cageymaru

Fully [H]
Joined
Apr 10, 2003
Messages
22,060
The PC DRAM market is experiencing high inventory levels and this oversupply is expected to cause notable price declines in the 1H19. Market demand remains weak due to the off-season and the substantial amount of inventory that has carried over from the previous quarter. Contract prices of DRAM products already declined 15% month-over-month (MoM) in January and are expected to slide lower in February and March. The PC DRAM market is expected to see a 20% quarter-over-quarter (QoQ) decline in Q1 2019 while the server DRAM market is forecast to decline nearly 30% QoQ.

According to the latest analysis from DRAMeXchange, the PC DRAM and server DRAM inventory problem will persist into Q2 2019. An extended period is needed to deal with the inventory issues even if some recovery in demand occurs. New emergent technologies that will fuel demand such as 5G and automotive electronics are in the early stages of development. Thus they won't have much influence on the DRAM market in 2019. DRAM suppliers have scaled back their capacity expansions and this is expected to narrow the gap between supply and demand. The price downswing is expected to moderate over the next several quarters. The mobile DRAM market will experience less of a price swing, but the demand for products is still too weak to prevent prices from falling.

With respect to the price trends in the major application markets, PC DRAM prices are the most sensitive to demand changes and often serve as an indicator of the overall price trajectory. The average QoQ decline in contract prices of PC DRAM products was already 10% in 4Q18 and is projected to be nearly 25% in 1Q19. Currently, the average contract price of mainstream 8GB PC DRAM modules is on its way to under US$45.
 
And this is why the reports of radeon viis costing near 650 to 700 for bom are not true.

The entire industry is showing a slowdown which have impacted all memory prices downward. This translates into excess capacity at memory makers which would in turn lower opportunity cost to make hbm2 and lower the price to make.

When HBM2 on vega was first made, drr4 prices were 220 to 200 dollars for 16gb for 2400 rated memory. Its half that at the moment. Some can be found in the 80 dollar range.

Yet somehow HBM2 prices(according to rumors) have stayed the same when vega demand has fallen off a cliff, samsung has since more than doubled its hbm2 manufacturing capacity(confirmed) and a huge over supply of memory from ram to ssd from slow down in every market including datacenter, phones, pcs and laptops on top of the disappearance of mining. Improving yields over time which is natural, should have an impact on HBM2 prices. With all this somehow hbm2 prices have gone up by 10 dollars according to fudzilla.

Absolute bullshit. Those rumors were leaked to make distract from the 699 price of radeon vii.
 
And this is why the reports of radeon viis costing near 650 to 700 for bom are not true.

The entire industry is showing a slowdown which have impacted all memory prices downward. This translates into excess capacity at memory makers which would in turn lower opportunity cost to make hbm2 and lower the price to make.

When HBM2 on vega was first made, drr4 prices were 220 to 200 dollars for 16gb for 2400 rated memory. Its half that at the moment. Some can be found in the 80 dollar range.

Yet somehow HBM2 prices(according to rumors) have stayed the same when vega demand has fallen off a cliff, samsung has since more than doubled its hbm2 manufacturing capacity(confirmed) and a huge over supply of memory from ram to ssd from slow down in every market including datacenter, phones, pcs and laptops on top of the disappearance of mining. Improving yields over time which is natural, should have an impact on HBM2 prices. With all this somehow hbm2 prices have gone up by 10 dollars according to fudzilla.

Absolute bullshit. Those rumors were leaked to make distract from the 699 price of radeon vii.

What? You do realize this has nothing to do with AMD, right? Samsung, Hynix, and Micron were sued for price fixing and inflating DRAM prices.
 
When HBM2 on vega was first made, drr4 prices were 220 to 200 dollars for 16gb for 2400 rated memory. Its half that at the moment. Some can be found in the 80 dollar range.

HBM and DDR4 aren't made on the same manufacturing processes, so there's nothing saying that they're going to follow the same price trendlines (nor production/demand). Do you have any hard evidence to the contrary or it is all purely speculation?
 
And this is why the reports of radeon viis costing near 650 to 700 for bom are not true.

The entire industry is showing a slowdown which have impacted all memory prices downward. This translates into excess capacity at memory makers which would in turn lower opportunity cost to make hbm2 and lower the price to make.

