During this week's podcast, Josh and the team went through an in-depth discussion of the new memory technology that AMD will be using on the upcoming Fiji GPU, HBM (high bandwidth memory). In case you don't regularly listen to our amazing PC Perspective Podcast, we have cut out the portion that focuses on HBM so that everyone can be educated on what this new technology will offer for coming GPUs.
Enjoy! Be sure to subscribe to the PC Perspective YouTube channel for more videos like this!
Josh and Allyn, What’s to
Josh and Allyn, What’s to prevent the migration of this concept to Non-Volatile DRAM? We are already seeing 3D Nand in volume manufacturing. My ignorance of process fabrication could be pointing me in the wrong direction. Your thoughts?
There is talk of using the
There is talk of using the interposer for multiple chip types. NAND is already being stacked, but not using TSVs. Using an interposer for a SOC type implementation could potentially be more cost effective than designing the entire chip to work on a single process node. So you have multiple chips made on multiple processes which are more efficient for that type of chip, but you get all of the performance of having them essentially on the same substrate.
As some one already said,
As some one already said, they already are stacked die. They stack up to 8 or maybe 16 dies in a package. This is done with wire bonding. The die are slightly offset and tiny wires are used to connect to pads along the exposed edge of the chip. This works well for flash memory since it does not need a wide interface. The number of wires is very small since they are only along the edge. The use case for flash does not require a wide interface, so there isn’t much reason to move to new tech like TSVs. Wire bonding has been around almost since the beginning of integrated circuits, so it is very mature technology. Most chips were connected by some type of wire bonding before they moved to flip-chip tech.
I am not sure what you are asking about, but 3D flash technology is referring to something completely different (sorry if you already know this). 3D flash is a different structure built on a single die. Instead of building the flash cell as a flat device, they build up many layers which results in the flash cells being laid out vertically. This gives a much larger area to store the charge which results in higher durability. The capacity is increased by going to more layers. These die are still stacked in a package in the same manner as other flash chips (wire bonded stack).
Using TSVs is a big win for DRAM since tha actual memory dies are almost entirely the array of cells. This makes them very small. The logic to control the memory is built into the bottom logic die, which could be built on a different process from the actual storage array. Flash chips probably do not need such high speed logic.
There isn’t much reason to move to TSVs for flash since the use case does not require that kind of speed. You wouldn’t want to lay out flash memory like DRAM. This is less dense and more expensive. Flash can not be used like DRAM because it would wear out very quickly. If they ever develope some type of NVRAM with durability as high as DRAM, then there would be reason to lay this out like DRAM; it would allow almost instant wake from deep sleep states among other benefits. They have been talking about such a type of memory for a long time, but an actual product has not been released.
> If they ever develop some
> If they ever develop some type of NVRAM with durability as high as DRAM, then there would be reason to lay this out like DRAM; it would allow almost instant wake from deep sleep states among other benefits. They have been talking about such a type of memory for a long time, but an actual product has not been released.
See also MIT Technology Review here re: HP:
I am not sure what specific
I am not sure what specific implementations these companies are working on. There are quite a few different types being researched, but as I said, they have been talked about for many years without much in the way of real products. We do not see any of these replacing flash or DRAM. It seems like the phase-change or magnetoresistive memory is the most promising. If you look at the wikipedia article on computer memory, there is a list of many different types in the frame on the right.
Don't know much yet but we will be looking into it.
Thanks, Jeremy. I apologize
Thanks, Jeremy. I apologize if my original question was off-topic. Nevertheless, I remain sincerely fascinated with the possibilities presented by non-volatile DRAM; and, I would be delighted to learn that you share that fascination too.
For example, I can easily see a truly INSTANT-ON PC, switching ON like a light switch. The companies manufacturing NVDRAM are certainly realizing its potential already, with specific applications.
AgigA Tech in San Diego are using super capacitors, but that solution seems (to me) much less elegant than Everspin’s ST-MRAM or the approach HP is taking:
For myself, I don’t see why a corporate policy to maximize return on R&D should impede experimentation with new ideas.
I seem to recall that HP started in a 2-car garage 🙂
“The NVDIMM SIG was formed to
“The NVDIMM SIG was formed to accelerate the awareness and adoption of NVDIMMs,” stated Jeff Chang, Co-Chair of the SNIA NVDIMM SIG. Co-Chair Arthur Sainio added, “The synergy between JEDEC and the NVDIMM SIG on the introduction of hybrid modules is accelerating the market acceptance of exciting new system configurations looking to take advantage of persistent memory at DRAM speeds. NVDIMMs are in production now by multiple suppliers, with many new product introductions in the coming months.”
With the Intel Xeon
With the Intel Xeon roadmap,mit sounds like Intel is planning on using fast NVRAM in Xeon servers. This makes a lot of sense since servers often need random access to a lot of data that is seldom overwritten.
