UMC Enters Volume Production for TSV Process that Enables AMD Radeon R9 Fury X

[btarunr]

United Microelectronics Corporation, a leading global semiconductor foundry, today announced that it has entered volume production for the Through-Silicon-Via (TSV) technology used on the AMD Radeon R9 Fury X, the flagship GPU in the recently announced Radeon R 300 Series of graphics cards. The AMD Radeon R9 Fury X GPU utilizes UMC's TSV process technology and die-stacking to fuse HBM DRAM with AMD's GPU on a silicon interposer, enabling the GPU to deliver unmatched memory bandwidth of 4096-bit and quadruple the performance-per-watt over the current GDDR5 industry standard.

"AMD has a successful history of delivering cutting-edge GPU products to market," said S.C. Chien, vice president of Corporate Marketing and co-chair of the TSV committee at UMC. "This volume production milestone is the culmination of UMC's close TSV collaboration with AMD, and we are happy to bring the performance benefits of this technology to help power their new generation of GPU products. We look forward to continuing this fruitful partnership with AMD for years to come."

Bryan Black, senior fellow at AMD said, "UMC's long track record of bringing innovative technologies from the R&D stage to volume production for customer products was a compelling reason for us to engage with them as our foundry for the interposer and associated TSV technology. They have again proven their ability to execute successfully with TSV on our latest high-performance GPU, and we are pleased to have them as a valuable supply-chain partner for our exciting new line of Radeon products."

AMD's GPU and stacked HBM dies are placed on top of UMC's interposer that employs a TSV process. Through a CMOS redistribution layer and advanced micro-bumping, these ICs communicate with each other on the interposer, thus enabling the cutting-edge performance and form factor of AMD's Radeon R9 Fury X. AMD's silicon interposer with TSV is manufactured at UMC's specialty 300mm Fab 12i in Singapore.

View at TechPowerUp Main Site

 

[the54thvoid]

So the very limited supply of Fury X should improve then? Good news for AMD.

 

[Easy Rhino]

Cool

 

[Ferrum Master]

Oh it means the Beta testing is over?

 

[Casecutter]

So, perhaps it wasn't question of TSMC not being able to deliver a chip with the transistor density of Fuji after all.

 

[xvi]

Perhaps this will solve the Fury X pricing that everyone (myself included, albeit quietly) is complaining about?

 

[xenocide]

TSV is not an easy process to nail, I wouldn't be surprised if it's still months or even years before their yields are acceptable.

 

[HumanSmoke]

So, perhaps it wasn't question of TSMC not being able to deliver a chip with the transistor density of Fuji after all.

How do you get to that conclusion from the press release?
The Fury/Fury X's short supply could be down to any number (and a combination of one or more) conditions - die yield per wafer, wafer starts devoted to Fiji, packaging tolerances (both for HBM and interposer micro-bump reflow), and assembly line throughput of interposer/HBM/GPU packages.
It is actually basic statistical mathematics as far as yield vs transistor density are concerned. If transistor defects cause inoperability, higher incidence of salvage parts, and lack of commercial production quantity, then it stands to reason that the higher the number of transistors, the higher the defect rate. It then becomes a matter of applying the math to the various components ( core/uncore/PHY density per die, die per wafer, modified Poisson statistical model for wafers per wafer start. This was set out years ago for complex IC's.

If die yields are in line with every other GPU in use, shouldn't we be wondering why the only other GPU with a density approaching 15 million/mm², namely Tonga (which Fiji is derivative of), is also seen in very limited applications and usage as a full core-enabled die as the M295X.

 

[Relayer]

How do you get to that conclusion from the press release?
The Fury/Fury X's short supply could be down to any number (and a combination of one or more) conditions - die yield per wafer, wafer starts devoted to Fiji, packaging tolerances (both for HBM and interposer micro-bump reflow), and assembly line throughput of interposer/HBM/GPU packages.
It is actually basic statistical mathematics as far as yield vs transistor density are concerned. If transistor defects cause inoperability, higher incidence of salvage parts, and lack of commercial production quantity, then it stands to reason that the higher the number of transistors, the higher the defect rate. It then becomes a matter of applying the math to the various components ( core/uncore/PHY density per die, die per wafer, modified Poisson statistical model for wafers per wafer start. This was set out years ago for complex IC's.

If die yields are in line with every other GPU in use, shouldn't we be wondering why the only other GPU with a density approaching 15 million/mm², namely Tonga (which Fiji is derivative of), is also seen in very limited applications and usage as a full core-enabled die as the M295X.

They don't seem to have a lot of defective chips either though.

 

[wiak]

Perhaps this will solve the Fury X pricing that everyone (myself included, albeit quietly) is complaining about?

X has watercooling thats is a atleast a $70 value...

 

[HumanSmoke]

They don't seem to have a lot of defective chips either though.

