Vega/Navi Rumors (Updated)

[maddie]

Not sure about the accuracy but since HBM2 is newer and has alot more customers (nVIDIA, AMD etc).. the cost might have been driven up substantially given demand/supply compared to HBM1. Plus these are going into high margin products so perhaps the manufacturers like SK Hynix etc are looking to recoup the cost in investment also.

So there could be some truth in this. Doesn't bode well for the consumer versions i.e. the gaming SKUs.

From $12 est to $80 for a same 4-Hi stack. Yes the new stack has 1GB memory chips instead of 256MB ones, but still no way I believe that price. A lot of FUD in my opinion.

 

[Valantar]

If this were even remotely true, then the FE card would have been on a smaller board. Even if you don't need the actual board room, a bigger card allows for more fans and better cooling.

That doesn't quite make sense. We haven't seen any board shots, and making a blower cooler capable of dissipating 250W is not an easy task. The FE might very well be a small card in a big shroud. After all, the Fury/Fury X were just 7.5" long, and they had both bigger dice and twice the HBM stacks of Vega.

I don't think the 1080Ti has 50% of power going towards the memory. Probably only 25% at absolute maximum.

HBM2 will mean AMD has lower TDP, or more headroom for an OC, but I think you're overestimating its importance here.

Where did you get those numbers from? Not from me, that's for sure. All articles I've seen looking into memory power consumption at the time seemed to imply that the Fury X's competitors spent around 50W of their TDP on memory, while the Fury X was around half that or a little less. The 1080Ti has 11GB of GDDR5X running high speed, so I would be very surprised if it didn't use at least that much power on RAM. On the other hand, Vega has half the HBM stacks of Fiji. This bodes well for power usage.

I see guru3d has a post headlined as:

AMD RX Vega HBM 2 8GB Memory Stack Reportedly Costs $160
Reports have been showing that a 4 GB HBM2 stack costs $80. AMD Vega GPUs make use of 8 GB HBM2, two stacks used is $160.


To think they're stating the just the HBM2 modules cost almost as much as the entire Fiji card was estimated to cost and it's being seriously considered as accurate in the forum.

From $12 est to $80 for a same 4-Hi stack. Yes the new stack has 1GB memory chips instead of 256MB ones, but still no way I believe that price. A lot of FUD in my opinion.

Yeah, that seems absurd. $20 per GB? In 2017? That makes no sense. Sure, HBM is more complex than GDDR5, and needs the interposer and all that jazz. But $80 for one stack? No way. Might they be mixing up per-stack price with the price of the RAM + the interposer + packaging?

 

[leoneazzurro]

There is no source nor evidence for the pricing rumor, so...

 

[antihelten]

If amd actually did what you suggested and had the wc cooled top end chip for $600 (significant discount from 1080ti) I'll break time to have it first. But I doubt it. I know I'll preorder this and be highly disappointed with the price /perf. I'm expecting $700 or even a full match to the 1080ti at $750.

Well, looking at the Fury X vs. 980 Ti, which is somewhat comparable, I think AMD has to go lower in price this time around for 3 reasons.

First off there's the obvious one with timeframe, Fury X launched just 3 weeks after the 980 Ti, whereas Vega RX will be launching ~4 months after the 1080 Ti, so a lot of the potential market will be gone already. Secondly we saw that AMD priced Fury X the same as 980 Ti even though it was a tad slower, probably because they thought that the inclusion of an AIO water cooler was enough of a value add to make up for the difference, however the launch of said cooler was anything but smooth, so I think a lot of the glamour of having an AIO water cooler is probably gone today, and as such AMD can't rely on it as a value add to the same degree. Finally Volta probably isn't all that far off, and AMD need to grab what market share they can until it arrives.

As such I think AMD has to price Vega lower than Fury X relatively speaking, so if Fury X was priced at the same price as the 980 Ti, then Vega probably has to be cheaper than the 1080 Ti. With the 1080 Ti at $700, I think $600 makes sense (although it could also be $650). But as I also said, this assumes performance roughly equal to the 1080 Ti, if AMD can actually beat it, then all bets are off.

Barely slower....
You could get fury x speed out of it if you wanted it just didn't make sense since you weren't gaining a lot of perf for a lot extra wattage.

The Nano was specced at 1000MHz, but in reality it only averaged about 875 MHz, or about 17% slower than Fury X.

Anyway I definitely think a Vega Nano would be nice, and the 1200 MHz version that has been showing up in benches might be a candidate for this (although 1200 MHz would be a much bigger gap than we saw with Fiji). To be honest though I think it basically comes down to whether or not it makes financial sense for AMD to create an official reference ITX card like this, rather than just leaving it to their partners (like with the Zotac 1080 Ti mini).