I guess this could drive
I guess this could drive switching flash to TSVs. I am curious as to how this will look to the system though. It says one mode is backup, where (I assume) a power failure would trigger write of DRAM into nvram. The other would view the nvdimm as a mass storage device. It will be interesting to see how this is implememented through a DDR4 memory interface. Using HMC for such connections makes a bit more sense, since the memory is more decoupled from the controller through a high-speed serial interface.
Scroll down to these few tidbits:
“Resistive RAM NVM Options”
(see table under this heading, at link above)
“Many candidate next generation NVM technologies.
“Offer ~ 1000x speed-up over NAND.
“Non-volatile memory isn’t limited to the NAND flash we have today. NVMe will eventually pair with other storage technologies. 3D NAND will last us for several more product generations, but new technologies are already in development to replace it.”
I was with NVIDIA since GTX
I was with NVIDIA since GTX 8800 and I love this company very much but I must to admit that HBM is the revolution in memory technology!!! I do think that next high end GPU from AMD shall be better and faster than GTX TITAN X in every way: more ROP, more texture unit, more shaders and a lot more faster memory. Of course we need to wait benchmarks, but I think, AMD Fiji will beat every card NVIDIA has to offer in that time(till Pascal came out, maybe???) !!!
How could you love any
How could you love any company! Companies are not concerned about you, they are only concerned with the fiduciary responsibility to their investors. You are a consumer forget about love, it does not factor into any economic equation. Instead concern your self with getting the best technology for your investment in whatever hardware that you buy as a consumer. Loyalty to any corporation is simply not going to get you anywhere, and even the corporation’s investors will drop that corporation at the drop of a hat, if that means more profits could be made elsewhere. Buy the best technology for the dollar and hope that both AMD and Nvidia will be around for a good while to compete and keep the technological improvements coming, AMD’s competition in the GPU market is definitely necessary to keep Nvidia from completely dominating the market, a very bad thing for the consumer if that should continue to happen. Nvidia is just removing too much functionality from their GPU product for the Graphics Processing units to be good for other uses, and Nvidia will charge a premium to add the functionality back, a sure sign that Nvidia has a little bit too much market share.
The same thing holds true for Intel, and SOC graphics will continue to suffer if AMD can not get more of its APU SKUs into laptops, and PCs, Intel is now subsidizing its x3, x5, and x7 series SKUs to try to dominate mobile, but there are ARM competitors, as well as AMD’s own x86 based APUs with much better graphics, and Intel is not going to offer its “best Graphics” on any of the subsidized SKUs.
It’s good that Apple has chosen an AMD GPU for their new MacBook pros, as Apple was not using the most up to date Nvidia SKUs in their product, so this will give AMD some more revenue to fund the R&D for both Zen, and its Newer GPUs, and Apple never really uses the most bleeding edge SOCs/GPUs in their x86 based SKUs anyways, unless there is a valid reason. For sure Apple is probably looking at AMD’s use of HBM, and the possibility of a Zen based APU with AMD graphics combined with HBM, and this will sure allow Apple to continue with its thin and light obsession for its products.
Apple is almost certainly interested in what Jim Keller’s design team has done with the Zen microarchitecture, and Apple could request a custom Zen based APU with HBM for its future line of MacBook Pros, and AMD graphics will not disappoint. Apple is always looking to reduce the complexity of the mainboards in its SKUs, so an HBM enabled Zen SKU, will be too good for Apple to ignore, and Apple has the spare cash to order a custom APU from AMD, as AMD has a record of producing some custom gaming console parts at affordable rates. As Zen gets closer to release keep an eye out for any Apple action behind the scenes, as Apple was none too pleased with Intel’s delays of parts for its previous update.
Nvidia’s sue ball tactics against some big OEMs, in helping Apple make the decision to rely more on AMD, as Nvidia’s lawsuit involves technology that Apple uses(PowerVR), so Nvidia is taking on some very large players in the OEM market while damaging any remaining good will it has among the OEMs. I do not own any Apple products, but I do hope that AMD can get a custom Zen x86 based APU design win with Apple, and Apple’s need for thin and light, as well as HBM’s ability to allow for great space savings, high bandwidth, and lower power usage could definitely attract some custom business that would benefit not only Apple’s products, but any AMD Zen/HBM APU based laptops in the future.
An interposer with GPU, CPU,
An interposer with GPU, CPU, and HBM would make a very powerful mobile part in a small space. The silicon interposer tech may also be quite useful even for smaller parts. Current phone and tablet parts are stacked to save space, but not with silicon interposer technology. They stack the memory and possibly other components with package on package tech. I would assume that the process tech for making all of the components of an SoC is not actually optimal, so we may actually see some SoC components moved to separate chips. The interposer would allow such components to be combined together in an efficient manner. They may also be able to move some components that would have been board mounted onto the interposer instead. Communication via the interposer consumes less power than communication through the PCB, so it is advantageous to move everything they can onto the interposer, even if the device doesn’t need the bandwidth provided by the interposer. AMD’s tech is a good fit for Apple, and for mobile parts in general. Hopefully they will get some mobile design wins.
sAMSUNG AND aPPLE ARE IN
sAMSUNG AND aPPLE ARE IN PREPARatory talks to buy out existing amd investeors, or at least take a large stake. They know this will give them priority on orders as HBM will take over greater and greater sections of the Dram market, and may be in short supply for a yaer or two.
