Poll: ** The AMD VEGA Thread **

[silversurfer]

Come on the average guy wouldn't be using LN2 for cooling.

They are extrapolating previous data to speculate on the missing Vega data. Presumed logic would be its not possible that Fury in any feasible extreme setup could outperform Vega at room temperature, so theres a baseline to work from

The reference FuryX has 512GB of memory, the best case scenario for vega is the same 512GB/s, but more realistic given recent HBM2 rumours would be in the 410-480GB/s range.

So is HBM2 smaller cheaper to make, must be some point missing. Chart below shows advantages are with 2nd gen ?



https://www.extremetech.com/extreme...-huge-size-advantage-of-hbm-over-gddr5-memory

(logic being 16GB would likely be on the higher end gaming card)

The highest amount of memory will be on the professional use cards.
https://www.digitaltrends.com/computing/nvidia-quadro-m6000-24gb/

 

[Howling]

Is the Frontier Edition not the professional card?

 

[five8five]

So is HBM2 smaller cheaper to make, must be some point missing. Chart below shows advantages are with 2nd gen ?

HBM2.0 and the price of HBM2.0 is already sky high, vastly more expensive than GDDR5X or GDDR6. All RAM, NAND at the moment is sky high, it is why RAM, SSD, Flash prices have double/tripled in comparison to where they were a year ago.

 

[five8five]

Is the Frontier Edition not the professional card?

Yes it is.

 

[D.P.]

They are extrapolating previous data to speculate on the missing Vega data. Presumed logic would be its not possible that Fury in any feasible extreme setup could outperform Vega at room temperature, so theres a baseline to work from


So is HBM2 smaller cheaper to make, must be some point missing. Chart below shows advantages are with 2nd gen ?



https://www.extremetech.com/extreme...-huge-size-advantage-of-hbm-over-gddr5-memory


The highest amount of memory will be on the professional use cards.
https://www.digitaltrends.com/computing/nvidia-quadro-m6000-24gb/

HBM2 is bigger. Vega will only have 2 chips at higher clocks compared to FuryX.

So yeah, best case is Vega has same BW as FuryX, most like somewhat less. So using LN2 and doubling the memory clock TS just not that useful.

 

[D.P.]

Those clocks aren't convincing. Makes little sense that FE has the highest clocks, and by quite a long way. 1000mhz on the 16GB card seems highly unlikely especially (logic being 16GB would likely be on the higher end gaming card)

Some are engineering cards so hard to tell what is close to production.

While conveniently we expect pro cards to be down locked slightly for power/heat/fan/stability/life reasons, there are a few rumours the gaming cards will be lower clocked. It could be very few cores but those speeds and AMD can sell them at massive markups and use the high theoretical compute numbers as a selling point, i.e it just pips the Pascal GP100 PCI cards. All chips not making 1600 get stockpiled for the gaming cards once HBM2 is in enough volume. Gaming cards maybe at 1500mhz for fastest and 1200 for lower model.

Just have to wait and see

 

[r7slayer]

Those clocks aren't convincing. Makes little sense that FE has the highest clocks, and by quite a long way. 1000mhz on the 16GB card seems highly unlikely especially (logic being 16GB would likely be on the higher end gaming card)

Could be a dual GPU card. Same core as the 8Gb 687F:C3. It has same memory clock and it will be listed as a 16Gb card but only 8Gb per GPU and core clocks knocked down to 1000Mhz for power TDP reasons such as efficiency.

 

[Rroff]

HBM2 is bigger. Vega will only have 2 chips at higher clocks compared to FuryX.

So yeah, best case is Vega has same BW as FuryX, most like somewhat less. So using LN2 and doubling the memory clock TS just not that useful.

HBM(2) is starting to look like a bit of a turkey - Micron is supposedly starting to move GDDR6 to a 12nm "like" node with volume production sometime around the end of Q1 next year I believe (not 100% on exact timetable) which would give it significantly enhanced capabilities.