Gap is not that big.RX580 with aftermarket cards is only 20-25% slower than GTX1070.

I assume that you're referring to my claim about the 1070 being 30-40% faster than the 480 here. We are basically saying the same thing, the 1070 being 30-40% faster is the same as the 480 being 25-30% slower, the remaining 5% is then simply due to the difference in clock rate between the 480 and the 580.

 

[Ancalagon44]

I assume that you're referring to my claim about the 1070 being 30-40% faster than the 480 here. We are basically saying the same thing, the 1070 being 30-40% faster is the same as the 480 being 25-30% slower, the remaining 5% is then simply due to the difference in clock rate between the 480 and the 580.

Mathematically, you could both be right even ignoring the difference in clock rate.

Look at it this way - compare the numbers 4 and 3.
4 is 33% higher than 3, but 3 is only 25% less than 4.

Same would apply to the performance difference between 580 and 1070. If the 1070 gets 40FPS and the 580 gets 30FPS, then the 1070 is 33% faster than the 580, and the 580 is 25% slower than the 1070.

 

[Valantar]

Mathematically, you could both be right even ignoring the difference in clock rate.

Look at it this way - compare the numbers 4 and 3.
4 is 33% higher than 3, but 3 is only 25% less than 4.

Same would apply to the performance difference between 580 and 1070. If the 1070 gets 40FPS and the 580 gets 30FPS, then the 1070 is 33% faster than the 580, and the 580 is 25% slower than the 1070.

Isn't that exactly what @antihelten is saying?

 

[antihelten]

Isn't that exactly what @antihelten is saying?

Yep. Although I did round the numbers a bit, so it's perhaps slightly less clear that the numbers overlap (my 25-30% is actually 23-29% without rounding which more obviously overlaps with Head1985's 20-25%)

 

[Head1985]

I assume that you're referring to my claim about the 1070 being 30-40% faster than the 480 here. We are basically saying the same thing, the 1070 being 30-40% faster is the same as the 480 being 25-30% slower, the remaining 5% is then simply due to the difference in clock rate between the 480 and the 580.

Reference rx580 dont exist
Those 1400+mhz cards are very fast.GTX1070 is only 25% faster or 22% with boost bios.In 1440p the gap is same.Also its same gap around 26% that is between GTX1070 and 1080.SO if Nano is at GTX1070 performance there is no room for more cards.They need replace polaris with small vega if they ever release it.

 

[antihelten]

Reference rx580 dont exist

I never said that it did, so I have no idea how that is relevant. All I said is that the performance difference between the 580 and the 480 is due to higher clocks on the 580.

Those 1400+mhz cards are very fast.GTX1070 is only 25% faster or 22% with boost bios.In 1440p the gap is same.Also its same gap around 26% that is between GTX1070 and 1080.SO if Nano is at GTX1070 performance there is no room for more cards.They need replace polaris with small vega if they ever release it.

Again, you're just repeating what I already said.

At 1080P the 1070 is 33% faster than the 480 and at 1440P it is 37% faster. This falls perfectly in line with my claim that the 1070 is 30-40% faster than the 480, and inversely that the 480 is 25-30% slower than the 1070.

The 580 will obviously be faster than the reference 480, doesn't matter which model we're talking about here, they all are (except maybe some low power models I don't know about). The gap in performance between the 580 (again any model) and the 480 is 5% or more which accounts for the difference between your claim that the 580 is 20-25% slower than the 1070 and my claim that the 480 is 25-30% slower.

There is no contradiction here.

 

[MrTeal]

That doesn't quite make sense. We haven't seen any board shots, and making a blower cooler capable of dissipating 250W is not an easy task. The FE might very well be a small card in a big shroud. After all, the Fury/Fury X were just 7.5" long, and they had both bigger dice and twice the HBM stacks of Vega.

 

[Mopetar]

Where did you get those numbers from? Not from me, that's for sure.

Your previous post seemed to imply 50%: Edit: I'm illiterate.

Still, the 1080 Ti mini has to dissipate 250W, not the 175 of the Nano - and it doesn't even have a vapor chamber to help it along. As for the Nano being handicapped due to concentrated heat, I disagree simply due to how much less power HBM uses compared to GDDR5(X). If your RAM is using half the power, and is close enough to be cooled by the same vapor chamber as the GPU itself, I'd count that as an advantage, not a handicap.