When HBM2 on vega was first made, drr4 prices were 220 to 200 dollars for 16gb for 2400 rated memory. Its half that at the moment. Some can be found in the 80 dollar range.

Yet somehow HBM2 prices(according to rumors) have stayed the same when vega demand has fallen off a cliff, samsung has since more than doubled its hbm2 manufacturing capacity(confirmed) and a huge over supply of memory from ram to ssd from slow down in every market including datacenter, phones, pcs and laptops on top of the disappearance of mining. Improving yields over time which is natural, should have an impact on HBM2 prices. With all this somehow hbm2 prices have gone up by 10 dollars according to fudzilla.

Absolute bullshit. Those rumors were leaked to make distract from the 699 price of radeon vii.
HBM is not expensive because of the cost of the materials or the ram itself, that is the cheap bit, it is the manufacturing method and complexity that makes it expensive. Vega demand may have fallen off but the HBM memory architecture is proving to be very good in a number of applications where real time visual analysis is required, primarily in things like self driving cars and such.
 
Hmm.

Firstly, prices are still not low enough. Wake me when they get down to between $3 and $4 per GB. Tech is supposed to get cheaper with time, not more expensive. DDR4 is still more expensive per GB than DDR3 was 8 years ago.

Can you imagine if this were the case with anything else? What if we were paying more for SSD's today than we did 8 years ago?

Secondly, the question is, do I buy my DDR4 RAM now in anticipation of my Ryzen/Threadripper 3xxx upgrade when it launches this summer in advance to take advantage of these prices, or do I wait and hope they stay low until more details are announced, and I have a better idea of supported RAM speeds, and whether or not I am going to want a Ryzen or a Threadripper....

Hmm....
 
Hmm.

Firstly, prices are still not low enough. Wake me when they get down to between $3 and $4 per GB. Tech is supposed to get cheaper with time, not more expensive. DDR4 is still more expensive per GB than DDR3 was 8 years ago.

Can you imagine if this were the case with anything else? What if we were paying more for SSD's today than we did 8 years ago?

Secondly, the question is, do I buy my DDR4 RAM now in anticipation of my Ryzen/Threadripper 3xxx upgrade when it launches this summer in advance to take advantage of these prices, or do I wait and hope they stay low until more details are announced, and I have a better idea of supported RAM speeds, and whether or not I am going to want a Ryzen or a Threadripper....

Hmm....
You pose a very interesting question! I've been accumulating parts for a new build. I strongly leaning towards Ryzen for this upcoming build...32 GB of ram would be nice. I think I'm going to wait and see unless a drastic price drop occurs.
 
The PC DRAM market is experiencing high inventory levels and this oversupply is expected to cause notable price declines in the 1H19. Market demand remains weak due to the off-season and the substantial amount of inventory that has carried over from the previous quarter. Contract prices of DRAM products already declined 15% month-over-month (MoM) in January and are expected to slide lower in February and March. The PC DRAM market is expected to see a 20% quarter-over-quarter (QoQ) decline in Q1 2019 while the server DRAM market is forecast to decline nearly 30% QoQ.

According to the latest analysis from DRAMeXchange, the PC DRAM and server DRAM inventory problem will persist into Q2 2019. An extended period is needed to deal with the inventory issues even if some recovery in demand occurs. New emergent technologies that will fuel demand such as 5G and automotive electronics are in the early stages of development. Thus they won't have much influence on the DRAM market in 2019. DRAM suppliers have scaled back their capacity expansions and this is expected to narrow the gap between supply and demand. The price downswing is expected to moderate over the next several quarters. The mobile DRAM market will experience less of a price swing, but the demand for products is still too weak to prevent prices from falling.

With respect to the price trends in the major application markets, PC DRAM prices are the most sensitive to demand changes and often serve as an indicator of the overall price trajectory. The average QoQ decline in contract prices of PC DRAM products was already 10% in 4Q18 and is projected to be nearly 25% in 1Q19. Currently, the average contract price of mainstream 8GB PC DRAM modules is on its way to under US$45.

