AMD Unveils More Info About Vega

[Vode]

19 hours ago, cj09beira said:

its weird no one is talking about this, i made some calcumalations and i got this:

14nm fury x 357mm2 (from 596)

in 536mm2 using 14nm it could have 6144 cores (using the same cu/mm2 of the rx 480)

 

i really believe the gpu shown was bigger than 4096 cus, how mush more is still to be determined thou 

You can't compare it like this because the new core are presumably different. We'll propbaly see 4000 core or even less than that.

[Space Reptile]

On 5.1.2017 at 10:30 PM, Zheren159 said:

 If AMD's top of the line GPU beats my 1070 im going to throw a brick at someone. And by  brick I mean my 1070 brick.

well it beat a 1080 in doom , so it shud mess w/ a 1070 in DX11

[cj09beira]

24 minutes ago, Vode said:

You can't compare it like this because the new core are presumably different. We'll propbaly see 4000 core or even less than that.

its a stretch to say that the new core needs almost 50% more space for the same amount of cores

[Vode]

35 minutes ago, cj09beira said:

its a stretch to say that the new core needs almost 50% more space for the same amount of cores

We don't know the die size yet. It could be smaller. More cache is pretty much guaranteed, takes up a fair amount of die space. Also we know nithing about clock speeds.

 

My guess is roughly a 400-450m2 chip. I just don't think we'll see 600m2 chips yet...

[DocSwag]

1 hour ago, cj09beira said:

well, first i calculated the racio between die size increase and core increase, using 280x to fury x, with that racio and  with the rx 480 as a base i calculated what would a 14nm fury be (357mm ) then i played around by increasing to core count until i got to 530mm2 which is the size of the vega chip (give or take).

280x                                                        fury x

352mm                                                    596mm

2048c                                                      4096

   ====  200% number of cores ===>

   ==== 169% die size  ==========>

racio 0.845

i attached a excel file with all the data

 

 

 

cards specs.xls

But you ignored all the weird variables there could be. The rx 480 only has 32 ROPs, and it's quite likely vega could have as much as 128 ROPs, which doesn't lineup with the CU increase (the percent gain of ROPs is greater than the percent gain of the CUs). The same could be applied to TMUs, Geometry Processors, cache, and even memory controllers (since vega uses hbm you need a much wider memory controller and so the memory controller could potentially be many, many times larger than the Polaris one). That's not to mention the changes made to vega such as the CUs, ROPs, Geometry processors, memory controllers, and probably even more stuff that we don't know about yet! That's not to mention the other stuff you ignored, which is the stuff that is universal to all chips and doesn't change from chip to chip (pcie controllers, crossfire stuff, display stuff, etc.).

 

Basically, what I'm saying is there are so many variables that you need to account for. There's stuff that doesn't scale from chip to chip and stays the same size like pcie controllers and stuff, there's ROPs and stuff that don't scale the same amount as CUs from chip to chip which throws off the die size, and there might even be special circuitry added on to the chip that we have no idea about that could throw off the measurements. So we have nowhere near the amount of info we need to make such estimates.