Info 64MB V-Cache on 5XXX Zen3 Average +15% in Games

[Kedas]

Well we know now how they will bridge the long wait to Zen4 on AM5 Q4 2022.
Production start for V-cache is end this year so too early for Zen4 so this is certainly coming to AM4.
+15% Lisa said is "like an entire architectural generation"

 

[Vattila]

I'm more or less curious why they've been drilling chips since the beginning if they were going to implement it this late in Zen 3's life.

The hybrid bonding technology used for V-Cache is cutting-edge, and the timing is most likely very dependent on TSMC's packaging roadmap (qualification and capacity build-out).


AMD 3D Stacks SRAM Bumplessly – WikiChip Fuse

Note that there is no "drilling" involved. The vertical connections needed to connect the V-Cache, i.e. through-silicon-vias (TSV), are made by deep etching with mind-boggling precision (through the use of directed ion beam sputtering). I suspect this step of the process may not be performed for the current CCDs without V-Cache.

What is the Bosch Process (Deep Reactive Ion Etching)?

www.samcointl.com

PS. Introductory paper on TSV:

http://www.ziti.uni-heidelberg.de/z...eminar/2012-Solience_Ngansso-3D_TSV-Paper.pdf

 

[Doug S]

All for using some analytical skills that pointed to AMD and TSMC have all the incentives in the world to put the highest priority on this product, which would likely result in it being shipped this year.

Why does TSMC have any incentive to put "the highest priority" on this? They don't get care whether AMD beats Intel or not, they manufacture their customers products as spec'ed in the order they've committed to. They have zero reason to invest extra resources to bring something for one customer in ahead of schedule.

 

[moinmoin]

Why does TSMC have any incentive to put "the highest priority" on this? They don't get care whether AMD beats Intel or not, they manufacture their customers products as spec'ed in the order they've committed to. They have zero reason to invest extra resources to bring something for one customer in ahead of schedule.

I somewhat disapprove of the wording choices you two made there. It's not about "highest priority", "extra resources", "ahead of schedule" etc. It's about introducing new tech, and usually TSMC, producing no chips of its own, is working together with some early adopter customers before the tech reaches general availability. All indications so far points to AMD being one if not the early adopter for what they dubbed X3D before. There is no "getting ahead of schedule" going on here, it's naturally happens and has to happen before the general availability of the tech. It's a step to getting there to begin with.

 

[Joe NYC]

The slide says Data Center Roadmap and underneath it says EPYC, so naturally it would be Genoa.
I doubt AMD will be open ever again for client roadmaps

In one of the interviews with Lisa Su post earnings, she phrased one of her answers in a manner from which I implied that Genoa will be leading Zen 4 product.

These new roadmaps point in that direction as well. AMD in fact removed Zen 4 from client road map (for now).

 

[Joe NYC]

The hybrid bonding technology used for V-Cache is cutting-edge, and the timing is most likely very dependent on TSMC's packaging roadmap (qualification and capacity build-out).
View attachment 48421

AMD 3D Stacks SRAM Bumplessly – WikiChip Fuse

Note that there is no "drilling" involved. The vertical connections needed to connect the V-Cache, i.e. through-silicon-vias (TSV), are made by deep etching with mind-boggling precision (through the use of directed ion beam sputtering). I suspect this step of the process may not be performed for the current CCDs without V-Cache.

What is the Bosch Process (Deep Reactive Ion Etching)?

www.samcointl.com

PS. Introductory paper on TSV:

http://www.ziti.uni-heidelberg.de/z...eminar/2012-Solience_Ngansso-3D_TSV-Paper.pdf

Yeah, it is more like shaving or thinning of silicon, which reveals the TSVs buried in the silicon.

 

[Vattila]

I was accused of [...blah blah blah..., all] for using some analytical skills

Nothing of the sorts, Joe! This emotionally charged blah, blah, blah was delivered in response to perceived ridicule. Ideally, you should have recognised that and defused the discussion with assurance or apology. Knowing you from another forum, I am pretty sure personal ridicule wasn't your objective — and that you were just enthusiastically arguing your view. But clarifying the subtext often helps.

