20nm Samsung and 20nm Yields

[Wynix]

The original article is "Samsung to build 20nm foundry capacity, sources say" but i feel the 20nm yields are more interesting.

Samsung Electronics plans to build up foundry capacity of 30,000 wafers operating at 20nm, signaling that it is joining rival TSMC in the process race, according to industry sources.
Samsung has stepped up the development of its 20nm process node, and already shifted a portion of R&D resources from 14nm to 20nm, said the sources. Samsung is looking to start offering 20nm process technology in early 2014, the sources indicated.
The sources did not identify how many options Samsung will offer at 20nm.
The current yield rate of Samsung's 20nm is estimated to be as low as 10%, the sources noted. At TSMC, its 20nm facilities are operating at 20-30% yield rates, the sources said.
TSMC has disclosed plans to initiate volume production of 20nm chips in early 2014. The foundry will offer only one process at the node.

 

[Alexko]

The problem is that yields expressed as absolute percentages without any qualification of the structures being manufactured aren't terribly meaningful.

 

[entity279]

Surely they aren't talking about complex CPUs here, so the numbers do sound terrible anyway, even if not precise

 

[Wynix]

The problem is that yields expressed as absolute percentages without any qualification of the structures being manufactured aren't terribly meaningful.

While you are somewhat correct, i believe it is safe to assume the structures are ARM based.
After all this is based on "sources" so we can be flexible about the technical details.

 

[Homeles]

Aren't these typical early-life yields? They still have several months to go before volume manufacturing begins, and even then, it's going to be early volume manufacturing where wafers are scarce, prices are high, and deisgns are few.

By the end of the year, TSMC will likely be yielding more than twice as well.

 

[Alexko]

While you are somewhat correct, i believe it is safe to assume the structures are ARM based.
After all this is based on "sources" so we can be flexible about the technical details.

It's just that when you see "leaked" information that doesn't really make sense, you have to seriously consider the possibility that it's just made up.

 

[Homeles]

This does make sense though... they're risk, or even pre-risk production yields. Nothing out of the ordinary here. The only interesting bit this yield information gives us, assuming that it's true, is that Samsung's 20nm node is a few months behind TSMC's, which shouldn't surprise anyone.

Remember how an AMD spokesperson said that Hawaii isn't 20nm, because it'd be slower? This is why. The process variability and yields are so poor right now, that there is literally no benefit for using the 20nm node. Simply put, it isn't ready, and we already knew that.

So no, there's nothing really interesting about the yields . The interesting bit, as per the article title, is that Samsung is expanding 20nm capacity. The mobile device war has had interesting implications on fab companies as well... TSMC is doing very well, and Samsung is trying to get a slice of the pie. Of course, so is GloFo. And Intel. And potentially others that I may have missed.

Someone's going to get hurt in all of this. It is unlikely, with all of the rapid expansion going on, that everybody is going to be a winner. So, who are the losers going to be?

 

[Silent_Buddha]

Samsung has been producing chips on 20 nm node for well over half a year now with chips actually in shipping products. 19 nm TLC NAND and 21 nm MLC NAND. I'm not sure they are actually behind TMSC. Although perhaps they are behind TMSC for ARM SOC production on 20 nm.

Regards,
SB

 

[Wynix]

Samsung has been producing chips on 20 nm node for well over half a year now with chips actually in shipping products. 19 nm TLC NAND and 21 nm MLC NAND. I'm not sure they are actually behind TMSC. Although perhaps they are behind TMSC for ARM SOC production on 20 nm.

Regards,
SB

This is obviously not talking about NAND products.

 

[Blazkowicz]

Yes flash is very different and more easy.
Likewise when you see yet again a CPU with most of its area as SRAM cache, the SRAM takes area but it is relatively trouble free next to logic parts.

 

[aaronspink]

Samsung has been producing chips on 20 nm node for well over half a year now with chips actually in shipping products. 19 nm TLC NAND and 21 nm MLC NAND. I'm not sure they are actually behind TMSC. Although perhaps they are behind TMSC for ARM SOC production on 20 nm.

You are comparing two completely different types of processes. A NAND/DRAM process has little relation to a logic process as far as yields and process maturity. Both the transistor and (intra)metal layer optimization and characteristics are so different they have little bearing on each other. Using 20nm DRAM/NAND transistors and metal stacks would result in performance so bad that it is unlikely to surpass 35/45nm logic process performance.

