Vega Rumors

Hahahaha.

The one game Shintai produced benchmarks for that shows the HBM based Fury X in a pretty bad light. I'd never heard of the game,but thanks for the clarity on its state of quality code.

Irony...

Tbh, I was looking into the Fury X and its performance in AAA video games, I found it doing quite poorly vs its counterpart the 980Ti in quite a few games. I stopped when I realized the amount of time and work could be better wasted on something enjoying. But hopefully someone else cares enough to check.

Thanks cageymaru for that post about Square Enix being shit (RIP Sleeping Dogs), I was about to give them points for letting IO leave with the Hitman IP, but not anymore.
 
there are lot of games that make the HBM suffer badly once you crank textures up and some quality visual settings up, Dying Light and Rise of The tomb raider comes to my mind also Doom.

Exactly. But that conflicts with the alternative reality that's made around HBM ;)
 
https://videocardz.com/69475/amd-radeon-vega-spotted-with-16gb-memory-and-1600-mhz-clock

we finally have something. Looks like final silicon for top end vega is indeed running at 1600mhz boost! This may be the limited edition all we know with 16gb of hbm2. We will find out.

https://compubench.com/compare.jsp?...&hwtype1=dGPU&hwname1=AMD+6864:00&D2=AMD+Fiji

Compared to Fiji:
upload_2017-5-15_12-55-11.png
 
$400 for 1070 tier is terrible. And then at the same time, the slight bump to $600 for 1080 Ti is amazing.

$400, $500, and $600 is really fucking dumb for 3 tiers. Nobody should be dumb enough to believe that.
Can't believe TPU fell for it. The guy posted on Reddit with no source, just made it all up.

Guru3D posted the same.
http://www.guru3d.com/news-story/am...posedly-leaked-vega-coreeclipse-and-nova.html

Booooooooooooo.
 
$400 for 1070 tier is terrible. And then at the same time, the slight bump to $600 for 1080 Ti is amazing.

$400, $500, and $600 is really fucking dumb for 3 tiers. Nobody should be dumb enough to believe that.
Can't believe TPU fell for it. The guy posted on Reddit with no source, just made it all up.

Guru3D posted the same.
http://www.guru3d.com/news-story/am...posedly-leaked-vega-coreeclipse-and-nova.html

Booooooooooooo.

True, when you look at it the way you do. That's why the most sensible explanation is that the $400 card will beat the 1070 easily, while the top end will underperform the 1080 Ti.
 
1600 on the core clock is impressive along with a double stack of HBM2 at 16GB. I could have sworn in the other thread that a few people said that wont happen. Looks like it will be a interesting launch soon.
 
True, when you look at it the way you do. That's why the most sensible explanation is that the $400 card will beat the 1070 easily, while the top end will underperform the 1080 Ti.
It doesn't really matter, three SKUs from $400 to $600 ain't gonna happen.
 
1600 on the core clock is impressive along with a double stack of HBM2 at 16GB. I could have sworn in the other thread that a few people said that wont happen. Looks like it will be a interesting launch soon.
Yeah, there are several know it alls who claimed that. Shiftai top of the list.
 
The only thing i don't get is where you've found 16 gigs of HBM2 at? To my eye it clearly reports 4GB and 8GB of VRAM.
 
I meant to also include a child comment attached with the one I just linked to.


Again, I don't actually know if its possible.
 
I meant to also include a child comment attached with the one I just linked to.


Again, I don't actually know if its possible.

Crossbar? No. You cant cram shell it either like with GDDR. If you want higher capacities you increase the HBM height.
 
Guys something is really wrong with that compubench detail.
Q2 2017 book from SK Hynix does not show anything but 4-Hi stack at 4GB 1.6Gbps.

And this book is showing Q4 '17 for their GDDR6; 12 and 14Gbps if anyone is interested.
So it is up to date.
Cheers

Edit:
And to also clarify, this book has not been referenced by many if at all due to it being quite recent.
 
Last edited:
As a few of us have been saying, 8-Hi stack is still a long way off and TBH going to be a nightmare to use with GPUs.
Cheers
 
Probably, I remembered Robert Hallock from AMD said Vega is compatible with GDDR5 or HBM. This could be a GDDR variant of Vega.
 
GDDR5X/6 would certainly make more sense. There's little reason to push for HBM for desktop parts as it is; AMD certainly hasn't shown that they've been willing to put enough compute resources into a consumer product to actually take advantage of HBM, and certainly not enough to live with its disadvantages (cost).
 
GDDR5X/6 would certainly make more sense. There's little reason to push for HBM for desktop parts as it is; AMD certainly hasn't shown that they've been willing to put enough compute resources into a consumer product to actually take advantage of HBM, and certainly not enough to live with its disadvantages (cost).

Even Hynix between the lines gave up on HBM with the GDDR6 showcase. Just as efficient and fast as HBM2 they say.
 
Guys something is really wrong with that compubench detail.
Q2 2017 book from SK Hynix does not show anything but 4-Hi stack at 4GB 1.6Gbps.

And this book is showing Q4 '17 for their GDDR6; 12 and 14Gbps if anyone is interested.
So it is up to date.
Cheers

Edit:
And to also clarify, this book has not been referenced by many if at all due to it being quite recent.