 

[Chuk_Chuk]

Wasn't HBM shortage merely a rumor? (One that I believe started at WCCFtech). Are we taking it as truth now?

Also reportedly samsung started mass production of this stuff last year.
https://www.digitaltrends.com/computing/samsung-mass-producing-hbm2/

 

[N19h7m4r3]

Wasn't HBM shortage merely a rumor? (One that I believe started at WCCFtech). Are we taking it as truth now?

Also reportedly samsung started mass production of this stuff last year.
https://www.digitaltrends.com/computing/samsung-mass-producing-hbm2/

Samsungs HBM2 is very slow. Their production from last year and even now runs at 1.4Gbps. That's what the Tesla P100s have been using.
According to videocardz the Tesla V100 will have HBM2 at 1.76Gbps.

Both are under the rated 2.0Gbps of the HBM2 spec's max.

AMD have been Partnered with SK Hynix for a very long time. Their 2.0 and 1.6Gbps was originally due Q3 2016. Then 1.6 moved to Q1 2017, and 2.0Gbps was removed from their production catalogue.

Then 1.6Gpbs was moved once again, to Q2 2017.

So yes, there has been shortages and delays.

If AMD used two 1.6Gbps modules Vega would have just over 400GB/s memory bandwidth.

2.0Gbps would give them 512GB/s, and they could never use Samsung's 1.4Gbps, as it would give them even less bandwidth, and they'd be competing against the GTX 1080s amount.

 

[Chuk_Chuk]

Maybe something changed but Samsung's press release on mass production states 256GBps (https://news.samsung.com/global/sam...on-newest-high-bandwidth-memory-hbm-interface)
which according to this (https://www.skhynix.com/eng/product/dramHBM.jsp) is 2Gbps per pin. Do you have any links showing samsung having problem with HBM2 speeds?

looking at the videocardz article (https://videocardz.com/65649/sk-hynix-updates-memory-product-catalog-hbm2-available-in-q1-2017)
They went Q3' 16 for both. Then they dropped 1.6 Gbps and only showed 2Gbps as TBD. The early this year 2.0 Gbps is dropped, 1.6Gbps return as Q1 '17.
First if they were having trouble with 2.0 Gbps why was 1.6 dropped first? Surely 2.0 should have dropped first. In fact why did they still continue to advertise a product that they are having problems with. That doesn't sound right.

To top it off why would you advertise something that your not confident in taking to the market. Considering this is the world of OEMs and multi million pound contracts false advertising is going to land you in a lot of legal trouble if you can't fulfill contracts. (Not like the consumer space where they can barely afford a lawyer).
Consider that SK hynix refers to them as availability, and this in the product catalog; it is possible that this is what they have to offer people to buy and stuff disappearing is simply because someone else bought the rest?

I have a hard time believing that SK hynix or Samsung is having trouble producing 2.0 Gbps chips because if they were they, would never have advertised them. I expect that they would only advertise products if they were confident that they could deliver. If they weren't confident then i expect the products to merely be back room talks.

If there is a shortage or speed issue i don't think it is the chips themselves, I think its another part of the system letting it down.

 

[N19h7m4r3]

There's not a single device out that runs HBM2 a 2.0Gbps or 256GB/s.

The Jedec spec for HBM2 is what Samsung and SK Hynix show on their sites. Just as you linked. It's not what they can manufacture in mass quantities.

The Tesla and Quadro GP100 all have Samsung 1.4Gbps. And Volta Tesla V100 is only 1.7Gbps.

The Quadro was released this year and is still only using Samsung 1.4Gbps, nothing from Samsung out in the world bar press releases a has 2.0Gbps

If they had 2.0Gbps AMD would be using that for Vega, as the original spec assumed 512GB/s and they can barely managed 480GB/s on the Frontier Edition.