 

[Blazkowicz]

You are comparing two completely different types of processes. A NAND/DRAM process has little relation to a logic process as far as yields and process maturity. Both the transistor and (intra)metal layer optimization and characteristics are so different they have little bearing on each other. Using 20nm DRAM/NAND transistors and metal stacks would result in performance so bad that it is unlikely to surpass 35/45nm logic process performance.

Also why eDRAM is very hard. It's a little marvel that it is possible at all (afaik, only IBM does it on 32nm?, and even 22nm on POWER8)

 

[willardjuice]

Also why eDRAM is very hard. It's a little marvel that it is possible at all (afaik, only IBM does it on 32nm?, and even 22nm on POWER8)

It's not hard per say, it just adds months to production (can be only a month if rushed iirc) and does not yield quite as well as normal dram. It's a cost/benefit problem.

 

[Wynix]

Not sure where to post this;

Samsung Mass Producing Industry's Most Advanced 4 Gb DDR3, Using 20 nm Process

Samsung's modified double patterning technology marks a new milestone, by enabling 20nm DDR3 production using current photolithography equipment and establishing the core technology for the next generation of 10nm-class DRAM production. Samsung also successfully created ultrathin dielectric layers of cell capacitors with an unprecedented uniformity, which has resulted in higher cell performance.

With the new 20nm DDR3 DRAM applying these technologies, Samsung also has improved manufacturing productivity, which is over 30 percent higher than that of the preceding 25 nanometer DDR3, and more than twice that of 30nm-class* DDR3.

In addition, the new 20nm 4Gb DDR3- based modules can save up to 25 percent of the energy consumed by equivalent modules fabricated using the previous 25 nanometer process technology. This improvement provides the basis for delivering the industry's most advanced green IT solutions to global companies.

What are the chances we will see desktop RAM use this?
Samsung released some cheap 30nm DDR3 1600Mhz @ 1.35V that was crazy good, but supply ran out once word got out how good it overclocked.

 

[Nebuchadnezzar]

http://blogs.barrons.com/techtrader...eased-risk-of-losses-to-samsung-says-bluefin/

Ramp up up to 12000wpm in July, "going to Qualcomm". Would be their 20nm MDM9x35 modems. The Exynos 5430 seems more and more dubious getting a large volume product win, and from what I've read in past few weeks from investor analyses, Apple truly will simply skip Samsung at 20nm due to them not committing to the node and concentrating on 14nm. The Austin fab is apparently the only one ramping up 20nm while the two other Korean fabs are migrating over to 14nm.

 

[McHuj]

http://blogs.barrons.com/techtrader...eased-risk-of-losses-to-samsung-says-bluefin/

I found this interesting as well:

Obviously, the potential loss of business from AAPL and QCOM would be bad news for TSM after recently losing the AMD (AMD) GPU business.

 

[MfA]

Wasn't that just the XB1/PS4 APU's? Doesn't mean much, AMD has contractual obligations to build APUs at GloFo (they got a temporary waiver) so that was almost inevitable.

 

[Alexko]

Wasn't that just the XB1/PS4 APU's? Doesn't mean much, AMD has contractual obligations to build APUs at GloFo (they got a temporary waiver) so that was almost inevitable.

AMD has clearly stated that some discrete graphics manufacturing will be moved to GloFo. Exactly how much is unknown, but there are rumors that all upcoming 28nm GPUs (which I guess means Iceland, Tonga and Maui) will be manufactured by GloFo.

Further, Beema and Mullins are made by GF. They're technically APUs, but also evidence that GF's 28nm process is at least comparable to TSMC's in performance and power-efficiency for AMD's graphics IP.

 

[Erinyes]

http://blogs.barrons.com/techtrader...eased-risk-of-losses-to-samsung-says-bluefin/

Ramp up up to 12000wpm in July, "going to Qualcomm". Would be their 20nm MDM9x35 modems. The Exynos 5430 seems more and more dubious getting a large volume product win, and from what I've read in past few weeks from investor analyses, Apple truly will simply skip Samsung at 20nm due to them not committing to the node and concentrating on 14nm. The Austin fab is apparently the only one ramping up 20nm while the two other Korean fabs are migrating over to 14nm.

Interesting article here as well - http://www.eetimes.com/author.asp?section_id=36&doc_id=1322399

TSMC's 20m capacity expected to be 60,000 WPM by Q4.