 

[Joe NYC]

Why does TSMC have any incentive to put "the highest priority" on this? They don't get care whether AMD beats Intel or not, they manufacture their customers products as spec'ed in the order they've committed to. They have zero reason to invest extra resources to bring something for one customer in ahead of schedule.

This is a perfect example of bad analysis from start to finish.

 

[jpiniero]

This is a perfect example of bad analysis from start to finish.

The mistake you are making is assuming that Zen 3D is that product and not Milan-X.

 

[Joe NYC]

I somewhat disapprove of the wording choices you two made there. It's not about "highest priority", "extra resources", "ahead of schedule" etc. It's about introducing new tech, and usually TSMC, producing no chips of its own, is working together with some early adopter customers before the tech reaches general availability. All indications so far points to AMD being one if not the early adopter for what they dubbed X3D before. There is no "getting ahead of schedule" going on here, it's naturally happens and has to happen before the general availability of the tech. It's a step to getting there to begin with.

Suppose you have a technology on the drawing board that can:
- double or more demand for your product
- leaps a generation ahead in technology that has been mired in expensive trench warfare
- sends your most dangerous competitor almost a generation back in the market place

Would you just wait for this to happen when it happens, or push it to top of the priorities, expand resources to expedite it from lab to full production?

 

[Joe NYC]

The mistake you are making is assuming that Zen 3D is that product and not Milan-X.

Do you mean Ryzen vs. Milan-X?

In theory, it is the same chiplet we are discussing. But AMD made it clear that Ryzen will be the leading product.

There was an interview with journalists post V-Cache announcement, and Anandtech Cutress reported from it that Ryzen will be the leading product for Zen 3D.

And on Milan-X, the answer was "No Comment" rather than "No". Which, IMO. means "Later"

 

[moinmoin]

Suppose you have a technology on the drawing board that can:
- double or more demand for your product
- leaps a generation ahead in technology that has been mired in expensive trench warfare
- sends your most dangerous competitor almost a generation back in the market place

Would you just wait for this to happen when it happens, or push it to top of the priorities, expand resources to expedite it from lab to full production?

You make it seem like a rhetorical question. The tricky part though is: The time to make that particular decision is long gone, it was a year ago or longer. So all the comparisons you do there are not based on current reality but on projections of the future from back then. If you seriously ask that question now though waiting for this to happen is the only possible answer on short term as all decisions done already take their courses. The only change that can still happen at this point is a delay due to unforeseen technical difficulties (a certain competitor of TSMC and AMD is an expert in this area).

 

[Joe NYC]

You make it seem like a rhetorical question. The tricky part though is: The time to make that particular decision is long gone, it was a year ago or longer. So all the comparisons you do there are not based on current reality but on projections of the future from back then. If you seriously ask that question now though waiting for this to happen is the only possible answer on short term as all decisions done already take their courses. The only change that can still happen at this point is a delay due to unforeseen technical difficulties (a certain competitor of TSMC and AMD is an expert in this area).

I agree. And I have not limited making this technology a priority to present and future. TSMC made a decision some time ago to build / expand a large packaging facility for 3D stacking.

Their newest and biggest packaging capacity dedicated to 3D Fabric technologies is supposed to come on line in H2 2021.

So the time is already ticking for this facility to be utilized.

 

[Doug S]

Suppose you have a technology on the drawing board that can:
- double or more demand for your product
- leaps a generation ahead in technology that has been mired in expensive trench warfare
- sends your most dangerous competitor almost a generation back in the market place

Would you just wait for this to happen when it happens, or push it to top of the priorities, expand resources to expedite it from lab to full production?

Those are all reasons for AMD to want to prioritize it, not for TSMC to want to. TSMC fabricates their customers' designs as contracted, it isn't their problem whether those chips or the products they go in are successful in the marketplace.