The AntiTrust authorities
The AntiTrust authorities would never allow “sAMSUNG AND aPPLE” to take over AMD, in Fact, no one with an already established SOC design/IP capacity would be allowed to takeover/acquire AMD, as AMD has an x86 license, and so no x86 maker would be allowed to acquire AMD. AMD also has an ARM license, and will be producing a custom ARM based competitor to Apple’s/other’s custom designs. You can forget about any takeover of AMD from already established competitors in the marketplace, as the main litmus test for the DOJ antitrust division for allowing/denying any acquisition is “Does the acquisition reduce competition in the marketplace!”, so stop dreaming about a takeover of AMD by any corporation/interest that already competes with AMD’s line of products. NO takeover, or controlling interest in AMD corporation will be allowed, they are free to purchase AMD stock, at less than controlling amounts, and if Samsung and Apple both purchased stock, they would not be allowed to pool there interests(if combined they had controlling interest) as they would be under watchful eyes of the DOJ antitrust division, You have not a clue of just how business works, and how the market is regulated, it is not a video game, or a sports match, get a clue!
Considering how things have
Considering how things have been going in the U.S., I wouldn’t trust the FCC to actually do their job. Apple buying AMD would be a disaster for the pc market. PC buyers are not going to switch to Apple products easily. Apple does not produce x86 processors. The ARM license is probably irrelevant. Many companies have ARM licenses to produce custom ARM chips. Samsung buying them might be okay. AMD could definately use the backing of a larger conglomerate. People seem to forget that AMD mostly comes in second, but it is essentially a two horse race. If AMD is acquired, they would need to renegotiate the x86 cross licensing agreement. I don’t think Intel can deny the license though, since this would give them a de facto monopoly in many markets. Also, technically, 32-bit x86 is deprecated, and x86-64 is really AMD64. Intel tried, and failed, to kill x86 in favor of IA-64. IA-64 is mostly dead, so attempting to deny a company buying or merging with AMD is not really an option. It would be very complicated, but I don’t think it is completely implausible.
Actually FTC, rather than
Actually FTC, rather than FCC.
So, if HP succeeds, fiber
So, if HP succeeds, fiber optics will replace copper wires, maybe?
“Case Study: RAID-on-Chip Journal Memory
“As well as for its fast-write without requiring load leveling or ECC overhead, Dell selected MRAM because MRAM non-volatility supports enhanced data center fault recovery, reduced system downtime and lower total cost of ownership. In addition, the fast write times of MRAM without load leveling or ECC overhead simplifies system design.
“MRAM in RAID Servers
“MRAM’s performance, high endurance and non-volatility allows Dell to provide industry leading RAID servers with enhanced data center fault recovery, reduced system downtime and lower total cost of ownership.”
“Case Study: Buffalo Memory SSD
“Buffalo Memory Company, Ltd., a subsidiary of Buffalo, Inc., a worldwide leader in the storage, memory and network equipment market is now developing SSD’s with ST-MRAM for the industrial market.
“Buffalo Memory has been working on a highly reliable SSD for the industrial market and will strengthen its product line by releasing ST-MRAM cached SSD with 2.5 inch form-factor and SATA3 interface in order to provide power fail endurance, higher speed and higher consistency.
“The new SSD will be launched for the applications in which ultimate reliability is required.”
“MRAM Replaces DRAM
“ST-MRAM is Replacing Persistent DRAM Solutions by Improving Latency while Reducing Design Complexity and System Cost”
“Why Is Flash Memory Running Out of Steam?”
“Crossbar Delivers a New Generation of Non Volatile Memory”
This is off topic. This is
This is off topic. This is not about alternatives to flash memory. Anyway, flash was running out of steam due to the cell size causing the durability to decrease. This was becoming worse with smaller process sizes. This isn’t that much of an issue with 3D flash though. It is built on a larger process node, with the size scaling up with more layers rather than by shrinking the size of planar cells. They will want to move to smaller processes eventually, but this extends the life of flash for a long time.
It is very difficult for new memory types (or really anything in the semiconductor industry) to replace entrenched types. This is because large amounts of research and development is being put into extending the life of tech currently in use. This is almost always cheaper than switching to a completely new tech. With the new technology, by the time they have a working product, it is often behind the old tech since so has gone into extending the old technology. This is similar to the process tech in general; it was thought that it would run out of steam a long time ago but it has been extended over and over again . We are still using 193 nm light to make 14 nm features due to immersion lithography and all kinds of other advancements. Fabs will be forced to switch to a new technology eventually, but they will not do it until they really can not push the current technology any further.