How do you arrive at that conclusion? I haven't seen any indication of wafer volumes.

What I have seen is:
The Tonga salvage part launching before the full die. How often does AMD do that?
A general paucity of Fury/Fury X parts in the retail channel (including some pointed talk of what amounts to drip-feeding the channel by etail reps such as OcUK's Gibbo - who if I remember correctly got the lions share of the UK's Fury X supply.
While there are as many Fury X reviews as those of the (non-X) Fury, the Fury X reviews were staggered over almost two weeks as review samples were shuttled around the globe. The (non-X) Fury - despite being limited to two AIB designs (Asus Strix and Sapphire Tri-X) - again, when was the last time that happened ( freezing out Gigabyte, HIS, MSI, VisionTek, XFX, and as far as I know, PowerColor's SKU hasn't hit retail yet in any numbers) had wider coverage on its launch day.

Ask yourself - if this was another vendor with limited supply, staggered reviews with the salvage part SKU receiving more site review coverage, and limited AIB involvement that cut out a number of IHV-specific AIB's, what would your initial conclusion be?
I don't know what the yields are - and I suspect you don't either - but there is a lot of circumstantial evidence surrounding not just Fiji, but Tonga also regarding availability of full die parts. HBM/Interposer likely play a large part in Fiji's availability, but Tonga suffers no such manufacturing issue....and if the reason a full die Tonga wasn't released for discrete desktop is because of weakness against incumbent parts, why reuse the architecture en-bloc for Fiji?

 

[Sony Xperia S]

The Tonga salvage part launching before the full die. How often does AMD do that?

Tonga full is used for mobile. It is a mystery tho what exactly happens there.

I have heard that AMD used quite a significant quantity of reserve structures in their chips, so I wouldnt be surprised that yield is not that bad to enable something with predermined lower specs.. For instance, Hawaii.

 

[SonicZap]

I don't know what the yields are - and I suspect you don't either - but there is a lot of circumstantial evidence surrounding not just Fiji, but Tonga also regarding availability of full die parts. HBM/Interposer likely play a large part in Fiji's availability, but Tonga suffers no such manufacturing issue....and if the reason a full die Tonga wasn't released for discrete desktop is because of weakness against incumbent parts, why reuse the architecture en-bloc for Fiji?

Some people have speculated that AMD hasn't released a fully enabled Tonga because it'd get powerful enough to cannibalize R9 390 sales. I don't know if that's true, but it sounds like something that could be true. Right now it looks like AMD doesn't want to challenge Nvidia (leading to a price war), they could be happy with competing with the GTX 960 and 970 and having a large gap between them, making customers buy the expensive Hawaii GPUs if they want faster graphics than the cut-down Tonga.

 

[HumanSmoke]

Some people have speculated that AMD hasn't released a fully enabled Tonga because it'd get powerful enough to cannibalize R9 390 sales. I don't know if that's true, but it sounds like something that could be true.

Tonga is 359mm², the 390X's Hawaii GPU is 438mm². Doesn't it seem counterproductive to keep a larger piece of silicon in production at the expense of a smaller one offering similar performance? The reality is somewhat more straightforward...

Right now it looks like AMD doesn't want to challenge Nvidia (leading to a price war), they could be happy with competing with the GTX 960 and 970 and having a large gap between them, making customers buy the expensive Hawaii GPUs if they want faster graphics than the cut-down Tonga.

The actual reality is that the full Tonga (in the M295X minus two IMCs) has less performance than the GTX 970M (which is basically midway between a GTX 960 and 970 in specification, with the formers 192-bit memory bus) in the same power envelope. The Fury X is essentially two full Tonga GPUs, yet can't topple a GM 200 (essentially a GM 204 + GM 206), so I doubt that a full Tonga would actually have the legs to dethrone a 390X, or GTX 980.
You are right in the respect that AMD don't want a price war. Both companies have reached an accommodation in recent years - dovetailing performance and price as well as staggering major releases. That truce seems to be on the verge of crumbling due to AMD's nosediving market share, and is why the Fury is pegged at the same level as the 980 Ti, and why the 390X is priced higher than the outgoing 290X. When relative parity exists in the market it becomes a case of mutually ensured destruction - but when one is clawing for every sale, a race to the bottom in pricing means the minor player bottoms out a hell of a lot faster.

 

[Casecutter]

The Fury/Fury X's short supply could be down to any number (and a combination of one or more) conditions

I made an extrapolation for part/process and offer none as "truth"; although you shot back "Undoubtedly it TSMC not delivering" and it's based on AMD and there pushing transistor density. Now you About-Face and say it can be a combination of one or more.
Just stop, you don't know, and your arguments while have a possibility... are not fact.

This press release is saying there where problems up until this point (fact) with making the interposer (as I ask in an open question and provide a link), but you spurned such information.

