AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

[Nebuchadnezzar]

On 5" wafers maybe. The common size is 12" (300mm) nowadays. That equals about 70,000mm² usable space per wafer .

Corrected, I was too hasty with the calculator. The point is that this isn't even a full day of fab capacity. It's not representative of yields over time.

 

[CarstenS]

I've been doing bin tracking for a whole lot longer than anybody here so don't take me for a fool.

Well, I will take your word for it.

Vega is supposed to be 475mm² and at 100% yield that is 24 chips per wafer.

Before there are further misunderstandings: What wafer size and chip size (width/length) are you basing this off?
For very roughly guesstimated 19.2×24.7 mm die on a 300mm wafer, the number of full dies per should be more like 116, if i can trust this online-calculator here: http://www.silicon-edge.co.uk/j/index.php/resources/die-per-wafer

Actual factory yields on ideal dies are probably in the 75% range. On a chip of this size the yields are probably closer to 50-60%. Even at those low yields all you need for 10000 units is a little over 800 wafers. That is just a couple of days on a fab line.
There are two options:

• AMD is voltage binning the chips to get better yields, this means that power is suffering and to get a proper median distribution and representative data on power we need a chip that has been in line production for a couple of weeks already. The proof of this is Polaris as you've clearly seen the power difference between a day 1 card and a card produced a year later. It doesn't matter if 14nm is "known", the yield curve for a new chip is something that only goes up in time, irrespective of process maturity. You can safely bet that the FE cards have the earliest silicon out of the fab.
• AMD is eating the yields and throwing away bad chips or keeping them for an even worse SKU, this means the power we see would actually be representative and the architecture is just disappointing.

Thank you, so again, this would be option #1.

 

[xEx]

Maybe we are lucky and gaming Vega will bring the full driver and will show 50% better performance while happily taking the performance crown. However whoever is responsible for the FE launch at AMD should be looking for a new job.

I know dreaming is free but 50% better performance seems like dreaming too hard for me...When was the last time that a product gain 50% more performance over a month? yes, exactly.

 

[Clukos]

Firstrike Ultra link: http://www.3dmark.com/fs/12988276

I tried to find a similar GPU score with a 1080 and the same CPU: http://www.3dmark.com/fs/9272181

What is interesting is this:

Vega FE:
GPU Test 1: 28.12 fps
GPU Test 2: 19.75 fps

1080:
GPU Test 1: 30.2 fps
GPU Test 2: 19.44 fps

Afaik (and I could be wrong) the first GPU test is focused on geometry + high tessellation factor while the second one is focused on testing memory bandwidth (at least on my 1080 Ti I've seen the biggest difference with higher memory clock on that test).

 

[chavvdarrr]

One has to wonder why AMD had chosen to use HBM ...

 

[SpaceBeer]

Why do you think HBM is an issue?

 

[T1beriu]

One has to wonder why AMD had chosen to use HBM ...

Size.Very big win in AIO, mobile & tiny PCs.

 

[seahawk]

but 300W TBP is not so good for that application.

 

[chavvdarrr]

Why do you think HBM is an issue?

- expensive
- unavailable
- not faster than gddr5x/gddr6

Its rambus all again

 

[Rodéric]

Gimme a Vega kthxbye !
(Will say how good it is if it supports everything I fancy!)

 

[Tofu]

Did Sebbis VEGA just explode?

 

[SpaceBeer]

- expensive
- unavailable
- not faster than gddr5x/gddr6

Its rambus all again

With four stacks of HBM, Fiji was ~$40 more expensive than R9 290X. I doubt difference in production cost between 2 HBM2 and 8 GDDR5X/6 stacks is bigger than that. It is probably even smaller. It will save some space and reduce power consumption (by 10 W maybe)

Regarding availabiltiy, there was no official information, only rumours.

 

[Deleted member 2197]

With four stacks of HBM, Fiji was ~$40 more expensive than R9 290X. I doubt difference in production cost between 2 HBM2 and 8 GDDR5X/6 stacks is bigger than that. It is probably even smaller. It will save some space and reduce power consumption (by 10 W maybe)

Regarding availabiltiy, there was no official information, only rumours.

https://forum.beyond3d.com/posts/1982501/

 

[silent_guy]

You don't know when the silicon came out of the fab, and even in mass production there can be large variations in power. You need proper high volume to be able to have a proper distribution and a median that you can draw conclusions from. How many FE's is AMD selling? 1000's? 10000's? That's too low to be representative. You only have to look at Ryzen to see the actual variations that come out of a factory. This is why binning even exists.

