Tezzaron: You’ve got our attention now

Tezzaron Founder and CTO Bob Patti delivered a rousing keynote at the Silicon Valley Magma User Group meeting back in 2010, talking about his company’s 3D memory technology and how it offered a solution to the increased demands for on-chip capacity. I spoke to Patti following that speech, with details of the conversation posted here.

In the past 3 years, things have only gotten worse with respect to memory demands, so one might think Tezzaron’s solution is in even greater demand today. The problem, of course, is that Tezzaron Semiconductor is not the only company offering something that looks like ‘3D memory.’

In fact, 3 weeks ago at MemCon 2013 in Silicon Valley, I attended a keynote given by Micron Technology’s Mike Black singing the praises of his company’s Hybrid Memory Cube. Sitting in the audience, I tried to compare and contrast the Micron technology with what I believed to be the Tezzaron solution.

Happily, Tezzaron had a booth at MemCon, so it was possible to talk to somebody from the company about how they viewed Micron’s competing technology. Unfortunately, Tezzaron VP David Chapman was surrounded by a mob of interested people when I got to the booth, so I took his card and arranged to talk to him later about my many questions.

My number one question: Given the size of Micron and the ecosystem of partners they’ve assembled, the Hybrid Memory Cube Consortium, is Tezzaron winning the battle of technology superiority, but losing the war for market share?

Tezzaron’s Chapman was not offended by my question. Instead, he started by framing his answer with a description of current market needs and continued from there. The following is a transcript of his comments.

********************

Bursty data dumps …

Per David Chapman, “The real distinction for Tezzaron is transaction rate, a figure of merit which may be unfamiliar to almost everybody. For the last 20 years at least, the memory industry has focused on improving bandwidth. Rightly so, because the memory industry has focused on computer applications, doing things that make things like PCs work well. In a broader context, making things that run microprocessors work well.

“Bear in mind that when I say computer applications, I don’t just mean PCs. I mean everything from mainframes to cellphones, things that run code on a processor of some description. Because they’re running code, the address pattern that’s going to memory devices demonstrates Locality of Reference, because computer programs do things like subroutines.

“So when running a program, there’s a real good chance that if you’ve asked for a piece of data at one address, pretty soon you’ll ask for data that’s near to that address, which is why caches work.

“As the processor designs have become more and more dependent upon caches that take advantage of that locality of reference, the task for main store has become more focused on doing line fills for the caches, and those are bursty transfers.

“In compute land, you issue one address and you get great big chunks of data back to do line fills and caches. The primary challenge has been to do those bursty data dumps into caches, and that’s where all the architectural work and memory guys’ work have been going in the last 20 years.”

********************

More than a fire hose …

“Networking customers, however, are up against a radically different problem. They have very little, if any, Locality of Reference in their memory traffic. Instead, they need much more random, small-transfer performance.

“As a consequence, for the last several years there have been a couple of different memory technologies that have been very useful to the memory guys that are of no use at all to the computer guys; specifically, discreet Fast SRAM and low-latency, high-transaction rate DRAM – for example, Micron’s RLDRAM.

“What those kinds of memories do well is random access. You can do small transfers, and you can launch small transfers very frequently – that’s what we mean by high-transaction rate. It’s like comparing a fire hose to a machine gun.

“Most of the memory guys have been working on a fatter and fatter fire hose, but what the networking guys need is a machine gun. They need lots and lots of very high rate individual small things to come out of memory.

“Micron has built a pretty good fire hose, but Tezzaron is building an awesome machine gun.”

********************

Radically different approach …

“For the last 40 years, everybody has been trying to do integrated circuits by putting more and more stuff into a single die, but broad industry consensus says this is getting really problematic. When you look at the memory performance trends over time, I think the results bear out that thought.

“Several years ago, Bob Patti saw an opportunity to do something completely different, to get off the integration horse, turn around and go in a different direction to get better performance. While everybody else doing 3D is taking full functional die, building them on a wafer, checking to find the good ones, and then cutting them up and stacking the good die, Tezzaron doesn’t do that.

“We have taken a radically different approach by building dis-integrated memory. There are no full-function memory die in the stack, only pieces of memory. We have the I/O circuits in the bottom die in the stack, the next die in the stack has the decoders and drivers and stuff like that, and all the die above that pretty much just have bits – bits and access transistors.

“If we took one of those bit die and put it on a tester, there would be nothing to test. It’s not a RAM, it’s just a sea of bits. The thing doesn’t become a memory device until it becomes stacked.

“Again, it’s a dis-integrated memory which allows us multiple advantages, not the least being that we get to build the various circuits that you need to have in a full functional memory in the type of silicon best suited to that type of circuit.”

********************

High interconnect density …

“Ask anyone if they want to build a 30 Gigabit SerDes on a DRAM process, and you’ll get an emphatic No! It makes no sense at all. You want to use a pretty hot logic process to do something like that, but we can because our bottom I/O die is in a CMOS process.

“The only die built on the DRAM process has access transistors and bits, so it’s about going to dis-integration. Focusing the circuits on the particular kind of process they need is allowing us to get tremendous performance advantages compared to what anybody else is doing.

“And this is only possible if you’re willing to do something that, by almost any standard of measure, is completely crazy – you have to be willing to bond untested wafers.

“The 3D process that I described earlier – the building of the wafer, the die on the wafer, testing the die to find the good ones, cutting it up and stacking the good die – again, Tezzaron doesn’t do that.

