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

I thought, we (as in the people in this thread) had mostly agreed upon that being the usual cherrypicking long before.

That's why I like B3D so much - apart from much to learn there is for me here: People are usually very properly mannered and everyone's bias is kept in check pretty well most of the time.

 

Rumor: AMD RX Vega HBM 2 8GB Memory Stack Costs $160, Guess The GPU Price


http://segmentnext.com/2017/05/24/amd-rx-vega-hbm2-memory-cost/
http://www.guru3d.com/news-story/amd-rx-vega-hbm-2-8gb-memory-stack-costs-160,2.html

 

According to some analysts, 4 GB of HBM(1) used on Fiji cost ~$40. Why should 4 GB of HBM(2) cost twice as much? It's manufactured on more advanced process, the price should be comparable or slightly lower, not double.

 

Time to buy some stock of Samsung and SK Hynix?

 

Was there any significant competition for HBM1?
The answer is no, but now there are GP100, GV100, Lake Crest, Vega and most of these are made for high margin markets

 

Not sure where you're getting at: Is the competition you imply good or bad for prices?

HBM came from SK Hynix alone, HBM gen2 from Samsung (until now, apparently).
Fiji used HBM, GP100 used HBM gen2.

Now, Fiji seems EOL, HBM demand as well. HBM gen2 will be used by four chips at least. Quantity and demand will go up. Supply? Not sure.

 

As you've already answered - "quantity and demand will go up. Supply? Not sure"
With limited supply and high demand for high margin markets, this comes as no surprise that prices might go up substantially

 

I'm fairly certain that cost is only relevant for 8 Hi stacks. The price doesn't rise linearly with stack height.

I expect 4 Hi stacks to be ballpark HBM1 prices.

 

Is that $80/stack supposed to be inclusive of the interposer?

 

I imagine hbm2 manufacturers would want to make profit and hence also cover r&d costs. Could be in low volume products big part of price is r&d cost needing to be covered. Would also make sense for 8hi to be more expensive due to yields and even lower volume.

 

I wondered myself, but hey, why not? Just look at wccftech.com for example, they are linked readily enough, also taken with a grain of salt as is proper for everyone. And they are based in Pakistan as well (interesting story about them, btw). Why should Pakistanis not have reliable industry sources?

Sure, it does not lend as naturally as people who are literally sitting on the manufacturers/designers doorsteps.

WRT to the 2010-era table I am also not sure. If legit, maybe it just servers the point of providing some clue how large discrepancies between even slightly different speed grades of memory or memory (HD 6950 vs. HD 6970) generation types (HD 5570 vs. 5670) can exist.

All things said, I don't buy 160 US-$ just for the memories just yet, but i would not rule it out completely as well.

More importantly - since even 160 US-$ is not a deal breaker for professional line of cards (Vega FE) - would be what two 4-hi stacks would cost that might be used in some RX Vega flavours. They are apparently available from Samsung as well, which introduces the competition.

Might be an interesting data point to compare, what the fast G5X in 12 or 16-GiByte quantities cost nowadays. They are from a single vendor as well, there is no competition and they are apparently profitably usable in sub-500-Euro cards.

 

There is a large uptick in general DRAM prices currently; Up from $3/GB a year ago to $7-8/GB in today's spot.

I don't think it is unreasonable for HBM to command at least a factor two more than DDR4, that would put the lower bound on HBM/2 at $16/GB or $128 for 8GB.

I'd expect GDDR5/X to follow the same general trend, so it won't have much of a price competive advantage.

Cheers

 

Wasn't the rumor $160 for 8-Hi stack (8 GB stack), not 2x4-Hi stacks?

 

It was related to Radeon RX Vega 8GB, so $160 for 2×4-Hi stacks:

 

Even more the reason to considering this rumor bogus - surely the price per GB can't be the same as HBM1, which was 1 4-Hi chip per GB

 

There doesn't even seem to be a source of the rumors, just popped up out of nowhere?

 

Well look like it is the guy who have do the article who have calculate thoses numbers, from where he have take them and think it was a good way to calculate that. that is another question.

I dont think anybody have seen any report about the price of HBM2, along that in fact, contract price could be way different ( specially for AMD who have been part of the developpement of HBM ( and so surely too of HBM2 ), meaning contract will reflect some R&D cost ..

 

Mostly, yes. Some things, especially concerning AMD, seem to originate at wccf as well, but do not ask me for a specific item please, it's just a feeling/remembering.

Ah. Misread the 2 in "HBM 2 8 GB stacks cost...". *doh*