New Macbook Pro with vega graphics coming soon

[koyoot]

Does AMD use the same Apple Pro branding elsewhere?

16 Polaris CUs would be equivalent to 12 Vega CUs. It is AMD's problem if they cannot upgrade the memory while keeping the price.

I would say 16 Vega CUs should cost the same as 16 Polaris CUs, as they come two years later.

If he was talking about 20 Vega CUs, it would be OK for the card to cost more than $99 (there's already a big premium to go from 8 to 11).

We have yet to see if Small Vega is front End bottlenecked like big Vega is. On paper it is not bottlenecked, by any means. You may be right, that in proper scheduling front end scenario vega may be faster than Polaris, in the manner you have posted. However ALUs in GPUs are always only ALUs. That is why 1.3 GHz 1280 CUDA core GPU, with 192 GB/s memory bandwidth, will perform exactly like 1.3 GHz, 1280 GCN core GPU with 192 GB/s. Other things like Tiled Based Rasterization, Operand Reuse Cache, only save bandwidth, Register File Size, and in result: power, and allow GPUs to increase core clocks higher. On graphcis side, if you are not severly bottlenecked elsewhere - ALUs will always play main role.

It's one way of looking at this. But it has to be commercially viable for them, or why would they sell the product to begin with. There is a reason why Nvidia does not use HBM2 — too expensive. Vega is a premium product in the sense that it uses very expensive tech. They can't really sell it to an average person. Apple is great partnership for them, since customers are ready to pay more for Apple products.

Let me break it down, a bit.

Single GDDR5 memory chip costs currently 7-8$, depending on capacity. It is 32 Bit, 512 MB-1 GB chip. If you want to build 4 GB, 256 GB/s memory subsystem, you have to pay 56-64$, for 8, 8000 MHz GDDR5 memory chips. GDDR6 memory chips will cost 25% more, than that.

Single 4 GB HBM2 stack costs 40$. That price does not include TSV's and interposer cost. Big Vega's 10 GPU, memory subsystem, with two stacks of HBM2, including TSV's and Interposer costs 100$, for dual 4 GB chips, and 320$ for 16 GB chips(single 8 GB HBM2 chip costs 150$). For Small GPU, like Vega 10, HBM2 plus TSV's and Interposer memory subsystem it may cost around 50$, at best.

However, GDDR6 memory, even with 25% price increase, will alleviate that price for 128 Bit memory subsystem, with 224 GB/s memory bandwidth, with 4 GB of Frame Buffer to just 40$.

 

[cube]

They do use Pro branding on their professional GPUs.

I'm not sure I share your theory on price formation

I know AMD uses Pro for other products.

If you think 16 Vega CUs should cost less than 16 Polaris CUs at launch, I will not disagree.

 

[leman]

If you think 16 Vega CUs should cost less than 16 Polaris CUs at launch, I will not disagree.

I do not know how much it should cost. I don't know any numbers related to the market, costs of materials or operating costs for producing these chips. The only thing I am pretty sure of is that Vega is significantly more expensive to make.

 

[cube]

I do not know how much it should cost. I don't know any numbers related to the market, costs of materials or operating costs for producing these chips. The only thing I am pretty sure of is that Vega is significantly more expensive to make.

Polaris and Vega are still both 14nm. It is the HBM stuff that is more expensive.

 

[koyoot]

I do not know how much it should cost. I don't know any numbers related to the market, costs of materials or operating costs for producing these chips. The only thing I am pretty sure of is that Vega is significantly more expensive to make.

Maybe not significantly, but it costs more .

10-15$ more it should cost more to make than Polaris 11/21.

 

[cube]

Is this Vega 12 chip 7nm too?

 

[koyoot]

Is this Vega 12 chip 7nm too?

No. 14 nm FF.

 

[e1me5]

It's mostly my "greediness" that makes me angry with Apple releasing the Vega options. I've saved more than enough to get the full monty of the 2018 15" MBP and no I feel a little bit cheated.
BUT a thought is brewing in the back of my head, what if they also redesigned the components, resulting in less throttling, resulting to a better performance of the whole package rather than only from the GPU front?
I would like Apple to give us the early buyers the option to trade in the 560X MBPs for the Vega ones, but in the same time, mine is a extremely stable unit with no KPs or other problems that plague other 2018 MBPs, so maybe a swap just for satisfying my lust for the "biggest and best" components, to trigger a divine intervention punishing me with a faulty unit.
Just my thoughts.

 

[cube]

Maybe not significantly, but it costs more .

