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

David's an analyst too, and one with quite an understanding of the technical aspects as well.

What other customers of HBM1 are known apart from AMD's Fiji GPU? From the top of my head, I cannot list (m)any.

 

Is Kanter using the FE's 16GB as the basis for that or the 64's 8GB?

 

they state that they ask other sources in the industry and all quote a price within 5 bucks from each other so they were really confident in it. That;s also the reason why they didnt speculate on the GPU cost cuz sources were within hundreds of dollars from one to another so it means they can't be trusted in that matter.

The price is consistent with Anandtech article of "2.5 times more than HBM1"

 

To be fair to AMD here, the pain is industry-wide and started long ago. There's a worldwide DRAM shortage right now, and anything that's not DDR4 or LPDDR4 is getting the short end of the stick. The wafers that could be used in making HBM2 or GDDR5 are simply too valuable elsewhere (at least unless you pay out the nose for it).

 

I can believe the 2x 8-Hi cost that 150, but 2x 4-Hi? I seriously doubt that

 

But surely AMD has been offered a contract and some sort of guaranteed supply. It's not like their number of HBM chips will decrease as newer buyers come to market. It will just make it hard for them to increase production

 

Well I doubt he would wait for confirmation sources and investigate all of that just to make a sensationalist article with inaccurate information just to explain a cost of a product using other's product numbers. Makes no sense to me.

 

Yes, and that's the point: paying insane prices for memories is a lot easier when you're at margins in the 90s, like Nvidia for example is with GV100 (close to a 1000 US-$ in production, allegedly 10k+ in resale with oftentimes no third parties involved). Whether or not they pay 200 $ for the HBM or 600 $ is of little concern and will only be visible in the margins not going beyond 60 in their next quarterly report. AMD additionally has to share margins on Vega with their AIBs and with the retailers.

 

You're not wrong, of course. The only point I'm making is that HBM isn't the cause of the situation; AMD is competing with everyone (particularly handset vendors) for DRAM, not just other HBM users. So AMD would be equally in trouble even if they used GDDR5. DRAM prices are going to be through the roof for the next year, regardless of the specific sub-type.

 

Well if the information about the cost of making vega are acurate its more like the AIB and retailers share the margins with AMD

 

But without taking into account the extra power usage wouldn't GDDR be more affordable since they can use/reuse the chips/board and controllers across its entire lineup or most of it?

 

I get your point and I bascially agree. Of course every product using DRAM right now has more margin pressure than, say, six months ago. But what I'm saying additionally is, that with HBM2 you are contending for a short supply with very high margin products, making the situation even worse. Because Joe Budgetsmartphone-Vendor would not go forth an buy 16 GiByte of RAM/NAND for, say, a 100 $ for his product selling for 200 $. Now, Intel, Google and Nvidia, whose products are going to sell or are already selling right now for four- to five-digit prices, they can easily guarantee availability to meet their respective demand by paying a premium price for preferred allocation. If not - as TottenTranz said - fixed pricing contracts beforehand. And why would Samsung managers do that when they already saw the prices for HBM2 for GP100 for about a year.

So, yes, G5 and G5X are getting more expensive as well and may even run into real shortages if allocations were not contractually fixed beforehand (allocation, not price!). And even with G5 or G5X every product trying to be priced very competitively would have margin problems.

 

I imagine that the factory overclocking of the HBM2 is the most troublesome part of the cost/supply for Vega at the moment. I'm not sure what kinda process they use to determine that which 800mhz/1.2V chips are stable at 945mhz/1.35V, but it certainly must be more complicated than receiving batches of chips that already went through QC for the specified speed/voltage like AMD was (presumably) originally planning. Even taking the raw cost of HBM2 supply out of the equation, they can't get around this problem until the suppliers can provide them with the chips they actually wanted.

 

I don't know if something like "factory overclocking" can even exist....since...OC is running the chip higher than the clock the factory sets(sometimes the max reliable clock/voltages and sometimes for market segmentation).

But I'm sure not a single factory would run a chip higher than its max reliable clock/voltage set from its maker since you will only get reliabilities problems although a company may do it for the first samples but not for production models.

 

I do not think they use OC'ed HBMgen2 from Samsung but instead they go up to where they feel safe on the maximum product spec. Heck, do you also think, Nvidia would use overclocked memory on GV100 which runs its memory also at 880 MHz+.

 

Hmm, thought I read that the Samsung product is 800mhz as well. But I haven't looked at any sort of spec sheet.

 

That's because Samsung hasn't released spec sheets. They're most likely 2 Gbps chips downclocked slightly, or then just 1.94 Gbps but I think that's less likely