VEGA DIESHOT

>>60765326
Hawt.

>>60765326
>which seems to be Samsung and NOT Hynix.

Fixed

>>60765326
Dang, that HBM is about 1/3rd the whole freaking die, hope it's worth it.

>>60765326
>2 stacks
>not 4
DOA. What a joke.

>>60765326
The block between the two memory controllers is likely the command processor, below that is likely the entire frontend, which might or might not include the ACE, which if I'm reading this correctly are now part of the CUs.

The interesting thing is the memory PHY like thing on the far right, which could be a IF link.

File: FijiBlockDiagram.png (1MB, 3999x2250px) Image search: [Google]

1MB, 3999x2250px

>>60765326
Fiji for comparison.

>>60765412
And the same amount of memory as a 4 stack GV100

Really makes you think.

>>60765712
Really weird Nvidia didn't wait for 8GB stacks for a product that clearly needs all the memory it can get.

How strange.

The new Mac is using a cutdown Vega with 56CU instead of 64

Vega confirmed for virtual launch

>>60766515
So the clocks remain unchaged? Still at 1600MHz at those 11tflops?

Pretty nice, Vega will be quite efficient.
Must be using a lot of Zen's power IP

>>60765326
Is that 4GB or 8GB total VRAM on there? each HBM2 chip has two stacks of 2GB memory on top of each other? I hope that is 8GB

>>60767956
It's 8GB to 16GB

A HBM2 chip can go from 2GB to 8GB

>>60767980
16GB sounds overkill for general use currently, but is good to have the option available.

>>60765326
It looks like tripple surface area of Polaris which makes it really promissing

>>60768045
They've said 16GB will mostly go for the workstation cards, maybe some more expensive gaming cards too if there's capacity.
But I figure gaming cards will be 8GB

>>60768080
I heard they told manufactorers to knock themselves out so you can expect it from 4(2?) to 16.

Also I feel like after the founders edition the gay men vega will arrive with custom coolers right away, the founders is the reference.

>>60765712
But half the bandwidth, which cripples the damn thing.

File: AMD-INSTINCT-VEGA-3-740x416.jpg (68KB, 740x416px) Image search: [Google]

68KB, 740x416px

>>60768151
>cripples the damn thing.
>Roughly 85% performance
>for a fraction of a price
>no bridges required
>comes in 3 petaflops racks

Volta is DOA

>>60768151
Kek no it doesn't, if your shaders need less bandwidth than your memory subsystem can provide, they're not crippling anything.
And in 95% of situations, they don't.

GV100 actually needs the bandwidth because it has enormous double precision performance and is generally freaking huge, around twice the size of Vega and the 1080ti
In summary: you can't compare bandwidth requirements between different architectures.

>>60768141
If their memory compression in hardware is effective for general use then they might be able to pull off 4GB at 4K just fine.
Rather stick with 8GB because that is the safer bet.
What I am really looking forward to is the elimination of GPU sag and dealing with smaller cards in general.

>>60768209
Speaking of bandwidth, due to bugs in Fury X's memory controller, it could only effectively use 360GB/s out of its 512GB/s

Thankfully Fiji was more of a HBM test run than anything else or AMD would have been fucked completely.

>>60768074
Polaris 10 is 230mm^2, Vega 10 is 500-550mm^2, so only about 2.2-2.4x.

>>60768791
>>60768791
Not him, but a significant part of the Polaris die is the memory controller, the two HBM2 controllers are tiny compared.
Hence more area going to shaders.

https://videocardz.com/newz/amd-releases-vega-gpu-die-shot

>The memory is codenamed GPA022GA2656, which does not correspond to SK Hynix nomenclature. Are those memory chips made by Samsung?

>Vega 10 HBM2 Bandwidth
>400 GB/s

>NVIDIA 1080ti GDDR5x Bandwidth
>480 GB/s

>NVIDIA GP100 HBM2 Bandwidth
>732 GB/s

Make you think. SK Hynix can't provide, that's why we waiting.

So when are the 580's coming back out?

File: Screenshot_20170606_032403.png (211KB, 1343x371px) Image search: [Google]

211KB, 1343x371px

>>60769061
The Vega in Apple's shit is underclocked for power reasons, and it's not a full Vega anyway, 56 out of 64CU, go to AMD's site , full Vega is 480GB/s for the workstation part, the gaming might be clocked higher, or not.

>>60769269
If there are lower clocked HBM2 cards there will be higher clocked as well, assuming any of the Vegas actually need more than 480GB/s bandwidth

>>60768791
It's below 500, maybe 450mm2 even

>>60768506
bingo, the original HBM test bed was going to be a Hawaii sized die but due to thermal problems (microbumps would crack from differing thermal expansion rates) they had to axe it in favor of Hawaii. Codename was Tiran.

>>60765326
Yea, thats got the d32b memory controller. Bottom of the barrel stuff really.

File: 1476679021786.png (9KB, 160x160px) Image search: [Google]

9KB, 160x160px

>dieshot
>it's just colored rectangles

>>60772063
When you see a die that shows visible micro structures of the architecture its because the top metal layers which would otherwise obstruct them have been removed. This image does not show those upper most layers removed.

I would fuck that die

>>60765397
Which is way smaller than the area taken up by GDDR* RAM.

>>60765326
It may be DOA but damn it's sexy

File: Die-Shot-AMD-Fiji-8.jpg (963KB, 2000x1689px) Image search: [Google]

963KB, 2000x1689px

>>60765559
Actually Fiji die shot for proper comparison

>>60773940
>DOA
kek https://www.youtube.com/watch?v=owL_KY9sIx8

>>60773956
That's a big die.

>>60773960
I like this guy.

Vega has been a fucking rollercoaster. I have gone through a cycle of despair and hope several times with all of the info. I wish Vega the best, but what I really want is Navi. They can scale well with CPU cores, I want to see them do that with GPUs as well.

>>60774037
Navi has a potential to change everything. it seems to me it would be a greater achievement than Zen.

>>60770667
Huh, I actually looked it up and found this. It's interesting to see just how aggressive AMD's 2009 graphics roadmap was. nVidia would have been crying if AMD had come out with a top down new lineup every 12 months.
>First gen GCN pushed back a whole year
>The entire of the original Northern Island lineup scrapped and for a single 40nm SKU (Cayman)
>Bali is scrapped and Cape Verde is on TSMC instead because GloFo's 28nm is late
> Sea Islands scrapped all apart from Bonaire, Volcanic Islands is launched instead with Hawaii.
Where they being way to ambitious with their roadmap? It's seems most of it got completely upturned due to delays to manufacturing nodes.

>>60774037
>>60774066
Considering Raja mentioned on twitter that Vega uses infinity fabric that could be quite interesting for the dual vega cards.

>>60773982
For you.

>>60765326
My gpu died when will this shit come out?

>>60774158
frontier edition end of june. normal editions end of july.