Just when you thought it was safe to start GPU shopping, with demand from miners dropping off somewhat, the NAND shortage is set to crank up prices again. First time miners have realize they are not about to become overnight billionaires and the dedicated miners have already picked up their GPUs; unless they just picked up this board, so there was some hope GPU prices might descend closer to their original MRSP. Unfortunately the suppliers of VRAM have shifted their production capacity more heavily in favour of server memory and RAM for smartphones which has lead to a dearth of VRAM. DigiTimes reports you can expect the price of NVIDIA cards to jump from 3-10% at the end of the month.
AMD's new offerings will not be effected by this; few and far between are the servers or phones which use HBM2. It would be interesting to discover that part of their original pricing took this into account; not that it matters overly as their original pricing statement has been tossed.
"With Samsung and SK Hynix cutting their memory supply for the graphics card segment, August quotes for RAMs used in graphics cards have risen to US$8.50, up by 30.8% from US$6.50 in July. Both memory suppliers have allocated more of their production capacities to making memories for servers and handsets, reducing output for the graphics cards segment and fueling the price rally."
Here is some more Tech News from around the web:
- Is it possible to control Amazon Alexa, Google Now using inaudible commands? Absolutely @ The Register
- A Functioning 3D Printer For 10€ @ Hack a Day
- Google Updates: Hardware, Firmware, The Firm @ The Inquirer
- World's first crowdsourced tablet, Eve V, is taking aim at the Surface Pro @ The Inquirer
- Gather round, kids, and let's try to understand the science of 3D NAND @ The Register
- A Game You Control With Your Mind @ Slashdot
- AVM FRITZ!Box 7560 AC1300 VDSL/ADSL Modem Router Review @ NikKTech
I grabbed a nice zotac 980 ti
I grabbed a nice zotac 980 ti omega for $280 in ebay… prices are still down too.
Although id rather get a Vega 56 and a freesync monitor, yeah its a lateral move towards AMD.
Miners should go to jail for
Miners should go to jail for running the biggest ponzi scam ever known to mankind
What?
What?
well the bollengekte was far
well the bollengekte was far bigger depending on how you look at it, but I see where you are going with this.
https://en.wikipedia.org/wiki/Tulip_mania
“AMD’s new offerings will not
“AMD’s new offerings will not be effected by this; few and far between are the servers or phones which use HBM2.”
Not really and you are only half right there with the phones not getting HBM2! BUT AMD is making some Epyc server deals in the Asian markets and the Radeon Pro WX/Radeon Instinct MI25 SKUs(At 16GB of HBM2 each) accelerator/AI SKUs will go into those Asian Search Engine companies Epyc servers and Asian Amazon equivalent companies server farms that will be using plenty of Radeon Pro WX/Radeon Instinct MI25 GPUs right along side the Epyc CPU SKUs.
So that’s Plenty of servers Getting HBM2 beased accelerator GPUs, but No phones, that will get those Radeon Pro WX/Radeon Instinct MI25 with their 16GB/Higher amounts of HBM2.
Look at your Vega 64 and 56 consumer SKUs and know that the best Vega 10 Dies are going for the Radeon Vega WX/Instinct SKUs to complement those Epyc CPU Server SKUs. And in that acceleration market and AI market that will get the HBM2 first and get the Very Cherry Vega dies, and HBM2 dies, going to the professional AMD GPU SKUs where the markups are way better that any consumer CPU market can produce. The Best Vega 10/HBM2 dies are already spoken for for quite some time as AMD’s Epyc CPU SKUs will get those Radeon Pro WX/Instinct accelerator SKUs to complement the Epyc CPUs in the server rooms in large numbers with HBM2 in short supply for a good while.
GPUs are only going up in price from both Nvidia and AMD until the HBM2 and GDDR/whatever supplies can catch up with demand.