You add more stacks. Fury had 4 stacks.

Even Hynix between the lines gave up on HBM with the GDDR6 showcase. Just as efficient and fast as HBM2 they say.

Cheaper HBM3 is around the corner.
 
You add more stacks. Fury had 4 stacks.



Cheaper HBM3 is around the corner.

Vega 20 in 2019 or so will have 4 stacks. You are not getting more than 2 stacks on Vega 10.

HBM3 isn't cheaper and its not around the corner. If lucky you see it around 2020. And I think you confuse it with the slower and cheaper HBM2 edition. Something that is not around the corner either.
SamsungHBM3.jpg

LowCostHBM.jpg
 
Vega 20 in 2019 or so will have 4 stacks. You are not getting more than 2 stacks on Vega 10.

Says who?

HBM3 isn't cheaper and its not around the corner. If lucky you see it around 2020. And I think you confuse it with the slower and cheaper HBM2 edition. Something that is not around the corner either.

You state HBM3 isn't cheaper and then post images stating that it is cheaper. HBM1 is HBM Gen1, HBM2 is HBM Gen2, and HBM3 is HBM Gen3.

By the time GDDR6 is out, that is around the corner.

It's also not slower. I think you should actually read the images you post.
 
Says who?

AMD.
https://videocardz.com/65521/amd-vega-10-and-vega-20-slides-revealed


You state HBM3 isn't cheaper and then post images stating that it is cheaper. HBM1 is HBM Gen1, HBM2 is HBM Gen2, and HBM3 is HBM Gen3.

By the time GDDR6 is out, that is around the corner.

It's also not slower. I think you should actually read the images you post.

You seem to mix up HBM3 and low cost HBM. Its 2 completely different things.

And GDDR5X is out, GDDR6 is coming this year. HBM3 isn't coming before 2020 or so. Neither is low cost HBM that is currently more a concept thought than an actual product plan.
 
Q2 2017 book from SK Hynix does not show anything but 4-Hi stack at 4GB 1.6Gbps.
So just use more 4-Hi stacks. Most PCs only have dual channel memory controllers, but I'd imagine almost everyone here isn't running just 2 DIMMs in their system. Same concept. Identical bandwidth and double the capacity at the cost of a single pin to indicate which chip is being accessed.

By the time GDDR6 is out, that is around the corner.
GDDR6 is just GDDR5 with an improved process. Not much holding back that old technology other than ramping up production. It could probably be used on existing products with little to no architectural changes.
 

That's your source? A leak that may or may not be real? That leak also states 2 stacks of 8GB of HBM2 which if SK Hynix book is correct, doesn't exist yet.

You seem to mix up HBM3 and low cost HBM.

And GDDR5X is out, GDDR6 is coming this year. HBM3 isn't coming before 2020 or so. Neither is low cost HBM that is currently more a concept thought than an actual product plan.

You seem to have an incapability to read things you post yourself, nor what I type to point out things your own images state. HBM3 is HBM Gen3, just as HBM2 is HBM Gen2. That low cost HBM is HBM Gen3.
 
So just use more 4-Hi stacks. Most PCs only have dual channel memory controllers, but I'd imagine almost everyone here isn't running just 2 DIMMs in their system. Same concept. Identical bandwidth and double the capacity at the cost of a single pin to indicate which chip is being accessed.

That's like asking to use bigger GDDR bus. Same issues applies. It cost a lot in power etc.
 
That's your source?
Well, that's better than your source, because at least we have seen it.
You seem to have an incapability to read things you post yourself, nor what I type to point out things your own images state. HBM3 is HBM Gen3, just as HBM2 is HBM Gen2. That low cost HBM is HBM Gen3.
Except that second slide has NO words about HBM Gen3, and first slide has NO words about cost. They are 2 distinct slides about 2 distinct things from 2 distinct presentations. How dense do you have to be to miss it?
 
That's your source? A leak that may or may not be real? That leak also states 2 stacks of 8GB of HBM2 which if SK Hynix book is correct, doesn't exist yet.



You seem to have an incapability to read things you post yourself, nor what I type to point out things your own images state. HBM3 is HBM Gen3, just as HBM2 is HBM Gen2. That low cost HBM is HBM Gen3.

Raja even showed the chip, with 2 stacks. How much more denial do you wish to be in?
AMD-Vega-GPU_2.jpg


Low cost HBM isn't HBM3. And its still a ~2020 product at best.
 
A great deal of people here will be sorely disappointed when Vega gets released.
 
Raja even showed the chip, with 2 stacks. How much more denial do you wish to be in?
AMD-Vega-GPU_2.jpg

They showed a chip on an interposer with 2 stacks. That doesn't rule out another chip on an interproser with 4 stacks.

Compubench shows a Vega with 16GB of memory. It either is two stacks of 8GB HBM2, or four stacks of 4GB HBM2. Which is the most likely to you?

Low cost HBM isn't HBM3. And its still a ~2020 product at best.

Yes it is.
 
What is stopping them from making a 4 stack interposer for high end or gddr... The chip can obviously do it for mi25.
Another thing is not all companies publicly list their juicy stuff. Many that I'm involved with don't.
 
Back
Top