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

Found some more SRAM, even though this is Scorpio. 1MB parameter cache.

 

About AMDs OpenCL implemantation for Vega:

clGetDeviceInfo for Vega seems to return:

CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF=1
CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT=2

Vega should be able to performs basic arithmetics with half2 and short2 in same time as with floats or ints, so the preferred vector size should be 2 for both. But it's one for halfs.

I wonder if the compiler is still missing the fast support for using those half vectors, or is this clGetDeviceInfo preferred width of 1 for halfs just a mistake?

 

http://www.tomshardware.fr/articles/radeon-vega-package-gpu-hbm,1-65141.html

tl;dr:
- AMD uses both Samsung & SK Hynix HBM, Samsung on Vega FE & Vega64, SK Hynix on Vega56 & Vega64
- South Korea packed Vega 10s don't have the filling between/around chips
- Taiwan packed Vega 10s have filling between/around chips
- The height difference between GPU & HBM is only on the South Korea packed Vega 10s, it's 40 µm, no issues when thermal paste fills the gap
- Taiwan packed Vega 10s are 0.1mm shorter than South Korea Packed
- In retail, Vega 64s are supposed to have filling, Vega 56's not

 

edit: nevermind. Delete.

 

What's the problem?

 

that some people likes to throw **** at amd and outright lie constantly like they have some agenda to fill
first hardocp then gibbo with the 100mh/s and then with the msrp and now they are crying over this

meanwhile people are already testing the new drivers while undervolted
https://www.hardwareluxx.de/index.p...vega-56-und-vega-64-im-undervolting-test.html

tl:dr 1070 power draw (a bit more) and 1080 perfomance (that is for the 56)

 

I wish it would work.

 

Vega is like the guide of how not to launch a product.

Coolers are more difficult to make cuz you need to make a design for both scenario or one for each one. Aftermarket cooler are harder also because of that. Why? because the chip hight is the main pivot point of the cooler and if you change it it will make the cooler not fitting properly.

 

Why do the coolers have to be different?

 

If only the power consumption and performance of cards other than AMD could be optimized by undervolting and/or overclocking.

 

What about the MSRP, who has lied about that (other than AMD themselves)? And I'm not any fan of Kyle, but what did HOCP lie about? Nothing, as far as I'm aware.

You seem a tad overly defensive methinks. AMD pooch-screwed this launch, there's no two ways about it. Undervolting your new GPU is fun and all for tweaker enthusiasts, but a product which is 14 years* late compared to the competition should be competitive straight out of the box, and vega clearly has issues to overcome there. Issues which haven't been made easier by AMD's constant fumbling of the ball one might add.

Latest fumbling in a long line is that launch drivers quietly still did not activate the new and improved rasterizers, something they kept from everyone until explicitly asked about.

*Edit: D'oh, well you know what I meant, right? lol!

 

Per the article, it at least requires adjustments to the screws and mounting methods, first for the height difference and then in the direction chosen for the mounting if the GPU and stacks are not the same height.
It's throwing off the rollout of custom cards as they figure out how to do manufacturing when AMD is throwing different stacks at them.
I didn't see a clear reason why this is happening.

Mechanically, getting this wrong may cause damage or lead to eventual reliability changes. They did not note thermal problems for the mismatched packages, but recommended not using low-viscosity TIM. I suppose that is because of the risk of it flowing out or forming voids with thermal cycling. The somewhat thicker interface might have a slight effect, but since this is going into the lower-specced Vega 56, it might not be as critical. That claim Hynix is going into all Vega 56 products, but also some Vega 64 products makes me curious. It seems like Hynix can make at least some of its memory fast enough. I'm curious how the chips and stacks are tested and when they are matched up to get all Vega 56 GPUs and Hynix stacks.
Either AMD knows enough ahead of time, or it's letting Vega 56 packages out with either the GPU or HBM being more capable but held back by the other.

Interestingly, the package height difference is greater than the non-molded GPU and stack heights. I'm not sure if Hynix is putting out two versions, since the molded Vega 64 is apparently using Samsung and Hynix stacks.
Which elements are different heights is not necessarily clear from the slides. If the interposer and substrate are not significantly different, does that mean the diagram's apparently showing the GPU being thinner is the reason?
That may explain the molding if the chips in Vega 64 were thinned and less rigid.

 

You realize what 0.1mm and 40um means, right?

 

Those differences were significant enough that AMD had to produce an informational slide to inform their partners so they could account for it in their products.

 

Per the article, potentially enough to matter for mounting assembly design and thermal compound flow.
The elements in question have heights measured in fractions of a mm.

The molded versus unmolded height difference is apparently being reviewed to determine what is appropriate for the height difference. In terms of uniformity, with baseplate variance specced to <=0.1mm, the height difference from one side of the package to the other is in the same order of thickness as the chips' heights. An HBM2 stack is listed in the standard as being 0.72mm tall.

That the GPU may be thinner may also speak to the molding being necessary. AMD may be assuming the GPU in the non-molded case is supposed to be stiff enough to take full mounting pressure, although preferably evenly distributed over the whole die.

 

Interesting that vega chips can apparently be binned without RAM precisely enough to determine if it will become a V56 or V64, so each flavor of GPU can be paired with its accompanying brand of RAM... That, or else there's a way to non-permanently bond vega dies with HBM for testing purposes. *shrug*

 

At what point has AMD lied about MSRP? It's $499 for just the card, $599 for Radeon Pack and $699 for Radeon Pack Aqua, it was that at launch, it is still that and it will be that in the future.
Now we have even AMD rep re-confirming it https://forum.beyond3d.com/posts/1997432/

 

They gave a price, and then shortly later it was a hundred dollars more expensive, without warning or explanation. The lower price was clearly never intended to be honored, so essentially a bait and switch (although perhaps not a blatant enough one to actually run afoul of the law.)

 

It has been $499 since day 1 and hasn't changed at any point, it didn't become $100 more expensive.

You're mixing "just the card" SKU and Radeon Pack SKU, which are priced $100 different. Some shops choose to sell Radeon Pack SKUs with just the games without even telling the customer that they'd get discounts on the ryzen/mobo-combo and eyefinity display, and ask that normal Radeon Pack price for it.

AMD is working on bringing all SKUs back to stock, including the $499 one, they first said it in the official statement and now we have it even more clearly stated by AMD representative. Of course if the shops or middlehands decide to hike up the price it sucks for customers, but AMD is doing their part of deal just like they should and can't do anything about the shops/importers/etc

 

I think the pricing thread has more of the back and forth on the MSRP issue.
There are various quirks to the discussion.