“AMD’s new offerings will not
“AMD’s new offerings will not be effected by this; few and far between are the servers or phones which use HBM2. ”
I suspect that Nvidia can sell as many GV100s as they can produce to meet demand and can effectively name their own price. Unless AMD have negotiated a new guaranteed-supple/preferred-client arrangement as they did with HBM1, they may be priced out of the market for HBM2 due to enterprise demand for high performance GPUs for inference.
Samsung can Supply enough
Samsung can Supply enough HBM2 for AMD’s and Nvidia’s professional market needs but it’s AMD consumer market that will have to wait for SK Hynix to get its HBM2 production up there. AMD is not going to worry about its consumer gaming market as much as AMD is needing to have an ample supply of HBM2 for its Radeon Pro WX 9100 and Radeon Instinct Professional GPU SKUs that will be sold alongside AMD’s Epyc Server CPU SKUs to those large cap Asian Companies that AMD/Server partners has made the big server room deals with out east.
I’m sure that both Nvidia and AMD want SK Hynix to get their HBM2 to market ASAP so Samsung will be pressured to lower its HBM2 pricing. The problem with all HBM2 production currently is, whatever the current amount of HBM2 is sold, both Samsung and SK Hynix have large R&D and production rampup costs related to HBM2 that have to be Amortized over a realtively small HBM2 production unit amounts. So both AMD and Nvidia will be having to pay more until that HBM2 R&D/Production Ramp-Up costs are fully amortized by both Samsung and SK Hynix.
The more HBM2 units sold the smaller amount of that R&D/Production Ramp-Up costs can be factored into the cost of each HBM2 stack sold. So the more users of HBM2 the better it will be for any future costs cutting as that R&D/Production Ramp-Up costs for Samsung and SK Hynix gets paid off.
Also with regards to Samsung’s better binned HBM2 dies for both AMD and Nvidia is that the best performng HBM2 stacks will go into AMD’s and Nvidia’s high margin professional SKUs while the consumer SKUs for AMD(currently) will get any dies that do not make the grade and have to be clocked lower than the top binned HBM2. Ditto for the very best thermally performing Vega 10 dies.
Nvidia has specific die variants for its professional SKUs that are different than its consumer variants but AMD’s Vega 10 dies are used across its entire GPU product stack and that’s one reason why AMD’s consumer SKUs have so much more compute relative to Nvidia’s consumer GPU SKUs. Maybe when AMD/RTG start getting enough revenues there can be a professional GPU only micro-arch and a stripped of compute consumer/gaming GPU micro-arch form AMD that can save on the power used by not having as much compute with loads of TMUs/ROPs to fling that frames out there like Nvidia does.
AMD will not be priced out of the market as AMD will pass those HBM2 costs along to its professional market GPU end users and Samsung has plenty of resources to meet Nvidia’s and AMD’s Professional market HBM2 needs. So until SK Hynix gets its HBM2 production online at full capacity then it will be AMD’s consumer Vega SKUs that will be in short supply. The professional market will pay to get its GPU needs met and that expense is just a tax write-off while the consumer markets can not afford to pay or write-off any GPU/other related costs.
Well yes, there is an accross
Well yes, there is an accross the board ram shortage, and its here to stay. Production cant sate the needs of advances in processing power.
It affects amd & NV equally…or does it?
AMD at least have a partial imperfect solution in vegas hbcc chip, which cleverly manages a pool of lesser resources like unused system memory, and juggles them to simulate a much larger gpu memory.
System ram is also limited in size, but hbcc also alows storage to be used in its resources pool. It seems foolishly slow for the job of simulating memory, until modern nvme raid array speeds are considered.
a raid array of 3x 500gb samsung 960 pro nvme ssdS should yield about 8GB/s read & 6GB/s seqential write, or about as fast as an 8 lane pcie card can interconnect w/ system memory.
It has merit as is, but the weak link is linking the components via the shared & circuitous pcie3 bus.
Vega has the ability, as shown in their production vega ssg cards, to bypass the pcie3 bus, and put the nvme raid array, directly on the gpu card and its Fabric bus.
In this form, amd have a card that gives a pretty good imitation of a vega card with 2TB of gpu ram, tho it only has 16GB of hbm2.