That's phrased as a hypothetical, so perhaps I've misread what you meant to imply, but it seems to suggest it's the case, or else why pick half as an example. Edit: never mind, I think I understand what you meant. Half the power as in using less than would would otherwise be used instead of half as in using 50% of the total amount.

All articles I've seen looking into memory power consumption at the time seemed to imply that the Fury X's competitors spent around 50W of their TDP on memory, while the Fury X was around half that or a little less. The 1080Ti has 11GB of GDDR5X running high speed, so I would be very surprised if it didn't use at least that much power on RAM. On the other hand, Vega has half the HBM stacks of Fiji. This bodes well for power usage.

I recall that GDDR5X has been stated to be more power efficient over GDDR5, though I don't know if that comparison is at the same speeds or how it is being made. My guess is that the 11GB GDDR5X on the 1080Ti uses about the same power overall as the slower 8 GB GDDR5 on the 980Ti. Essentially NVidia traded the better efficiency for better clock speeds and more memory.

The better question is how much power does the HBCC save. Letting you get by with less VRAM is good and all, but it takes power to run the HBCC so it's not as simple as just cutting out 20W due to lower memory use.

 

[Valantar]

Your previous post seemed to imply 50%: Edit: I'm illiterate.



That's phrased as a hypothetical, so perhaps I've misread what you meant to imply, but it seems to suggest it's the case, or else why pick half as an example. Edit: never mind, I think I understand what you meant. Half the power as in using less than would would otherwise be used instead of half as in using 50% of the total amount.



I recall that GDDR5X has been stated to be more power efficient over GDDR5, though I don't know if that comparison is at the same speeds or how it is being made. My guess is that the 11GB GDDR5X on the 1080Ti uses about the same power overall as the slower 8 GB GDDR5 on the 980Ti. Essentially NVidia traded the better efficiency for better clock speeds and more memory.

The better question is how much power does the HBCC save. Letting you get by with less VRAM is good and all, but it takes power to run the HBCC so it's not as simple as just cutting out 20W due to lower memory use.

I guess I should have said "half as much power"

And yeah, the power consumption of the HBCC is indeed interesting. As for GDDR5X power, I think I remember reading that it was supposed to be something like 20% less power hungry as high end GDDR5 at the same speed (8GT/s? Can't remember.) I would assume similar power per chip for 11GT/s GDDR5X, and thus a total RAM power consumption increase of ~30-40% for Big Pascal. Might be that I'm pessimistic here or remembering wrong, though. Still, HBM2 should be noticeably less power hungry.

 

[Mopetar]

I guess I should have said "half as much power"

And yeah, the power consumption of the HBCC is indeed interesting. As for GDDR5X power, I think I remember reading that it was supposed to be something like 20% less power hungry as high end GDDR5 at the same speed (8GT/s? Can't remember.)

I think I read something that said either 66% or 75% of the power usage of GDDR5, but I can't recall and can't seem to find wherever I'd read it.

80% would make sense assuming that RAM uses the same equation for calculating power use as CPUs where power scales with the square of the voltage as GDDR5 is 1.5 V and GDDR5X is 1.35 V and (1.35/1.5)^2 is .81 (so about 20% less), but I'm not sure if RAM operates under the same principles as does a CPU when it comes to determining how many watts of power it consumes. I don't think that the speed of the RAM matters as much as the voltage, otherwise RAM power usage probably wouldn't have dropped as much as it has over time.

 

[antihelten]

I think I read something that said either 66% or 75% of the power usage of GDDR5, but I can't recall and can't seem to find wherever I'd read it.

Actually, according to Micron GDDR5X is actually similar or slightly more power hungry than GDDR5, not less. This is of course largely due to GDDR5X running at significantly higher speeds, and if you actually normalized for speed GDDR5X would use less, but no one uses GDDR5X at such low speeds so that's not really relevant.

From anandtechs GDDR5X article:

Due to lower voltages and a set of new features, power consumption of a GDDR5X chip should be lower compared to that of a GDDR5 chip at the same clock-rates. However, if we talk about target data rates of the GDDR5X, then power consumption of the new memory should be similar or slightly higher than that of GDDR5, according to Micron. The company says that GDDR5X’s power consumption is 2-2.5W per DRAM component and 10-30W per board. Even with similar/slightly higher power consumption compared to the GDDR5, the GDDR5X is being listed as considerably more energy efficient due to its improved theoretical performance.

 

[tential]

@antihelten
What's your opinion of topend Vega if it doesn't release at the discounted $600 price that makes sense, but at the $750 price to match the 1080Ti just like last year? Obviously performance slower than a 1080Ti. No way I'm going to bet on Vega being faster.

 

[w3rd]

Vega is more powerful than Pascal.

