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

[Kaotik]

TBH the 16GB is an alarm bell for me.

Some may be interested on a few details from the newest SK Hynix Q2 2017 catalogue (all previous news was on Q1 book).
HBM2 still only 4GB 4-Hi stack at only 1.6Gbps.
GDDR5 10Gbps Q4'17
GDDR6 12&14Gbps Q4'17

Cheers

As the point was raised above, having something and being public about it can be different things. The fact is that 2 Gbps chip were available for a while in the catalogue, and you don't suddenly just lose the capability of producing them when you've already been making them, do you? So could it be as easy as AMD hogged all 2 Gbps chips from SK Hynix when they first came available > they got removed from catalogue because AMD took it all, and now the same with 8-Hi stacks?

 

[iMacmatician]

I singled-out the important part, which is no company is making 8-Hi stacks according to publicly released info.
There seems to be quite a lot of stuff being produced in secret nowadays. GDDR5X came out of nowhere with specs and production being announced just over a month before GP104 cards were in the hands of reviewers. TSMC had 12FFC volume production scheduled for 2018 but turns out nvidia has either been stocking chips from risk production or co-developed a secret second 12FF with TSMC.

Question about the possibility of 8-hi (4 GB) stacks: if they exist then why doesn't the Tesla V100 use them, since it uses 4 stacks for 16 GB? I assumed that the V100 would use the largest HBM2 capacities available. Could AMD have gotten exclusivity of 8-hi stacks for a period of time (or they bought all of them)?

 

[CSI PC]

As the point was raised above, having something and being public about it can be different things. The fact is that 2 Gbps chip were available for a while in the catalogue, and you don't suddenly just lose the capability of producing them when you've already been making them, do you? So could it be as easy as AMD hogged all 2 Gbps chips from SK Hynix when they first came available > they got removed from catalogue because AMD took it all, and now the same with 8-Hi stacks?

They were in the book before they ever got close to production, it was an estimation.
Same way Micron adjusted their GDDR5X over time.
Just reality of estimated spec and what can be done by when in reality.
Nvidia hogged all the HBM and GDDR5X, but they still showed up in catalogues because there is a calendar schedule for ordering large scale.
What really matters is status of when production orders can be accepted, if someone buys all current stock it does not mean they stop manufacturing or taking orders.
Or do you think AMD can afford to buy all 2Gbps and 8-hi stack HBM2 stock now and to end of 2017 from SK Hynix and that would have to be in advance for it not to be advertised for Q4'17?
Cheers

 

[Kaotik]

Question about the possibility of 8-hi (4 GB) stacks: if they exist then why doesn't the Tesla V100 use them, since it uses 4 stacks for 16 GB? I assumed that the V100 would use the largest HBM2 capacities available. Could AMD have gotten exclusivity of 8-hi stacks for a period of time (or they bought all of them)?

Well there are 2 companies producing HBM-memories, and NVIDIA has so far only used Samsung for theirs, maybe they want to stick with them and their schedules?

 

[gamervivek]

After thinking about it, we don't know how Vega's new High Bandwidth Cache Controller behaves in the eyes of the OS. It could be a Vega that simply shows direct "control" of 16GB of system RAM with the HBM2 being used as cache, and this could be completely transparent to anything but the driver.




I singled-out the important part, which is no company is making 8-Hi stacks according to publicly released info.
There seems to be quite a lot of stuff being produced in secret nowadays. GDDR5X came out of nowhere with specs and production being announced just over a month before GP104 cards were in the hands of reviewers. TSMC had 12FFC volume production scheduled for 2018 but turns out nvidia has either been stocking chips from risk production or co-developed a secret second 12FF with TSMC.

Or maybe it's just being reported to OpenCL, the values not necessarily true of the underlying hardware, especially the memory of a card presented in windows.

 

[Deleted member 13524]

Just because Apple buys all current availability does not remove it from the actual catalogue that covers quarters and the whole year (hence why it shows Q4'17 for GDDR6 and estimate for 14Gbps).

This actually counters your whole point.
The Radeon Pro 460 (Polaris 11 with all 16 CUs enabled and tuned for 35W TDP) was kept in secret until after the new Macbook Pros were unveiled in late October 2016, although other Polaris 11 graphics cards had been in the market for a months with only 14 CUs enabled.

Apple buying all availability did remove the Radeon Pro 460 from AMD's catalog until the laptop was formally announced.

Question about the possibility of 8-hi (4 GB) stacks: if they exist then why doesn't the Tesla V100 use them, since it uses 4 stacks for 16 GB? I assumed that the V100 would use the largest HBM2 capacities available. Could AMD have gotten exclusivity of 8-hi stacks for a period of time?

It's possible that AMD is entitled to exclusivity from whatever they want out of SK Hynix's HBM portfolio for some time because its development was (maybe still is?) a joint venture between the two companies.