In a gross sense TSMC needs their customers to be successful so they will continue to place orders and keep their fabs full. But any one customer (except perhaps Apple) could vanish tomorrow and it wouldn't impact TSMC's fortunes. They don't care if AMD sells CPUs that beat Intel's or not, TSMC gets paid the same regardless of how how much money AMD is able to sell those chips for.

If TSMC is able to deliver on AMD's orders ahead of schedule because TSMC's commitments were kind of a "worst case" timeline to protect them, and AMD is willing to take delivery early, great. But they aren't going to expend additional resources making that happen unless 1) it doesn't negatively impact the schedule for their commitments to any other customer and 2) AMD compensates them for making that happen.

 

[french toast]

Those are all reasons for AMD to want to prioritize it, not for TSMC to want to. TSMC fabricates their customers' designs as contracted, it isn't their problem whether those chips or the products they go in are successful in the marketplace.

In a gross sense TSMC needs their customers to be successful so they will continue to place orders and keep their fabs full. But any one customer (except perhaps Apple) could vanish tomorrow and it wouldn't impact TSMC's fortunes. They don't care if AMD sells CPUs that beat Intel's or not, TSMC gets paid the same regardless of how how much money AMD is able to sell those chips for.

If TSMC is able to deliver on AMD's orders ahead of schedule because TSMC's commitments were kind of a "worst case" timeline to protect them, and AMD is willing to take delivery early, great. But they aren't going to expend additional resources making that happen unless 1) it doesn't negatively impact the schedule for their commitments to any other customer and 2) AMD compensates them for making that happen.

I agree with part of that, it is AMD who would want to prioritise their product rather than Tsmc.
However you are forgetting that Intel also fabs their own chips for the most part, Intel is TSMCs biggest competitor and it would serve TSMC better if AMD won the chip battle if it meant more chips being bought from their fabs as opposed to Intel owning the market and funding their own fabs.
Also as Intel revenue declines and their fabs fall behind in technology Intel would need to source more of their own wafers from TSMC - as they have already started doing.

As for customers I think TSMC wants the x86 high powered market, Samsung isn't really in that market and is considered inferior in high powered chips, TSMC would want to develop tech and keep their leadership there, backing AMD and Nvidia over Intel serves their interests long term, even if for anti competition purposes and wooing Intel they wouldn't want to be seen as obstructive in any way.

 

[lobz]

Joe! Properly applied common sense beats fancy words most of the time. In the bleeding edge semiconductor industry, where planning and making timelines must easily be one of most difficult jobs, when in the past decades has any company managed to start producing and therefore release a product before they said they would in a timeframe of ~6 months? No need for any further personalization of the matter, just think about this one. Was there most probably a very good reason that Lisa Su did not use the wording '2H' even though she could have without technically lying, but decided to further specify the timing of production of the new tech to happen by the end of the year instead? There is nothing else behind my thoughts on this topic, only common sense. Moreover, AMD REALLY SHOULD have this product ready by the time ADL launches, they just sadly won't, so all the 2 pages talking about why is releasing it instantly in this or that company's best interest makes no difference. It's an interesting discussion, I'll give you that! I've read through it with great pleasure Only it has nothing to do with the topic where you thought I was not appreciating your analysis.

By your narrative, Intel would never have gone into their specific 10nm rabbit hole, EVER. They would either have launched their new and newer architectures with backporting sooner, or realized how unattainable the original 10nm goals were much sooner than they did - and of course I know that engineers realized that immediately, they just had different orders and mandates, ever so sadly... What should happen and what's in the best interest of a company in any other area than short to mid-term financials is seldom a reliable indicator when you try to make projections from the outside about a big joint-stock corporation with lots of layers and levels of management and decision-making.