Also 1.6Gbps wasn't dropped first. Re read that article. 2.0Gbps was dropped and changed to TBD, while 1.6Gbps was moved several times. Never dropped.

TBD means to be determined. As they weren't able to determine a mass production date anymore, but they could for 1.6Gbps.

Here's the latest on available HBM2 that's been found.
https://videocardz.com/69504/sk-hynixs-updated-memory-catalog-features-hbm2-and-gddr6

 

[akarypid]

It definitely seems like there were delays and issues (which may have impacted Vega), but surely these should be ok now?

I mean, how bad can things be right now if Apple is putting Vegas in iMac Pros? Even if they're "pro" boxes, Apple must be selling significant volumes of those right? Surely they must feel like things are stable enough now?

Or maybe it's could be a matter of clout (as in, partners were dragging their feet because they thought there would not be enough sales)? In that sense, Apple would have no concerns (they probably have more clout than any other OEM).

Either way:

Less than 18 days left now chaps! Vega FE will soon be with us!

 

[N19h7m4r3]

It definitely seems like there were delays and issues (which may have impacted Vega), but surely these should be ok now?

I mean, how bad can things be right now if Apple is putting Vegas in iMac Pros? Even if they're "pro" boxes, Apple must be selling significant volumes of those right? Surely they must feel like things are stable enough now?

Or maybe it's could be a matter of clout (as in, partners were dragging their feet because they thought there would not be enough sales)? In that sense, Apple would have no concerns (they probably have more clout than any other OEM).

Either way:

Less than 18 days left now chaps! Vega FE will soon be with us!

The main thing about Apple is they announced the iMac Pro with Vega is coming end of the year. Not within the next few months.

Since we know the Intel 12-18 core CPUs aren't due till end of the year as well, it could be a delay because of that, or Radeon Pro Vega.

Although...Apple lists the memory bandwidth of 400GB/s, and that means AMD are putting 1.6Gbps HBM2 on those cards. Since the Frontier Edition is 480GB/s.
Which means if HBM2 above 1.6GBps is still an issue, Apple might not have an issue using cards made using those for their needs.

So it could be that Apple are getting specially made Radeon Pro Vega cards just for there use. Similar to how Apple's Radeon Pro 460 was a fully unlocked Polaris 11 a full year because AMD launched the RX 560 on the desktop side.

 

[akarypid]

The main thing about Apple is they announced the iMac Pro with Vega is coming end of the year. Not within the next few months.

Since we know the Intel 12-18 core CPUs aren't due till end of the year as well, it could be a delay because of that, or Radeon Pro Vega.

I really hope it's just the CPUs that are the issue. I mean, we know that AMD will be selling the Vega FEs already in July, so that means that on that side it can only be a matter of building up inventory, which Apple doesn't really mind (like you said, CPUs are end of year anyway).

Or it could be both, so while Apple waits for Intel CPUs, AMD is building Vega inventory.

But that's also partly the point I was trying to make. I'm sure that if it was a matter of HBM supply, the deal with Apple would automatically fix it. If there's any OEM who can 'call Hynix' and say "Stop whatever you're doing: you're now ramping up on HBM2" it would be them.

 

[N19h7m4r3]

But that's also partly the point I was trying to make. I'm sure that if it was a matter of HBM supply, the deal with Apple would automatically fix it. If there's any OEM who can 'call Hynix' and say "Stop whatever you're doing: you're now ramping up on HBM2" it would be them.

I agree. Apple has the clout to pull strings and Hynix and Samsung in regards to HBM2.

 

[Chuk_Chuk]

There's not a single device out that runs HBM2 a 2.0Gbps or 256GB/s.

The Jedec spec for HBM2 is what Samsung and SK Hynix show on their sites. Just as you linked. It's not what they can manufacture in mass quantities.

The Tesla and Quadro GP100 all have Samsung 1.4Gbps. And Volta Tesla V100 is only 1.7Gbps.

