To be honest I think it'd be a sensible use of their relatively limited resources.

Doing a huge prestige style 390/X is obviously worthwhile, especially with HBM to include and the chance to hit really big double precision compute performance.

The mid/upper mid range? I honestly don't see where they'd get the return on it right now. They were so very late to respond to the 970/80 that the majority of the people in that section of the market have either already got one of those or are going to wait until the huge performance increases due in the next gen.

With Tonga already existing due to the iMac contract you can see them deciding to just push it as their mid range card for a bit. They'll live.

If they're 6-9 months late to the 14/16nm/HBM 2 generation as well? Not likely to be at all pleasant.

 

[coercitiv]

All I know is the forum has never been so divided on any GPU launch.

HBM is a disruptive tech. Combine that with utter silence regarding product performance and you have the perfect catalyst to create forum controversy.

I have a feeling the new Macbook Pro might settle this before reviewers do.

 

[3DVagabond]

I'm not saying what will happen but you can definitely see how they could end up with Tonga as the 380/X and a massively faster 390/X. To be honest I think it'd be a sensible use of their relatively limited resources.

Doing a huge prestige style 390/X is obviously worthwhile, especially with HBM to include and the chance to hit really big double precision compute performance.

The mid/upper mid range? I honestly don't see where they'd get the return on it right now. They were so very late to respond to the 970/80 that the majority of the people in that section of the market have either already got one of those or are going to wait until the huge performance increases due in the next gen.

With Tonga already existing due to the iMac contract you can see them deciding to just push it as their mid range card for a bit. They'll live.

If they're 6-9 months late to the 14/16nm/HBM 2 generation as well? Not likely to be at all pleasant.

IF it's not that long until 16FF then I can understand not releasing a complete lineup. Other than that though they've had plenty of time to do more than one new chip. Their resources aren't that low. It's only been very recently that they've reduced their R&D below nVidia's. That might be because of the delay in new modes and they've got stuff waiting as much as anything else.

 

[RussianSensation]

RS, I never said we should go by futuremark or TPU, or toms.

Just saying, There is good reasons why different sites have different results.

I don't believe futuremark is very useful. Gpus throttle just because the exe is detected. Its not a good measure for anything other than what a card uses in futuremark.

A lot of times people will say this site is wrong or that site can't be trusted because they don't like the results, I think that comes from a lack of understanding. Its not only the games or apps that vary fro. Review site to review site, the locations and action of the benchmark run can make one sites results wildly different than another, even if they benchmark the exact same game and settings.

I agree with you. I was just pointing out how using TPU max against any other site is not really relevant. Otherwise, I agree with your point that we should consult many different sites to get a more unbiased viewpoint.

I honestly don't see where they'd get the return on it right now. They were so very late to respond to the 970/80 that the majority of the people in that section of the market have either already got one of those or are going to wait until the huge performance increases due in the next gen.

GTX970/980 only came out in Sept 2014 but 14nm/16nm cards of that level prob won't be out until late summer/fall 2016. That's still 15-18 months from now - a long time. 960 came out towards late January which means it's not even 4 months old. GM200 6GB hasn't even launched. 750/750Ti haven't even turned 1 year old, which means this generation isn't even half way yet. AMD needs new products because there are at least 4-5 quarters left in the tank before 14nm/16nm drop (Q3/Q4 2015 + Q1-Q3 2016).

If anything, I would argue the complete opposite of your point. If AMD is extremely tight on resources, they should have never done 390/390X $500-700 level cards and focused all their efforts on the $100-400 desktop and mobile dGPU segments - that's literally 80% of the entire GPU market. If AMD spent hundreds of millions on dollars just to make 2 good desktop chips in the form of uber expensive 390/390X/395X2 cards that only 5% of the market cares about, and everything else for the desktop and mobile is old crap, they are going to lose market share every quarter until 14nm/16nm. Also, given how R9 290/290X hardly sell at $240-280, this strategy would be an automatic fail and defeats the purpose of relaunching R9 300 series. Might as well do nothing, literally save their precious millions for 14nm/16nm and sell R9 200 series for another 15 months. I still don't understand the point of taking Pitcairn, Hawaii and Tonga and just re-labeling their boxes and replacing the number R9 2xx with R9 3xx. This sounds just too crazy to believe that AMD spend all its efforts on just 1 chip (and it's cut-down R9 390 version due to yields).

 

[3DVagabond]

I agree with you. I was just pointing out how using TPU max against any other site is not really relevant. Otherwise, I agree with your point that we should consult many different sites to get a more unbiased viewpoint.



GTX970/980 only came out in Sept 2014 but 14nm/16nm cards of that level prob won't be out until late summer/fall 2016. That's still 15-18 months from now - a long time. 960 came out towards late January which means it's not even 4 months old. GM200 6GB hasn't even launched. 750/750Ti haven't even turned 1 year old, which means this generation isn't even half way yet.