The GV100 probably uses Samsung's HBM2 like the GP100 before it and we don't know if Samsung is capable/willing to produce 8-Hi HBM2.
Regardless, that GPU uses 4 stacks not only because of the memory amount but also because that's the number of stacks needed to achieve their target VRAM bandwidth.
Besides, NVLINK 2 does 25GB/s duplex and GV100 has 6 of those, so at 150GB/s duplex from a larger RAM pool the GPU may not even need to have more than 16GB.

 

[CSI PC]

This actually counters your whole point.
The Radeon Pro 460 (Polaris 11 with all 16 CUs enabled and tuned for 35W TDP) was kept in secret until after the new Macbook Pros were unveiled in late October 2016, although other Polaris 11 graphics cards had been in the market for a months with only 14 CUs enabled.

That is a full product not a manufacturing component.
Micron originally had 14Gbps GDDR5X in their catalogue with future estimated date but then removed it for over 6 months.
So are you saying Nvidia purchased all of the 14Gbps that never saw the light of day and still MIA?
In reality it finally went back onto the catalogue a couple of months ago, the reason is that Micron had issues with product manufacturing and hitting specs-rating (known they had problems) and now looks like it will be in sampling and product status in Q4 - estimate but looks more confident this time.

 

[Deleted member 13524]

That is a full product not a manufacturing component.

On one hand you have a fully enabled GPU working at lower voltages, on the other you have higher-clocked memory.
Both are dependent on higher binning. Both were/are being kept away from the publicly available catalogs until a certain flagship product was/is announced.

The comparison could hardly be more similar.

 

[Jawed]

Nvidia hogged all the HBM[...]

Evidence?

 

[CSI PC]

On one hand you have a fully enabled GPU working at lower voltages, on the other you have higher-clocked memory.
Both are dependent on higher binning. Both were/are being kept away from the publicly available catalogs until a certain flagship product was/is announced.

The comparison could hardly be more similar.

Again for a very similar comparison:

Micron originally had 14Gbps iGDDR5x n their catalogue last year but removed it for over 6 months.
Are you also saying then Nvidia purchased all of those 14Gbps GDDR5x that still is MIA?
MIcron only added it back to the catalogue a few months ago but still with no guarantee although estimate is Q4.
That is the reality like I said, and applies to HBM2 as well.

Same can be said regarding Samsung HBM2, it was still possible to see as product status even though Nvidia was purchasing what seemed all stock.
As I said, unless one is buying the whole of the years production supply in advance it will be in the catalogue, especially as the catalogue goes up to Q4 2017 and has dates around that time for GDDR6.

 

[CSI PC]

Evidence?

And you do not want evidence about AMD buying a whole years supply of HBM2 running at 2Gbps and 8-Hi stack that many seem to think is happening and a reason for it to be taken out of the catalogue?

The logic from some is it was removed from SK Hynix catalogue because it has all been purchased by AMD, but then one has to then accept the 8-Hi stack as well that originally existed in the catalogue with estimate date and was then removed, followed 1 or 2 quarters later by the 4-Hi 2Gbps stack.

 

[Deleted member 13524]

Again for a very similar comparison:

Micron originally had 14Gbps iGDDR5x n their catalogue last year but removed it for over 6 months.
Are you also saying then Nvidia purchased all of those 14Gbps GDDR5x that still is MIA?

Not all stuff that disappears from a catalogue must disappear for the very same reason.
14Gbps GDDR5X might have disappeared because GDDR6 was pushed forward and brings lower power consumption while starting at those speeds.
HBM2 @ 2Gbps may have disappeared because it's a higher-binned product that will be coming up in a flagship product and AMD might want to keep specifics under wraps. Just like apple wanted to keep the Pro 460 under wraps until the new Macbook was announced.

And you do not want evidence about AMD buying a whole years supply of HBM2 running at 2Gbps and 8-Hi stack that many seem to think is happening and a reason for it to be taken out of the catalogue?

I don't know who claimed that AMD necessarily "bought a whole years' supply".

AMD may have simply asked SK Hynix to put HBM2 2Gbps and 8-Hi stacks under wraps until Vega's formal release so that they could show-off during their presentation.

In return, AMD could just assure the purchase of the first successful batches (not necessarily a whole year's supply) so that SK Hynix wouldn't hurt from not advertising the product sooner.


Besides, it's not like there's a whole bunch of companies lining up for purchasing HBM2. nvidia is using Samsung for GP100 and GV100 and will most probably use GDDR6 for their consumer cards.
Who else besides AMD could be biding for SK Hynix's HBM2 at the moment?