Anyone who is saying different really doesn't understand technology, or are bias and just trolling.

 

[Malogeek]

Anyone who is saying different really doesn't understand technology, or are bias and just trolling.

Or simply aware that it isn't out yet to actually compare and study.

 

[OatisCampbell]

Or simply aware that it isn't out yet to actually compare and study.

I'll have to agree with this one for now.

I hope it it far more powerful than Pascal though, AND could use a GPU win and we could all appreciate more power for 4k.

 

[IllogicalGlory]

Vega is more powerful than Pascal.

Anyone who is saying different really doesn't understand technology, or are bias and just trolling.

In what way exactly? Maybe I don't understand technology. Can you give an explanation?

 

[GoodRevrnd]

So many people in this thread with the blinders on. I would not be at all surprised if Vega was a good bit faster than Pascal. I also would not be at all surprised if it fell well short.

 

[tential]

The more telling thing is, what does it say about AMD if Vega is slower given AMD basically had 2 generations of improvements of GPU tech to go against Pascal?

 

[IllogicalGlory]

Remember when they showed Ryzen beating the 6900K side by side? Months before launch? That was pretty cool. Remember when they showed Vega beating the 1080 Ti side by side? I don't either.

 

[antihelten]

@antihelten
What's your opinion of topend Vega if it doesn't release at the discounted $600 price that makes sense, but at the $750 price to match the 1080Ti just like last year? Obviously performance slower than a 1080Ti. No way I'm going to bet on Vega being faster.

If AMD releases Vega at $700 (not $750, the 1080 Ti launched at $700), and doesn't have better performance than the 1080 Ti or beat it in efficiency (by a significant margin), then that to me would reflect that AMD has essentially given up on Vega 10 as a market share grabber and are just looking to recoup expenses and are looking ahead to Navi instead.

This in a way is arguably what happened with Fiji. AMD clearly didn't price Fiji to grab market share, and the sales numbers reflected this. To me this reflected that Fiji was to a large extent seen as a pioneering effort and a pipe cleaner (for HBM tech) for AMD. Vega might end up in a similar place given the above scenario, with AMD hoping to make a proper play for the market with Navi. One slight difference here though is that AMD seems far more keen to target the professional market with Vega than they did with Fiji, which could change the calculus by a significant degree.

By the way I wouldn't completely rule out Vega being faster than 1080 Ti (although I wouldn't bet on it either). As @crisium pointed out earlier, if Vega can match Hawaii PPF and we assume perfect scaling, then Vega beats the 1080 Ti by about 15% (at 1600 MHz). Obviously you never get perfect scaling, but even so beating the 1080 Ti by 5-10% shouldn't be impossible. Of course I also expected Polaris to match Hawaii PPF and thus end up closer to the R9 Fury in performance instead of where it actually ended, so who knows.

The more telling thing is, what does it say about AMD if Vega is slower given AMD basically had 2 generations of improvements of GPU tech to go against Pascal?

I would say that this simply reflects the fact that a generational leap for AMD is no better than a generational leap for Nvidia (even though a lot of people like to disparage Nvidia's generation and refer to Pascal as a 14nm version of Maxwell and so on), which really shouldn't come as a surprise to anyone, given the disparity in development budgets.

In a sense I think Nvidia and AMD have had the same number of leaps here though. Nvidia had two in Kepler to Maxwell and Maxwell to Pascal, and AMD had GCN 3 to Polaris and Polaris to Vega. Now some might say that AMD had three with GCN 2 to GCN 3, (which roughly coincided with the Kepler to Maxwell leap), but then again I think that was a small enough leap (apart from HBM on Fiji), that if you count that then you might as well also count Maxwell 1 to Maxwell 2.

 

[beginner99]

Remember when they showed Ryzen beating the 6900K side by side? Months before launch? That was pretty cool. Remember when they showed Vega beating the 1080 Ti side by side? I don't either.

True bit different divisions and competitor being in different situation. For example AMD new about intels roadmap. Kaby-lake was just releases and Skylake-x and coffeelake still far away. Nothing intel could do except lower kaby prices or as they did speed up release of new processors.

In case of NV AMD knew that the 1080 Ti will release and leaking information before it's release like Vega being ultra fast, then NV could lower prices prematurely forcing AMD to release vega on the cheap as well. It's better to make your competitor react rather than reacting yourself.

 

[Head1985]

We already know vega performance.They showed it in SE4.It was slower than aftermarket 1080TI.SE4 is best case scenario where furyx is 22% faster than GTX1070.So in average it should be somewhere between GTX1080 and 1080TI.

Edit:my tip is this