R9 380x rumor and speculation thread

[tviceman]

I think also nvidia is just intelligent enough to realize just how hard it will be to sell last gens gpus after people see the nodeshrink gpus. If it's truly the performance leap some people believe it to be, they'll be hard pressed to sell any remaining 970 or higher cards after their new gpus launch.

Remains to be see whether ShintaiDK's prediction that Nvidia and AMD will go with optimal perf/mm2 (i.e. lower transistor budget than the typical ~1.7-2x increase with new nodes) or stick with the typical cadence. If developing new architectures on finfets isn't exponentially more difficult, then I predict Nvidia and AMD will stick to the typical increases in transistors over predecessor GPU's. So if GP104 is 80% faster than GM204, then we're looking at about 45% faster than a stock reference 980 TI and probably new & higher mid-die prices to go with it.

 

[Techhog]

Remains to be see whether ShintaiDK's prediction that Nvidia and AMD will go with optimal perf/mm2 (i.e. lower transistor budget than the typical ~1.7-2x increase with new nodes) or stick with the typical cadence. If developing new architectures on finfets isn't exponentially more difficult, then I predict Nvidia and AMD will stick to the typical increases in transistors over predecessor GPU's. So if GP104 is 80% faster than GM204, then we're looking at about 45% faster than a stock reference 980 TI and probably new & higher mid-die prices to go with it.

Remember that certain people here have biases/agendas and will base predictions on those.

 

[AtenRa]

So if GP104 is 80% faster than GM204, then we're looking at about 45% faster than a stock reference 980 TI and probably new & higher mid-die prices to go with it.

For that to happen the GP104 should have almost 2x the transistor count and almost the same die size as GM204. For a "104" low price product i highly doubt it.

It should be more in line for the GP104 to match the GM204 at half the die size.

 

[Silverforce11]

For that to happen the GP104 should have almost 2x the transistor count and almost the same die size as GM204. For a "104" low price product i highly doubt it.

It should be more in line for the GP104 to match the GM204 at half the die size.

Mid-range dies of next gen has typically replaced current high-end. It's quite normal for a major node shrink to deliver that performance, on top of a new improved uarch.

For the same die area, 60-80% performance gains should be expected. Less than that implies either a poor uarch design that doesn't scale and/or drivers/current games have not been able to fully utilize the new uarch.

I'm hoping 2 top GPU on 16nm ff will be enough for a great 4K experience without compromising on IQ settings.

 

[AtenRa]

Mid-range dies of next gen has typically replaced current high-end. It's quite normal for a major node shrink to deliver that performance, on top of a new improved uarch.

For the same die area, 60-80% performance gains should be expected. Less than that implies either a poor uarch design that doesn't scale and/or drivers/current games have not been able to fully utilize the new uarch.

I'm hoping 2 top GPU on 16nm ff will be enough for a great 4K experience without compromising on IQ settings.

Sorry my bad, I thought he was talking about the low-end skus.

Yes the GP104 at the same die size should be 80-100% faster than GM204.

ps: does NV will revert to 100 series (GP100) again or they will stick to 200 (GP200) ??

 

[Techhog]

For that to happen the GP104 should have almost 2x the transistor count and almost the same die size as GM204. For a "104" low price product i highly doubt it.

It should be more in line for the GP104 to match the GM204 at half the die size.

So, you're expecting next gen to have no improvements compared to current gen other than power consumption? Bold.

EDIT: Nevermind, I read the post above.

 

[Techhog]

Sorry my bad, I thought he was talking about the low-end skus.

Yes the GP104 at the same die size should be 80-100% faster than GM204.

ps: does NV will revert to 100 series (GP100) again or they will stick to 200 (GP200) ??

I think the first number is basically a revision number.

 

[AtenRa]

I think the first number is basically a revision number.

ahh ok thanks

 

[Mondozei]

So next week?

It's launching soon™.

 

[tential]

Can we please change the thread name... To just? 380x rumor and speculation? It's not finally here...

 

[Techhog]

Can we please change the thread name... To just? 380x rumor and speculation? It's not finally here...