I'm really hopeful prices come down. It sure feels like 2009 with sky high DDR3 prices. At least back then DDR3 was new but right now it isn't the cast with DDR4 going for 4th year or something. Maybe I'll end up refreshing my gaming/workstation box if 64GB doesn't cost $500. Not to mention shortage and inflated pricing for 9th gen Intel processors and video cards. Though I think I'll be keeping my 1080ti's for the foreseeable future.
 
Maybe it'll get cheap enough to convince me to side grade to ryzen in July from my Intel system.
 
I have a feeling it'll spike back up a bit when 7nm arrives and supply gets tighter again. I think til then it's hit a bit of lull in part due to Intel's production issues and peoples reluctance. The latter part will probably fade a fair bit from the performance benefits the new Ryzen will offer and Intel might have offset a bit of it's production issues better as well by that point.
 
Remember. The more you buy, the more you save!

Oh wait, I was thinking overpriced video cards...
 
Last edited:
HBM and DDR4 aren't made on the same manufacturing processes, so there's nothing saying that they're going to follow the same price trendlines (nor production/demand). Do you have any hard evidence to the contrary or it is all purely speculation?

The memory industry is all interconnected because FABS want 100 percent utilization rates which is why we see over supplies and prices crashes at regular intervals. And in this case HBM2 and DDR4 are connected directly.

HBM2 is made on 20nm process.

https://www.anandtech.com/show/11643/samsung-increases-8gb-hbm2-production-volume

20nm is hardly cutting edge and has been used to make DDR4 since 2014.

https://www.extremetech.com/computi...m-ddr4-clears-the-way-for-massive-128gb-dimms

As a result, 20nm is the most common manufacturing to be used for DD4 currently which is why we must talk about opportunity cost.

When Vega 64 was released, DDR4 was selling in the consumer markets for 200+ for 16gb. This was because there was a DD4 shortage because capacity at memory makers was being fully utilized(because the increased demand in the telecom industry with Apple finally deciding to increase RAM in the iphone) and all memory made was sold rather than accumulated. What it meant is demand exceeded supply.

As a result, there was considerable opportunity risk for HBM2 because DDR4 was selling for high prices and was extremely profitable. Because these products are both made on 20nm, they competed with each other in production. As a result making more HBM2 meant less DDR4 which was extremely profitable at the time.

So what does this mean? HBM2 prices had to have a high selling cost to exceed the lost opportunity in revenue from making HBM2 instead of DDR4. This was reflected by the high initial cost during the Vega Generation on top of the lack of maturity of making the chips at the time.

For a while now, the price of DDR4 has plummeted and as a result the opportunity cost of making HBM2 vs DDR4 is not longer very high. Giving up production of DDR4 to make HBM2 is no longer needed because now there is excess DDR4 on the market because supply exceeds demands. What does this mean, production of HBM2 goes up, cost go down etc.

As far as complexity of HBM2, these things are not made by hand or something and its the original design and engineering that have high cost like CPU/GPU which is the fixed cost which is high. The variable cost which is the cost associated with ramping up production decreasing per unit, particularly with the price of DDR4 being so low at the moment.

HBM in the long run, was supposed to be cost competitive with traditional DDR technology, low cost HBM was supposed to bring pricing parity with traditional DDR. What this means is HBM is not some mythical unmanufacturable unicorn, it possible for yields to get better and manufacturing cost to get lower and it has certainly happened since the release of Vega.

Given that there has been about 3.5 years since HBM2 started getting manufactured, yields and performance have gotten better and this is reflected by the 50% increase in performance from the 1st HBM2 modules to the latests ones. So on top of what I already mentioned on top of the combined ramping up of production and the competition between HYNIX and Samsung, HBM2, production and selling costs have lowered.

How do we know this with real world proof? Look at the selling price of Vega 56/64. From taking a loss at 500/600 dollars to selling at 350/400 dollars for Vega 56/64 respectively.

If Vega 56/64 can be sold at 350/400 and not lose money while having 2 stacks of HBM2, Radeon VII is certainly generating a heavy profit at 700 dollars since there is 300/350 dollars to pay for 2 more stacks HBM2 and the smaller 7nm chip with whatever being left over being profit.