Sorry, but none of this matches reality.

First of all, it's not as if Vega is produced on some super immature process. The characteristics and the spread of all parameters should be very well by now.

Second, all silicon has a bunch of test structures to determine where those parameters are located for a particular piece of silicon.

Third, you immediately order corner lots where the process skewed towards the various extremes of the process.

With the combination of all that, you only need a limited amount of samples to determine how many samples you're going to get for each bin.

100000 not enough? Ridiculous, and time to fire your statistician immediately. With a few 100, you'll be very close to the truth.

 

[silent_guy]

Why do you think HBM is an issue?

I don't know if there's an issue with it, but there is simply no point in using it if you can't benefit from its advantages.

 

[sonen]

There are two options:

• AMD is voltage binning the chips to get better yields, this means that power is suffering and to get a proper median distribution and representative data on power we need a chip that has been in line production for a couple of weeks already. The proof of this is Polaris as you've clearly seen the power difference between a day 1 card and a card produced a year later.

No, I have not seen clear differences. RX 580 does not bring anything into perf/W, even dips compared to RX 480, as expected due to higher clocks.
Unless you are referring to magical Polarises using 120W overclocked sky high. Those I've seen, but they are coming from users using GPU-Z readings which does not give the entire card's power consumption.

 

[silent_guy]

With four stacks of HBM, Fiji was ~$40 more expensive than R9 290X. I doubt difference in production cost between 2 HBM2 and 8 GDDR5X/6 stacks is bigger than that. It is probably even smaller. It will save some space and reduce power consumption (by 10 W maybe).

We've talked about this before. In the table that you link to, they estimate a cost of $15 for one PCB and $5 for the other. Those prices are ridiculously out of whack. And any outfit with a phone and somebody who speaks Mandarin can figure that out.

If such an easy to obtain number is completely wrong, why do you trust a number, the price of HBM, that is much harder to figure out?

 

[DuckThor Evil]

Vega is supposed to be 475mm² and at 100% yield that is 122 chips per wafer. Actual factory yields on ideal dies are probably in the 75% range. On a chip of this size the yields are probably closer to 50-60%. Even at those low yields all you need for 10000 units is a little over 20 wafers. That's not even a day in fab line.

I think the calculator is still giving wrong numbers. I think you wanted to say over 200 wafers.

Corrected, I was too hasty with the calculator. The point is that this isn't even a full day of fab capacity. It's not representative of yields over time.

It might not be representative of yields over time, but it should be fairly representative of power characteristics of the chip. These chips are working within AMD-specs. specs that have been released to public, 300W for the air cooled and 375 for the liquid cooled version. It's very unlikely that these power figures are going to drastically go down when the volume grows, unless a major redesign is done. The process itself is also already fairly mature.

 

[SpaceBeer]

We've talked about this before. In the table that you link to, they estimate a cost of $15 for one PCB and $5 for the other. Those prices are ridiculously out of whack. And any outfit with a phone and somebody who speaks Mandarin can figure that out.

If such an easy to obtain number is completely wrong, why do you trust a number, the price of HBM, that is much harder to figure out?

Ok, that might be wrong. But Fury had $550 launch price and ~$500 several months later. I think we can assume production cost of Hawaii/Grenada is ~200$ at most. There is no way they would go for HBM it is cuts their profits at half, i.e. if HBM increases production cost for additional ~$150 and doesn't bring a lot of performance. I mean, it's the high-end product that brings the money, why whould you lower your margins.

 

[Kaotik]

Ugh... "Gaming" FE drivers are dated 6/27/2017. RX is supposed to launch sometime before 8/3/2017. This is more or less a month of the final phase of driver development which has been ongoing for years...

Do you really think a month of SW development should bring 50% of perf?

And AMD has been saying since the beginning that FE in gaming mode is not the same as RX