Well, we at least have actual pics now and seems to be in the hands of reviewers...

 

[AtenRa]

The asus looks nice, custom large PCB, CU II Strix cooler and back plate.

 

[tential]

How close in performance will this card be compared to the 290(since the 390 is a little bit faster)

 

[Techhog]

Confirmed 100% at the end of this video: https://www.youtube.com/watch?v=rjpefXm4YYs

 

[Mondozei]

If we take a look at TPU's performance index we see this:




Notice how the 280X is ahead of the 380. So say the 380X is faster than the 380 by 10%(not the same as 5 percentage points). It'll be a smidgen slower than the 280X, a 4 year old GPU by now.

In Sweden, I can buy a new 280X for $199. For the 380X to be competitive, it has to match that, since I don't think we'll see a 20%(or more) performance improvement over the 380. Of course, it would have a lower TDP than the 280X, but is that worth a $50 premium?

Now, if it can be faster than the 280X by 5-10% and still draw 60 watts less than it for a knockout $199, then we're talking.

 

[Techhog]

Notice how the 280X is ahead of the 380. So say the 380X is faster than the 380 by 10%(not the same as 5 percentage points). It'll be a smidgen slower than the 280X, a 4 year old GPU by now.

In Sweden, I can buy a new 280X for $199. For the 380X to be competitive, it has to match that, since I don't think we'll see a 20%(or more) performance improvement over the 380. Of course, it would have a lower TDP than the 280X, but is that worth a $50 premium?

Now, if it can be faster than the 280X by 5-10% and still draw 60 watts less than it for a knockout $199, then we're talking.

Are you seriously implying that you think there's a chance of it being slower than the 280X? Are you crazy?

 

[happy medium]

Are you seriously implying that you think there's a chance of it being slower than the 280X? Are you crazy?

No , I think he's saying that @ 250$ it best be faster than a 280x because the 280x has been under 200$ for quite a while now.

 

[.vodka]

Can we please change the thread name... To just? 380x rumor and speculation? It's not finally here...

Agreed, thread name changed. Sneaky card.

How close in performance will this card be compared to the 290(since the 390 is a little bit faster)

I suppose a throttling reference blower 290 is still going to be definitely faster than a full Tonga. Despite GCN 1.1 vs 1.2, there's just too much extra hardware in Hawaii to overcome by clock speed alone (suppose, a 1.1GHz Tonga vs 900 MHz 290/390 Hawaii at 95°C)

2048 vs 2560sp (1.1GHz 2048sp ≈ 890MHz 2560sp in raw TFLOPS)
128 vs 160 TMUs
32 vs 64 ROPs (even considering Tonga's improved ROPs and color compression making up for the narrower memory bus)

As shown by the 285 being faster than the 280x in certain workloads and consistently ahead of the 7950; a full Tonga not clocked as conservatively as in the 285, should easily pull away from the 280x (which is usually at the 1GHz spot) I suppose by at least 5-10%. It should fill that spot between 380 and 390 nicely. GCN 1.2 also handles tessellation much better, so that's another plus over prehistoric 7970/280x/Tahiti.


In that TPU chart I'd place it by the 780. Price will make or break this card.

 

[Mondozei]

Are you seriously implying that you think there's a chance of it being slower than the 280X? Are you crazy?

If 380X is 10% faster on average than 380, then it would be slower than a 280X at 1080p. If it's 15% faster, it would just barely be faster and cost $50 more.

If you look at the difference in leaked specs between 380 and 380X, it doesn't look like a typical 20% increase (or more) in performance. More like the difference between 290/290X, which was 10-15% if you'd remember.

I'm surprised you're shocked that the 380X isn't probably going to be faster than 10-15% over the 380. Even if we take the upper range, 15%, the extra performance over a 280X would be quite small for a premium of $50. Basically what you'd be paying for is lower wattage mostly.

Is the market big enough for such a niche?

 

[poofyhairguy]

If 380X is 10% faster on average than 380, then it would be slower than a 280X at 1080p. If it's 15% faster, it would just barely be faster and cost $50 more.