10-15$ more it should cost more to make than Polaris 11/21.

How much less does it cost to make Polaris 21 now than in 2017?

 

[koyoot]

How much less does it cost to make Polaris 21 now than in 2017?

Not really much, because even if the wafer prices went down, DRAM(GDDR5) prices went up.

 

[cube]

Not really much, because even if the wafer prices went down, DRAM(GDDR5) prices went up.

I said Polaris 21, not RX 560.

Did your price difference include VRAM?

 

[koyoot]

I said Polaris 21, not RX 560.

Did your price difference include VRAM?

RX 560 is Polaris 21 based...

 

[cube]

RX 560 is Polaris 21 based...

Polaris 21 is a chip, not a card.

You still did not say if the $15 added cost was for a whole 4GiB 20CU Vega card.

 

[Thysanoptera]

Vega 12.

Since you were looking at those Linux drivers, is the PCI ID of it 0x69A0? There is something like that in AOTS database, entries saved between November last year and January of this. Quite bad, RX555 levels or even below. Maybe they had the chip for a long time but couldn't get the drivers right.
https://www.ashesofthesingularity.c...-details/14e7cec5-0833-4eaa-9f15-d01042d36132

 

[koyoot]

Polaris 21 is a chip, not a card.

You still did not say if the $15 added cost was for a whole 4GiB 20CU Vega card.

This is what you have precisely asked...

How much less does it cost to make Polaris 21 now than in 2017?

 

[cube]

This is what you have precisely asked...

Yes, I asked about the chip, and you talked about GDDR5.

 

[koyoot]

Yes, I asked about the chip, and you talked about GDDR5.

It is inherent part of the chip. Without GDDR5 the GPU is not working.

Count yourself the price difference of wafers: 6500$ in 2016, vs 4500-4900$ in 2018.

 

[cube]

It is inherent part of the chip. Without GDDR5 the GPU is not working.

It is not part of the chip. @leman said Vega was much more expensive to make than Polaris.

I don't think it is more expensive at the same performance level (without VRAM). And about 30% more expensive at the same CU count, which could be the same cost as when Polaris launched.

4900 * 1.3 = 6370

 

[koyoot]

It is not part of the chip. @leman said Vega was much more expensive to make than Polaris.

I don't think it is more expensive at the same performance level (without VRAM). And about 30% more expensive at the same CU count, which could be the same cost as when Polaris launched.

<facepalm>

Memory subsystem is inherent part of GPU manufacturing cost. You cannot value its manufacturing costs seperately.

 

[cube]

<facepalm>

Memory subsystem is inherent part of GPU manufacturing cost. You cannot value its manufacturing costs seperately.

The controller in the chip is not the expensive part of HBM.

 

[koyoot]

The controller in the chip is not the expensive part of HBM.

Nobody cares. Apple would not buy Polaris GPUs separate from GDDR5 memory, as well as they would not buy Vega 12 or Vega 10 GPUs separate from their memory subsystem.

 

[cube]

Nobody cares. Apple would not buy Polaris GPUs separate from GDDR5 memory, as well as they would not buy Vega 12 or Vega 10 GPUs separate from their memory subsystem.

The point is that a Vega chip itself is not more expensive than an equivalent Polaris one. Then you can have more power at the same cost as two years ago.

 

[leman]

It is not part of the chip. @leman said Vega was much more expensive to make than Polaris.

I was talking about the entire GPU package. I don't think there is much point in talking about the price of the chip alone if you also need the RAM and the interposer.

 

[cube]

As @koyoot said he does not care about the price of the GPU, I assume he meant a whole 20CU 4GiB Vega package, which is a lot more powerful than a 16CU 4GiB Polaris package, only costs $15 more.

Then his suggested card price of $150 seems fair.

 

[koyoot]

The point is that a Vega chip itself is not more expensive than an equivalent Polaris one. Then you can have more power at the same cost as two years ago.

How can it not be more expensive when it is bigger in die size? It has at least 30% more xTors.

Let me give you a comparison. 123 mm2 die on 300 mm2 wafer gave you 575 working dies, with 95% of them fully working. Vega 12 appears to have at least 160 mm2 die size, and that results in 333 working dies on the same process.

It means that Vega 12 right now when each wafer, is almost 30% less expensive, costs AMD almost two times more to make! 7$ vs 13$.

Now lets add all of it: 7$ die, plus 32$ for memory = 39$ for Polaris 11, without validation, assembly of package, materials of the package.
Vega 12: 13$ die, 45$ memory subsystem = 58$ without validation, assembly of package, materials of the package.