These 700 manufacturing costs were likely fabricated to mitigate the fact that Radeon VII as a gaming card does not offer superior price to performance vs Nvidias offerings which were criticized for their poor price to performance recently and how the market has typically picked Nvidia over AMD for price to performance. In the past AMD has typically tried to sell their cards over Nvidias with better price to performance, last gen this advantage had largely evaporated when looking at the launch price cards of both cards. As a result, AMD heavily tried to leverage the freesync vs gsync cost. Now even with that advantage gone, AMD has gone on to unnerfing some of the pro performance on top of this high BOM narrative that is getting passed around.

You have to understand AMD is using viral marketing to combat Nvidia with fake rumors that tend to benefit them. E.g the CES launch of Navi with cards with vastly superior price/performance over RTX just before the christmas shopping season. Instead Radeon VII gets launched with the high BOM cost rumors to alleits viate the poor price to performance from the value player in the videocard market. To double down on this marketing we are now getting rumors of Navi being delayed to end of year to 2020 which means its better to buy Radeon VII now since navi is too far away. The original Vega rumors were absolutely key in holding off pascal sales when they had no cards on the market.
 
I like the thoroughness of your analysis, and appreciate the effort you've gone into writing it. There's a few logical jumps that you make that I don't think can be stated in confidence. For the record: I am not making a claim on HBM2 cost, but a claim that we, as the public, have no idea of true cost of HBM2. Similarly, the analysis you make may bear out over the next 6 months, rather than retrospectively.

1.) Nothing I've read suggests that fabs can readily swap between HBM2 dram cells and DDR4 dram cells, as there's some major differences in fabrication demands even on the same node (TSV's being the major one), in fact the very press release you link points to a dedicated HBMx factory having been set up by Samsung. This would largely decouple HBM2 manufacturing costs from DDR4.
2.) Provided I'm wrong about 1, production of memory (or most any leading edge process, we're not talking 300 nm node on 200mm wafers here) is highly forecasted and pipelined (to keep that high utilization level). Retooling for HBM2 is not something done casually. Production is not fungible and timelines for demand-leveling are quite long.
3.) Similarly, demand for DDR4 and HBM are not fungible and have long development pipelines and architectural constraints (e.g. you won't see AMD rapidly coming out with Vega chips on DDR4 given the surplus). There also appears to be a number of specialty parts (AI dedicated systems) that rely heavily on HBM-based architectures. E.g. even with 4 wheels and similar overarching factory demands, the cost of a Terex Titan is loosely correlated with the cost of a Toyota Corolla. (Okay, I'm exaggerating for a bit of entertainment)

None of this is to say *what* the true HBM costs are, simply that there is absolutely no reason for them to directly correlate with DDR4's supply/demand curves.
 
It doesn't do memory makers any favors when review sites routinely don't take into account true latency in comparisons and assessments made despite the fact that it makes a fairly tangible difference depending upon how far apart TS is in different DRAM frequency/CL test examples in benchmarks. A lot of people undervalue memory because it's the middle man in performance impact on a system. It's a amazing middle man, but that's squarely where it resides it's not mass storage nor blazing speed of CPU it's in between the two, but take it away and rely on just the two and it's a ugly experience because the gap in between is far and wide. Pricing is a bit high though and has been the last few years as Ryzen gained traction hampering supply a bit and DDR4 as well become more beneficial to the system itself in it's importance. I'm thankful I future proofed while it was probably a bit under priced because a lot of people didn't exactly see it's value enough readily at that point versus today where things have changed a bit and it's usefulness has only become more obvious with better hardware complimenting it. I think it'll dip back down close to where it was priced with Skylake over time, but a bit though maybe slightly above it until DDR5 is introduced. The demand is higher today and it's relative impact more in demand. Once DDR5 arrives or shortly prior it'll probably creep down below it then shoot way up as production halts and stops. DRAM is a fickle market really.
 
When manufacturers decide it's a great idea to launch new *mainstream* cell phones and GPUs (as just a couple examples) in 4-digit price tag territory...WTF, did they think their blatant price gouged products were going to fly off of shelves?

And the DRAM manufs aren't absolved of guilt at all: every one of them conspired together and cut back production to create a false demand and jack prices up...and they are STILL stuck with excess inventory - LMAO!
 
Back
Top