 

[CSI PC]

Not all stuff that disappears from a catalogue must disappear for the very same reason.
14Gbps GDDR5X might have disappeared because GDDR6 was pushed forward and brings lower power consumption while starting at those speeds.
HBM2 @ 2Gbps may have disappeared because it's a higher-binned product that will be coming up in a flagship product and AMD might want to keep specifics under wraps. Just like apple wanted to keep the Pro 460 under wraps until the new Macbook was announced.

Right,
so it is a different reason when it is nothing to do with AMD.....
Can you explain why they continued develop 12Gbps and very recently into manufacturing production for GDDR5X 12Gbps very recently?
They were doing this at the period you feel they would back off for GDDR6.
Bearing in mind the very 1st GDDR6 is also 12/14Gbps from all parties....

 

[Jawed]

And you do not want evidence about AMD buying a whole years supply of HBM2 running at 2Gbps and 8-Hi stack

No, I'm asking you for evidence of what you said. You have a habit of posting facts instead of links to facts.

 

[CSI PC]

No, I'm asking you for evidence of what you said. You have a habit of posting facts instead of links to facts.

And the same can be said about anyone posting about AMD buying all the 2017 production schedule for SK Hynix HBM2 stock apart from the only one actually shown in the international catalogue that is a lower rating-spec that itself only has a Q2'17 schedule (slipped another quarter from Q1'17 catalogue).

Anyway this was back in 2016 not now regarding Nvidia and Samsung, although they are providing over 41,000 V100 just for 3 HPC supercomputer projects; just over 21,000 of which need to be in place by EOY for 1st stage, let alone other HPC-smaller projects and DGX-1 type purchases.
So demand is going to be pretty high between P100 and V100 even with higher production from Samsung.
I guess you also are complaining I never provided links many months ago for the reasons I said V100 would be announced at GTC2017 and available by late mid-summer for certain clients when everyone kept on about it not appearing in any form until 2018.

 

[Kaotik]

And the same can be said about anyone posting about AMD buying all the 2017 production schedule for SK Hynix HBM2 stock apart from the only one actually shown in the international catalogue that is a lower rating-spec that itself only has a Q2'17 schedule (slipped another quarter from Q1'17 catalogue).

No-one said they have bought it. Me (and others?) have speculated that AMD could have done it, which would explain why the 2 Gbps chips disappeared from the catalog.

They were in the book before they ever got close to production, it was an estimation.

They appeared in SK Hynix Q3/16 Catalog after they updated it during Q3/16 in August 2016 with availability set to Q3/16. That's no estimation anymore.

I guess you also are complaining I never provided links many months ago for the reasons I said V100 would be announced at GTC2017 and available by late mid-summer for certain clients when everyone kept on about it not appearing in any form until 2018.

Probably because NVIDIA representative said it after they gave first glimpse of Volta-Tegra ages ago?

edit: clarification in one point

 

[CarstenS]

Regarding GDDR6: Power consumption supposedly only goes down compared to GDDR5 non-X. SK Hynix (the only official source until there will be a JEDEC standard) says "GDDR6 is a next generation graphics solution under development of standards at JEDEC, which runs twice as fast as GDDR5 having 10% lower operation voltage." (source)

GDDR5X already is operating at said 1.35 V, which apparently also holds true for the upcoming 12 gbps chips according to the Micron specsheet. G5X does use another voltage of 1.8v (Vpp, "pump voltage") which could be lower with GDDR6 allowing for less power consumption.

As of now, though, everything points into the direction that G6 might not be using less power than G5X, only as G5 operating at 1.5 or 1.55 volts.

 

[chris1515]

From a french site

https://www.cowcotland.com/news/578...e-eclipse-nova-5-juin-tarifs-399-499-599.html

399 dollars for core model, 499 dollars for eclipse and 599 for Nova and launch date 5th of June

 

[ImSpartacus]

From a french site

https://www.cowcotland.com/news/578...e-eclipse-nova-5-juin-tarifs-399-499-599.html

399 dollars for core model, 499 dollars for eclipse and 599 for Nova and launch date 5th of June

I'm a little suspect of that.

If you follow the sources through, you go to:

WCCFTech 5/15/17 http://wccftech.com/amd-radeon-rx-v...lineup-includes-3-skus-priced-at-599-499-399/

TPU 5/15/17 https://www.techpowerup.com/233325/entire-amd-vega-lineup-reportedly-leaked-available-on-june-5th

"Digiworthy" 5/14/17 http://digiworthy.com/2017/05/14/amd-rx-vega-lineup-fastest-vega-nova/

And that's curiously close to the baseless post on reddit documented here: https://forum.beyond3d.com/posts/1980465/

So my guess is that this "Digiworthy" site just copied the reddit rumor.

 

[fehu]

4096@1.5GHz = 10.5TF
aaaaaaaaaaaand
3584@1.5GHz = 12TF
???

Completely madeup or just inverted?