Reticle limits the size of the interposer, not the yield associated with it. 65nm is a very old process, and all that is required is that the there are sufficient layers to lay down the metal traces.

So we now know this might be more the issue, but no one knows how wide-ranging. You can write all the "fluff" you want, but it has no bases in fact... until we have the next press release.

Just sit back have a nice cup of...

 

[xvi]

X has watercooling thats is a atleast a $70 value...

True, the watercooling justifies part of the price, but still, I think a lot of people were expecting something better than just matching the value of a GTX 980.
I think most people are going to look at it in a stock configuration vs stock configuration aspect. I love me some AMD, but the GTX 980 (basically) matches speed, matches price, beats on power consumption, and doesn't require finding a spot to mount a radiator. AMD's only advantage is noise (which is a nice advantage certainly) and temperatures. Despite being on water with temperature headroom, overclocking results were somewhat poor.

Steering this back on topic a bit (my fault), I wonder if high volume production will make any improvements in quality to the chips. Perhaps we'll see better overclocking potential on the Fury X.

 

[Casecutter]

It is a mystery tho what exactly happens there.

No mystery, what basically drives "Amethyst" production (full-Tonga) is how many iMac with Retina 5K displays get configured with Radeon R9 M295X Mac Edition? How many like that (~$3,000 USD) do they sell... IDK.

The Mac Edition receives the 850Mhz clock/1362Mhz memory, which is like 17.5% more clock than a regular R9 M295 (723mhz), while add 9% clock in memory for a Mac Edition.

AMD's first production of Amethyst was just to fulfill Apple's demand, and whatever AMD got as geldings they sold as R9 285 Tonga, which really never took on any huge presence in the market being still scrunch between the mass of Tahiti parts. Tonga was not in short supply because of any yield issue or density, it's actually 13.6 mill/mm² working the math, a little more than 5% higher than a GM204.

The R9 380 now such chip are called Antigua (aka Tonga), I don't know how they work the mix, at this point. One would believe TSMC is offering good yields, while Apple isn't buying near the volume, so not sure how AMD culls enough Antigua's. They probably bid their time with enough Amethyst production to sustain Apple, while slowly binning few a Amethyst chips at this point. AMD will just need some to via the anticipated release of whatever Nvidia offers between a 960-970. AMD can hold off with a 380X (Full part) at this point, Nvidia it would seem they're looking at GM204 salvaged.

than just matching the value of a GTX 980...
a stock configuration vs stock configuration aspect.... the GTX 980 (basically) matches speed,

I believe you meant the 980Ti as there's a good 13-15% difference in a stock 980 and FuryX.

 

[HumanSmoke]

I made an extrapolation for part/process and offer none as "truth"

Just as well, since your strike rate is in the single-digit percentage area.

; although no you shot back "Undoubtedly it TSMC not delivering" and it's based on AMD and there pushing transistor density.

TSMC also didn't deliver on GF 100...but then, how could they when the design of the GPU hadn't been tuned for the limits of the manufacturing process. You seem to have a very narrow focus, or an almost complete lack of understanding regarding silicon design and production.

you don't know, and your arguments while have a possibility... are not fact.

No. They aren't known fact, but I'll put my strike rate up against yours any day of the week. What is more, I at least offer some known supporting evidence, while you actually quote a source that you use to support an hypothesis that TSMC is to blame...

Is it a question of TSMC not delivering?

And the first sentence in the post you use to support your hypothesis points to design NOT fabrication:

...yields of AMD's Fiji GPU, which powers the Radeon R9 Fury series, are rather low. Exact details aren't known but due to the complexity of the chip and the high transistor density yields are believed to be considerably lower than those of NVIDIA's GM200 GPU, making it hard for AMD to lower the pricing of its Fury lineup.

NEWSFLASH!: TSMC isn't responsible for Fiji's design, AMD is.

So yeah, TSMC might not be delivering, but it doesn't necessarily follow that TSMC is to blame. In fact the article you used as supporting evidence points to AMD's design and the transistor density they shoehorned into it. Yet, conveniently, even though the article YOU quoted stated this, you declined to list AMD's design failing as a possible cause for the shortfall - instead placing the fault at the feet of everyone except AMD
If I were you, I'd research the subject before advising people to STFU...especially since that particular edict is something you are demonstrably showing you should be observing.

 

[xvi]

I believe you meant the 980Ti as the a good 13-15% difference in a 980 and Fuzy

I initially thought 980 Ti as well, but I checked Wizard's review and figured relative performance lines up with the 980, although you're right for a majority of resolutions (I forgot to check higher than 1440x900 ). The GTX 980 Ti is likely the closer comparison. Even then, I think the point still stands. Low temps and a locked voltage doesn't give us enthusiasts much.