“What Tezzaron does is build a wafer and then bond it to another wafer, and another and another…over and over again. We build the entire stack in wafer form first, and then cut it up. Invariably, the first time people hear this they say, ‘You guys are nuts – that cannot possibly work!’

“And if we did things the conventional way, that might be true, but we don’t. We don’t do things the conventional way. For starters, we don’t use TSVs, at least not what people normally consider to be TSVs.

“Instead, we use tungsten contacts. We can because we bond our wafers together first – we know how to do that – and then we thin them. Also, while everybody else is trying to handle thinned wafers, which are devilishly hard to handle, we can thin one wafer, but leave the other one full thickness.

“Our thin wafers are only 5 or 5.5 micron thick, and because they’re so thin we don’t have to use copper TSVs. Because our wafers are so thin, we can use tungsten contacts instead. They’re way smaller – way, way smaller – so our contacts are roughly only one micron wide.

“Where copper TSVs are more like 10 microns wide and 50 microns tall, ours are only 5 microns tall. That means ours are electrically more attractive. They have radically lower parasitics associated with them, and we can have whole lot more of them [than in a typical TSV-based device].

“It’s because our interconnect density through the stack is so high that we can do this crazy thing of assembling wafers together first, without having gone and found the good die. It’s because the interconnect density is so high, we can also do effective post-assembly repair.”

********************

Doing post-assembly repair …

“Tezzaron uses a lot of different methods to do the repair. For starters, we use a simple repair method, we do traditional eFUSE repair. We have enough routing and interconnect through the stacks, spares and extra channels, that we can identify bad spots and route around them.

“In addition, there are CAMs in the stack that assist in the repair process. In all, there are a series of different techniques that come into play – some traditional, and some less traditional.

“The end result is, for everybody else when they build a stack, their yield gets worse and worse. But with Tezzaron, the more wafers in the stacks, the better the yield. Our yield improves as we add more wafers to the stack, because that’s how we pick up more spares.

“And, if you’re going to be in a 3D world, clearly you want things to get better with 3D, not worse.”

********************

You’ve got our attention …

“Yes, it’s true Micron and other companies have built large ecosystems around their memory solutions – and that means we’re facing an interesting problem – but the performance we’re getting is so dramatically better than anything anybody else is getting, there’s really no competition.

“The fact is, unless someone is willing to adopt the Tezzaron methodology for building memory, they don’t have a prayer of matching our performance. It’s not as if someone can go build a device in a traditional 3D way and expect to get these kinds of results. That’s simply not possible.

“So, if by an ecosystem, you mean having the ability to second source a product, someone will have to step up to the plate and say they’re willing to do this crazy thing, bond untested wafers. So far, nobody’s been willing to do that.

“But again, the cool thing is that the performance we’re getting is to radically high, from a transaction rate point of view, compared to anything else on the horizon from any other source, I think we’re going to see significant adoption going forward, despite the technology being sole sourced.

“Yes, I realize this is a wild claim, but we’ve talked to lots of networking guys and they all see the potential. To quote one of them that I cannot name, he looked across that table at me recently and said, ‘Well, you’ve certainly got our attention!’

********************

Just-in-time technology …

“Of course, that is not a commitment to buy, but we’re showing them something that they simply have not anticipated was even remotely possible and it’s having a radical impact on their thinking about what can be achieved. The Tezzaron technology is just stupendous compared to what they’ve seen up until now.

“The space and power savings are gigantic, and the performance differences are so huge our potential customers will be leaving themselves at a tremendous disadvantage if they don’t do this. They’re asking themselves, ‘Can I risk not building a system with this technology, if my competitor is going to use it?’

“We know there are a lot of big guys out there – most of them running projects and back-up projects – and what we’re seeing here in 2013.  We will end up on one of those projects. We expect it will work out great and at that point, there will be pressure to establish an ecosystem and that’s how the Tezzaron technology will propagate.

“You ask about winning the battle, but losing the war. Actually, I think the battle hasn’t really happened yet. This technology is going to be in the hands of commercial customers by 2015, and we’ll be seeing prototypes well before that.

“In fact, we’re arriving just in time for these customers. Up until now, these types of companies couldn’t even conceive of building a Terabit router, for instance. Tezzaron is arriving just in time to make that a reality for them.”

********************

The fun factor …

“The craziest thing about Tezzaron is that we have recently bought a fab…in Austin, previously owned by SVTC. We were running stuff at SVTC, and when they decided they were going to make other arrangements, the Tezzaron guys came in and bought it.

“Now the outfit is called Novati Technologies, a wholly owned subsidiary of Tezzaron, whose primary responsibility is to be an excellent fab. They run several lines of business, with the 3D work for Tezzaron being just one of them.

“This is actually an old Sematech facility in Austin, and it’s really big! It’s not a mega-fab, but it’s a good-sized facility sitting on 40 acres, with the associated physical plant capable of handling a substantially larger clean room operation than it’s currently connected to. We believe the Novati facility will provide us with all the manufacturing capacity we will need for some time to come.

“At Tezzaron we’re not worried about winning a war. We can see we’ve already won. So we’re having fun, and we’re having lots of it.

“And I’m telling you, in all my years in semiconductors, I have never had this much fun!”

********************

Tags: Bob Patti, bursty data dumps, David Chapman, DRAM, eFUSE repair, Fast SRAM, Hybrid Memory Cube Consortium, Locality of Reference, MemCon 2013, Micron, Novati Technologies, Tezzaron, Tungsten contacts

This entry was posted on Thursday, August 29th, 2013 at 10:00 pm. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.