Speculation: Ryzen 4000 series/Zen 3

[french toast]

October should be quite revealing in one way or another.

Why October? Have AMD got an announcement planned?

 

[xilli_fiberbit]

There are data about the AMD patent about X3D for Genoa :

https://twitter.com/x/status/1290055998351130624

P.S. : i should put this on zen 4 thread

 

[DrMrLordX]

Why October? Have AMD got an announcement planned?

Allegedly, that is when two Vemeer parts drop (12c, 16c). We will see.

 

[raghu78]

Still quite amazed at how few leaks we have on Zen3 desktop. This is incredible NDA enforcement by AMD if there's no delay...

Yes I am surprised at the lack of any Zen 3 benchmark leaks.

 

[LightningZ71]

Just thinking out loud... Is it possible that there won't be any Vermeer 4-8C products for desktop? Cezanne seems to have a slightly shorter development cycle than Renoir, retaining VEGA to get ZEN3 in to monolithic ides sooner rather than later. What if AMD goes the route of pushing the Renoir parts, 4300/4600/4700, for desktop, and holds Vermeer to 4900/4950. Then, when Cezanne hits, it's a replacement for the low end desktop parts from 5300/5600/5700? The Zen3 CCDs don't show up in the lower desktop market at all, or at least until MUCH later in the cycle, when they have reached full production volume and have their server/HEDT/Prosumer market covered and have amortized the package further?

Looking at their competition, Intel doesn't have an appreciable performance lead with their 4C/6C and 8C desktop parts now, and won't have replacements for those products until next year in Q3 at the earliest. That section of the market commands the smallest margins and lowest ASPs. Renoir desktop has to be cheaper to make than Matisse in the 4C-8c segment. If Zen3 is all that its supposed to be, and if the "improved" 7nm process that Cezanne is going to be built with has any sort of decent improvements, then its fair to assume that Cezanne will be competitive in its market environment.

 

[Shivansps]

Renoir leaks are looking so poorly in price/performance in desktop that Intel actually has a LEAD with their 10100 and 10400 for eveything that is not IGP gaming, and AMD APU IGP gaming took a hit with the Vega cutdown and segmentation.
(this is with the info we already know, it may change)

With this in mind getting Cezanne out of the door FAST is a good idea, specially with Renoir falling short in expected IGP and CPU results. In fact, if Renoir stays as OEM only and Cezannes comes out in H1 2021 it would not be so bad.

 

[DisEnchantment]

if the "improved" 7nm process that Cezanne is going to be built with has any sort of decent improvements

Two years is a lot of time for TSMC to not improve anything for a node which is being sold as a long node. Both TSMC and Samsung don't go big bang like Intel.
They continuously add and improve scaling boosters.

 

[KompuKare]

Renoir desktop has to be cheaper to make than Matisse in the 4C-8c segment.

Not sure I understand you there?
Renoir is 156mm², and Renoir's packaging is easier (cheaper).
Matisse is 74mm² per chiplet and the IO dies for Ryzen is a fixed cost (although it is possible that they may use some otherwise defective IO dies when only connecting to one chiplet in a 4 core Ryzen).
Plugging that into a die yield calculator, leaving defects at the default of 0.2 #/sq. cm:

Die Yield Calculator - iSine

DIE YIELD CALCULATOR Use this online calculator to figure out die yield using Murphy’s model. You’ll need to know the die size, wafer diameter, and defect density. iSine is your complete resource for ASIC design – from concept to manufacturing and testing. We have expertise in system...

www.isine.com

Gets a yield of 676 per 300mm wafer with a 86.2% yield for a Matisse chiplet.
And a yield of 269 per 300mm wafer with a 73.8% yield for a Renoir die.
Ergo, Renoir is far more expensive to produce.
Even Cezanne is probably going to have worse yields than Matisse chiplets although it should be easier to package. Guess that all depends on the size of Cezanne, but cutting the CCX in half and cutting back CUs still leaves fixed size stuff like the memory controlller, chipset areas, display controllers. Probably around 100mm² which is still big for 7nm compared to mobile SOCs.

 

[maddie]

Not sure I understand you there?
Renoir is 156mm², and Renoir's packaging is easier (cheaper).
Matisse is 74mm² per chiplet and the IO dies for Ryzen is a fixed cost (although it is possible that they may use some otherwise defective IO dies when only connecting to one chiplet in a 4 core Ryzen).
Plugging that into a die yield calculator, leaving defects at the default of 0.2 #/sq. cm:

Die Yield Calculator - iSine

DIE YIELD CALCULATOR Use this online calculator to figure out die yield using Murphy’s model. You’ll need to know the die size, wafer diameter, and defect density. iSine is your complete resource for ASIC design – from concept to manufacturing and testing. We have expertise in system...

www.isine.com

Gets a yield of 676 per 300mm wafer with a 86.2% yield for a Matisse chiplet.
And a yield of 269 per 300mm wafer with a 73.8% yield for a Renoir die.
Ergo, Renoir is far more expensive to produce.
Even Cezanne is probably going to have worse yields than Matisse chiplets although it should be easier to package. Guess that all depends on the size of Cezanne, but cutting the CCX in half and cutting back CUs still leaves fixed size stuff like the memory controlller, chipset areas, display controllers. Probably around 100mm² which is still big for 7nm compared to mobile SOCs.

IO is free? The more complex packaging with Matisse is free?