To make it short: for the sake of buzzing and bubbling competition we can all - and most of us probably do - wish for new architectures and technologies from competing companies to be launched/released within a couple of weeks of each other. In this particular case it's just highly improbable. Please note my wording again: HIGHLY IMPROBABLE. We all know that nothing is impossible, all molecules could gather in a single corner of our rooms any time. Let's just hope they won't ever

 

[jpiniero]

And on Milan-X, the answer was "No Comment" rather than "No". Which, IMO. means "Later"

AMD hasn't said anything publicly about Epyc getting the stacked dies. Only "Zen 3 based processors". Hence the no comment.

 

[Mopetar]

When in the past decades has any company managed to start producing and therefore release a product before they said they would in a timeframe of ~6 months?

That almost never happens simply because products are rather complex and part of a larger manufacturing and supply chain. If chips yield much better than expected there's still a necessity for other components and work to be done and shifting anything else downstream in the schedule to hit an earlier ship date. Not all of the customers would want an early delivery either as they have plans of their own that coincide with previously set delivery dates. What good is shipping a new CPU half a year early if none of the motherboard manufacturers have been able to release a product yet?

In the cases where something goes right, a company may just add in more components, etc. for downstream activities so that there's more supply at launch. It's much easier to do that than to dramatically shift the date ahead by that much time. At best you might see a release moved around within a quarter just to try to get out ahead of a competitor or perhaps to steal some of their thunder, but given the long lead times of these products it's something that was already in the pipeline for some time.

The only companies that are both capable of doing something like this and might actually do it are ones that are doing small runs of products where it's actually possible to make large adjustments like that due to not having to deal with the logistics of hundreds of thousands or even millions of parts.

 

[lobz]

That almost never happens simply because products are rather complex and part of a larger manufacturing and supply chain. If chips yield much better than expected there's still a necessity for other components and work to be done and shifting anything else downstream in the schedule to hit an earlier ship date. Not all of the customers would want an early delivery either as they have plans of their own that coincide with previously set delivery dates. What good is shipping a new CPU half a year early if none of the motherboard manufacturers have been able to release a product yet?

In the cases where something goes right, a company may just add in more components, etc. for downstream activities so that there's more supply at launch. It's much easier to do that than to dramatically shift the date ahead by that much time. At best you might see a release moved around within a quarter just to try to get out ahead of a competitor or perhaps to steal some of their thunder, but given the long lead times of these products it's something that was already in the pipeline for some time.

The only companies that are both capable of doing something like this and might actually do it are ones that are doing small runs of products where it's actually possible to make large adjustments like that due to not having to deal with the logistics of hundreds of thousands or even millions of parts.

This pretty much sums up what I meant when I was disagreeing with the hopes of V-Cache products launching soon, regardless of how perfect the timing could be, judging from the outside.

 

[Doug S]

I agree with part of that, it is AMD who would want to prioritise their product rather than Tsmc.
However you are forgetting that Intel also fabs their own chips for the most part, Intel is TSMCs biggest competitor and it would serve TSMC better if AMD won the chip battle if it meant more chips being bought from their fabs as opposed to Intel owning the market and funding their own fabs.
Also as Intel revenue declines and their fabs fall behind in technology Intel would need to source more of their own wafers from TSMC - as they have already started doing.

As for customers I think TSMC wants the x86 high powered market, Samsung isn't really in that market and is considered inferior in high powered chips, TSMC would want to develop tech and keep their leadership there, backing AMD and Nvidia over Intel serves their interests long term, even if for anti competition purposes and wooing Intel they wouldn't want to be seen as obstructive in any way.

Intel isn't TSMC's competitor at all, not until they enter the foundry market for real. Intel is already behind TSMC, and TSMC's best move to keep them there is to keep delivering new nodes on schedule - and Apple is the one primarily funding that, not AMD. Delivering a high end niche AMD product a quarter or two ahead of schedule is irrelevant in the big picture.

The US government is unlikely to allow Intel to lose this battle. They may not win, but they will never wave a white flag. They will get billions flowing from the government both in terms of purchase guarantees from the Trusted Foundry program plus giveaways like the "CHIPS Act" to help fund construction of new US based fabs. Sure TSMC is / will be getting in on that too, but as a US headquartered company Intel will always have the advantage when it comes to pork.