The Quadro was released this year and is still only using Samsung 1.4Gbps, nothing from Samsung out in the world bar press releases a has 2.0Gbps

If they had 2.0Gbps AMD would be using that for Vega, as the original spec assumed 512GB/s and they can barely managed 480GB/s on the Frontier Edition.

Also 1.6Gbps wasn't dropped first. Re read that article. 2.0Gbps was dropped and changed to TBD, while 1.6Gbps was moved several times. Never dropped.

TBD means to be determined. As they weren't able to determine a mass production date anymore, but they could for 1.6Gbps.

Here's the latest on available HBM2 that's been found.
https://videocardz.com/69504/sk-hynixs-updated-memory-catalog-features-hbm2-and-gddr6

Its the availability date not the mass production date, unless they mean the same thing in the tech industry (which i doubt). I don't think they are the same thing because i don't see how it benefits SK hynix to publicly reveal and give a quarter by quarter update on the start of mass production for their products, seems like behind closed room stuff to me. If I was buying stuff from them, I wouldn't care when they start mass production I want to know when they can deliver the stuff to my door. Based on that, I am assuming that, that is what the availability date is, and it is for customers who order from the catalog.

Also would you really offer something to customers that you weren't confident in being able to mass manufacture? It makes no sense, why offer something to customers when you are struggling to produce them?

You are right, there is something strange as to why none of those cards run at 2.0 Gbps. But I don't think it is down to supplies/manufacturing capabilities of SK hynix/Samsung. If there is a problem with 2.0 Gbps HBM2 I don't think it is to do with supplies, it something else that can only be understood by those on the inside.

 

[N19h7m4r3]

Also would you really offer something to customers that you weren't confident in being able to mass manufacture? It makes no sense, why offer something to customers when you are struggling to produce them?

You are right, there is something strange as to why none of those cards run at 2.0 Gbps. But I don't think it is down to supplies/manufacturing capabilities of SK hynix/Samsung. If there is a problem with 2.0 Gbps HBM2 I don't think it is to do with supplies, it something else that can only be understood by those on the inside.

Who says they weren't confident, and then issues appeared? Those catalogues are out out along with the Quarterly earnings reports and investor calls.

Most are estimations hence TBD, and saying broad terms like Q2 2017; like AMD's growth and future margins, and performances are discussed at those. They can't always deliver on it all though.

What we do know is Samsung does not even put out similar catalogues like Hynix, and so far the only HBM2 manufactured by them has been 1.4Gbps modules. Nothing else has appeared anywhere.
The only other information we have on Samsung's latest is that the Tesla V100 specs show HBM2 at 1.76Gbps.

Nevermind Hynix failing to deliver on their own roadmaps and catalogues.

Remember Vega DOOM demo was from December last year, showing performance equal to an overclocked GTX 1080. Surely if HBM2 was in ready supply we'd have those cards?

https://videocardz.com/64706/amd-vega-doom-4kultra-gaming-performance-demo-possible-specs

https://videocardz.com/63700/exclusive-first-details-about-amd-vega10-and-vega20

Even back then Vega was stated to have 512GB/s memory bandwidth according to AMD employees at the AMD Tech Summit. To get that Vega needs 2.0Gbps HBM2.

Yet THE best Vega card for specs, from leaks, previews, demos, engineering samples, and Apple, shows under 512GB/s. The best is Vega Frontier at 480GB/s, and Apple's Radeon Pros Vegas will only have 400GB/s.

AMD needs every advantage they can get, and if they had access to 2.0Gbps as per their original Vega spec, they'd be using that on their Halo product launching end of this month.

 

[akarypid]

AMD needs every advantage they can get, and if they had access to 2.0Gbps as per their original Vega spec, they'd be using that on their Halo product launching end of this month.

Wait for... "Vega FEp"?

 

[N19h7m4r3]

Wait for... "Vega FEp"?

Does the p stand for Patience?