If you look at the difference in leaked specs between 380 and 380X, it doesn't look like a typical 20% increase (or more) in performance. More like the difference between 290/290X, which was 10-15% if you'd remember.

I'm surprised you're shocked that the 380X isn't probably going to be faster than 10-15% over the 380. Even if we take the upper range, 15%, the extra performance over a 280X would be quite small for a premium of $50. Basically what you'd be paying for is lower wattage mostly.

Is the market big enough for such a niche?

How well does Tonga overclock?

 

[Mondozei]

Just to clarify even further. 42*1.15=48.3
48.3/47=1.0276

So around 2.8% faster than a 280X if it's 15% faster than a 380. If it's 10% faster than a 380 it would be slower. I don't expect it to be slower as a base case, but the possibility is there. And even if it'll be faster, it will more or less trade blows with a 280X unless AMD blows it out of the water with a 20% or more performance improvement(highly unlikely given the leaked specs).

So yeah, a GPU which will be, at best, 2-3% faster than the 4 year old 280X for $50 dollars extra(if those prices hold). The value proposition isn't amazing, to put it mildly.

How well does Tonga overclock?

The 280X also overclocks nicely. This isn't Hawaii.

Edit: Maybe it's unfair to compare the 380X against the 280X. After all, the 280X is a legendary GPU by now. But still, we can't use kids gloves.

 

[tential]

Agreed, thread name changed. Sneaky card.



I suppose a throttling reference blower 290 is still going to be definitely faster than a full Tonga. Despite GCN 1.1 vs 1.2, there's just too much extra hardware in Hawaii to overcome by clock speed alone (suppose, a 1.1GHz Tonga vs 900 MHz 290/390 Hawaii at 95°C)

2048 vs 2560sp (1.1GHz 2048sp ≈ 890MHz 2560sp in raw TFLOPS)
128 vs 160 TMUs
32 vs 64 ROPs (even considering Tonga's improved ROPs and color compression making up for the narrower memory bus)

As shown by the 285 being faster than the 280x in certain workloads and consistently ahead of the 7950; a full Tonga not clocked as conservatively as in the 285, should easily pull away from the 280x (which is usually at the 1GHz spot) I suppose by at least 5-10%. It should fill that spot between 380 and 390 nicely. GCN 1.2 also handles tessellation much better, so that's another plus over prehistoric 7970/280x/Tahiti.


In that TPU chart I'd place it by the 780. Price will make or break this card.

Ya I took a second look at the specs comparison between the current card, the new card, and the 390(290), and ya, .

Oh well, didn't need this card anyway. I guess a second R9 290, then undervolting them so they fit within my PSU is the way to go! I really was expecting overall GPU power usage to go down just like how intel's CPU power has been going down. I guess I miscalculated and need a new PSU now.

 

[Techhog]

If 380X is 10% faster on average than 380, then it would be slower than a 280X at 1080p. If it's 15% faster, it would just barely be faster and cost $50 more.

If you look at the difference in leaked specs between 380 and 380X, it doesn't look like a typical 20% increase (or more) in performance. More like the difference between 290/290X, which was 10-15% if you'd remember.

I'm surprised you're shocked that the 380X isn't probably going to be faster than 10-15% over the 380. Even if we take the upper range, 15%, the extra performance over a 280X would be quite small for a premium of $50. Basically what you'd be paying for is lower wattage mostly.

Is the market big enough for such a niche?

I'd expect the difference to be similar to the difference between the 280 and 280X. There is no chance whatsoever of it being slower than the 280X. Your math is just bad. By your logic, the 380 should be slower than the 280.

 

[Elixer]

So next week?

It's launching soon™.

I wonder if this means, no reference design from AMD, so, no review from this site...

AMD just needs to price this card right, and it will sell. Simple as that.

A few more days, and we will see.

 

[littleg]

Just a case of waiting to see how they manage to screw up this launch

If I was them i'd be pricing it for volume. $250 is a little high, around $220 and it becomes a great proposition and one that would gain them back market share and, more importantly, get people into their ecosystem.