 

[Thibsie]

Pork?

 

[Hitman928]

Pork?

I believe it is a term used only in the U.S.A.

Pork barrel - Wikipedia

en.wikipedia.org

 

[CHADBOGA]

Pork?

Yes, a very popular food item.

 

[Joe NYC]

Those are all reasons for AMD to want to prioritize it, not for TSMC to want to. TSMC fabricates their customers' designs as contracted, it isn't their problem whether those chips or the products they go in are successful in the marketplace.

In a gross sense TSMC needs their customers to be successful so they will continue to place orders and keep their fabs full. But any one customer (except perhaps Apple) could vanish tomorrow and it wouldn't impact TSMC's fortunes. They don't care if AMD sells CPUs that beat Intel's or not, TSMC gets paid the same regardless of how how much money AMD is able to sell those chips for.

If TSMC is able to deliver on AMD's orders ahead of schedule because TSMC's commitments were kind of a "worst case" timeline to protect them, and AMD is willing to take delivery early, great. But they aren't going to expend additional resources making that happen unless 1) it doesn't negatively impact the schedule for their commitments to any other customer and 2) AMD compensates them for making that happen.

You make it sound like Intel (to TSMC) is just another company out there, and AMD (competing head to head with Intel) is just another customer of TSMC.

Those assumptions are not correct, therefore any analysis that is supported by these assumptions is not going to be correct.

 

[Joe NYC]

I agree with part of that, it is AMD who would want to prioritise their product rather than Tsmc.
However you are forgetting that Intel also fabs their own chips for the most part, Intel is TSMCs biggest competitor and it would serve TSMC better if AMD won the chip battle if it meant more chips being bought from their fabs as opposed to Intel owning the market and funding their own fabs.
Also as Intel revenue declines and their fabs fall behind in technology Intel would need to source more of their own wafers from TSMC - as they have already started doing.

As for customers I think TSMC wants the x86 high powered market, Samsung isn't really in that market and is considered inferior in high powered chips, TSMC would want to develop tech and keep their leadership there, backing AMD and Nvidia over Intel serves their interests long term, even if for anti competition purposes and wooing Intel they wouldn't want to be seen as obstructive in any way.

Additionally, Intel is lobbying Congress to fund Intel fabs so that Intel can take on TSMC, and at the same time lobbying Congress NOT to give any support to TSMC fab in Arizona, even though it will be on US soil.

So after a period of truce, Intel (under Pat Galsinger) is back to being at war with TSMC.

 

[Joe NYC]

Joe! Properly applied common sense beats fancy words most of the time. In the bleeding edge semiconductor industry, where planning and making timelines must easily be one of most difficult jobs, when in the past decades has any company managed to start producing and therefore release a product before they said they would in a timeframe of ~6 months? No need for any further personalization of the matter, just think about this one. Was there most probably a very good reason that Lisa Su did not use the wording '2H' even though she could have without technically lying, but decided to further specify the timing of production of the new tech to happen by the end of the year instead? There is nothing else behind my thoughts on this topic, only common sense. Moreover, AMD REALLY SHOULD have this product ready by the time ADL launches, they just sadly won't, so all the 2 pages talking about why is releasing it instantly in this or that company's best interest makes no difference. It's an interesting discussion, I'll give you that! I've read through it with great pleasure Only it has nothing to do with the topic where you thought I was not appreciating your analysis.

The latest leaked road map has Zen 3D for 2021:

To make it short: for the sake of buzzing and bubbling competition we can all - and most of us probably do - wish for new architectures and technologies from competing companies to be launched/released within a couple of weeks of each other. In this particular case it's just highly improbable. Please note my wording again: HIGHLY IMPROBABLE. We all know that nothing is impossible, all molecules could gather in a single corner of our rooms any time. Let's just hope they won't ever

If the "highly improbable" does happen, then you didn't see it coming, and I did.