If anything, I would argue the complete opposite of your point. If AMD is extremely tight on resources, they should have never done 390/390X $500-700 level cards and focused all their efforts on the $100-400 desktop and mobile dGPU segments - that's literally 80% of the entire GPU market. If AMD spend hundreds of millions on dollars just to make 2 good chips in the form of uber expensive 390/390X/395X2 cards that only 5% of the market cares about, and everything else is old crap, they are stupid.

TPU uses one game for the avg. and peak and furmark for the max. As much as TPU is referenced for these measurements if I were AMD and/or nVidia I would optimize my drivers for best efficiency over performance in those benches (If they aren't doing that then they are dumb.). They should take an avg of all the games they use, just like they do with overall performance. Can't cheat optimize for efficiency in all of those benches and not hurt your overall performance results.

 

[Qwertilot]

Quite a lot of sales/profit in the compute markets?

They could very well be chasing those to a large extent with the 390/X. NV aren't half as strong there right now and the cards are big/expensive enough to let them use early HBM so they'll have a very real technological advantage vs NV too.

Looks a decent market opportunity to aim at.

Then yes, I was thinking they might have decided to basically save money and keep selling the 2xx stuff for a good while.

We'd all love them to really gun for the 970/80 and trigger a price war, but a big R&D spend needed to repurpose the 20nm designs at 28nm and not obviously much return on it. They're clearly not drowning in money with all those cancelled CPU projects.

If anything, I guess they might need the 3xx range beneath the 390/X to let them finally kill off those very cheap 290/X cards. They can't be doing their image/sales of their other cards any good.

We'll see

 

[arandomguy]

Regarding TPU power measurements, they have mutliple measurement categories which are explained. "Average" and "Peak" is a gaming measurements while "Maximum" is the Furmark measurement.

Those who take "Maximum" results against Crysis 3 results from Anandtech for example only have themselves to blame.

750/750Ti haven't even turned 1 year old, which means this generation isn't even half way yet.

Was launched with retail availability in Feb. 2014 - http://www.anandtech.com/show/7764/the-nvidia-geforce-gtx-750-ti-and-gtx-750-review-maxwell

 

[Glo.]

I'm not making stuff up, it's listed as OEM on AMD's page + slides.

This looks like HD 8000 all over again with an incoming renaming.

This is exactly what Im thinking about this next gen.

 

[3DVagabond]

Regarding TPU power measurements, they have mutliple measurement categories which are explained. "Average" and "Peak" is a gaming measurements while "Maximum" is the Furmark measurement.

Those who take "Maximum" results against Crysis 3 results from Anandtech for example only have themselves to blame.

The Avg and Peak on TPU are with Metro LL. So yet a different set of results.

 

[Qwertilot]

With the 750ti being 15 months old there has to be a chance of a similarly 'early' small 60w variant of Pascal turning up. Won't replace anything directly of course but it will flag what is coming.

 

[3DVagabond]

With the 750ti being 15 months old there has to be a chance of a similarly 'early' small 60w variant of Pascal turning up. Won't replace anything directly of course but it will flag what is coming.

I really really doubt we'll see Pascal on 28nm.

 

[Qwertilot]

True But they can sanely push out that sort of small die much sooner on 14/16 than they can the bigger ones. Won't need enormous volumes either.

 

[JDG1980]

http://media.bestofmicro.com/H/6/474954/original/03-Overview-Torture.png

Tom's has the R9 285 max power at 187 watts...R9 280 at 241 watts.

I guess it all depends on who you ask.

Tom's Hardware uses an unspecified GPGPU load for their "torture test". TechPowerUp uses FurMark, which is the only reliable way to get maximum power consumption out of a card.

 

[JDG1980]

But, there are rumors about unsold amd inventory stuck on the market. Maybe amd wants to clear that before 14nm GPUs come out next year (?), because after that, those old parts will not find a new home.

That would be a terrible decision and a classic case of the sunk cost fallacy. You don't let your company fall way behind its competitors because you've got unsold back stock. You blow it out at fire sale prices and move on.

 

[ocre]

I would say nvidia sells more low skus like the 750 than they do all others combined.
So, yeah.......they need large quantities.

What I find strange is, what's up with them anyway. We have the 900series out but the 750ti is still in the line up. That has got to be confusing to the average joe.

As for futuremark being the only way for max power, I don't think so.
I believe the throttling has an effect.

Didn't mining on stock volts use a heck of a lot of power?

 

[LTC8K6]

Tom's Hardware uses an unspecified GPGPU load for their "torture test". TechPowerUp uses FurMark, which is the only reliable way to get maximum power consumption out of a card.

I'm guessing the lack of an overclock is a big factor.