My view is that 125mm² 12nm IO + 74mm² 7nm + complex package routing for Matisse is comparable to 156mm² 7nm with simpler package.

If you use the die calculator, you might not know this, but the math equations underpinning it, give differences in yield for smaller versus larger die that are exaggerated as defect density is increased. In reality the defect density is under .1, closer to 0.07-0.08 (TSMC numbers).

 

[Shivansps]

Im really tired of this "7nm is expensive" talk... AMD is able to provide a 7nm 8C Zen2 APU with 36 CU RDNA2 in order to make next gen consoles viable.

But somehow, on desktop, 7nm means every price has to go up. Im sorry but i dont belive that for one second.

 

[jpiniero]

AMD is able to provide a 7nm 8C Zen2 APU with 36 CU RDNA2 in order to make next gen consoles viable.

52 CUs you mean.

It's just the kind of thing where the overlap of people on desktop who want high perf gaming but won't buy a card isn't much.

 

[Kenmitch]

Im really tired of this "7nm is expensive" talk... AMD is able to provide a 7nm 8C Zen2 APU with 36 CU RDNA2 in order to make next gen consoles viable.

But somehow, on desktop, 7nm means every price has to go up. Im sorry but i dont belive that for one second.

How much does AMD charge Sony and Microsoft for the APU's?

 

[Hitman928]

Im really tired of this "7nm is expensive" talk... AMD is able to provide a 7nm 8C Zen2 APU with 36 CU RDNA2 in order to make next gen consoles viable.

But somehow, on desktop, 7nm means every price has to go up. Im sorry but i dont belive that for one second.

If you are willing to pay AMD millions on the front end to support APU R&D and also promise to buy tens of millions of chips on the back end, I'm sure you could get each chip for a very good price as well.

 

[Kenmitch]

If you are willing to pay AMD millions on the front end to support APU R&D and also promise to buy tens of millions of chips on the back end, I'm sure you could get each chip for a very good price as well.

I'll buy one from him as long as he only charges cost. Would be silly to complain about pricing and expect to make a profit on me.

 

[Shivansps]

If you are willing to pay AMD millions on the front end to support APU R&D and also promise to buy tens of millions of chips on the back end, I'm sure you could get each chip for a very good price as well.

That also means wafer demand go way up. There is a lot of components in a product final price, production cost is just one part.

The "order 500 millon chips" helps to reduce R&D costs over a larger quantity of chips, what is important in a custom chip, yes, but you still need to supply enoght wafers for that. And as bigger the order you need to get more wafers and supply is limited. SO, NO.

On this forum people are arguing me that asking more perf at the same price in 7nm is crazy... THIS NEEDS TO STOP RIGHT NOW.

 

[Hitman928]

That also means wafer demand go way up. There is a lot of components in a product final price, production cost is just one part.

The "order 500 millon chips" helps to reduce R&D costs over a larger quantity of chips, what is important in a custom chip, yes, but you still need to supply enoght wafers for that. And as bigger the order you need to get more wafers and supply is limited. SO, NO.

On this forum people are arguing me that asking more perf at the same price in 7nm is crazy... THIS NEEDS TO STOP RIGHT NOW.

You need to stop asking for more performance at the same for 7 nm chips RIGHT NOW.

Did that work? Is that how this forum works now?

Again, you missed the part about Sony and Microsoft funding R&D and not just buying chips, but guaranteeing tens of millions of sales. It's a pretty sweet setup for AMD. They can't have it be their only business, but it's a great component for them to have and is really the only thing that kept them alive during the Bulldozer era.

I understand that desktop Renoir doesn't meet your needs as well as you'd like and it will probably impact your business. That sucks. But don't expect AMD or the rest of the market to bend to your will unless you come forward with millions of orders to be fulfilled.

 

[maddie]

If you are willing to pay AMD millions on the front end to support APU R&D and also promise to buy tens of millions of chips on the back end, I'm sure you could get each chip for a very good price as well.

So if the R&D has been heavily subsidized by the console makers, should we then expect even lower prices than would normally be the case? Have cake = no eat it.

 

[Shivansps]

You need to stop asking for more performance at the same for 7 nm chips RIGHT NOW.

There is nothing wrong with my argument. You cant keep using the same argument over and over that "just because" 7nm is more expensive it is justifed to have higher prices. PRICING DOES NOT WORK THAT WAY. Go and learn how prices are formed, production costs is just one element.

Is even worse when you try to argue that ordering "a lot of chips" should result in a lower price when there is something called "supply and demand" and 7nm demand is very high at this point. You need to see the whole picture.

 

[LightningZ71]

I don't think that people understood my point. I believe that Renoir should be cheaper than Matisse at the 8, 6 and 4 core segments because:
The individual dies have to be assembled onto an MCP. At least one set of those dies must be shipped a good distance to get there.
The package assembly process is not 100% perfect. There are errors that can, and do, occur there, and processors are lost.
Matisse requires more time and energy, per completed processor package, than Renoir, due to the extra assembly steps in its production.

All of that adds cost to Matisse. Cost that Renoir either does not have, or has less of.

I will grant you that the individual 8 core CCDs that Matisse uses cost less to produce than each larger Renoir die. I grant you that the smaller 8 core Matisse CCD has a higher yield as a percentage of dies per wafer than Renoir. This is a fact of life in silicon production. However, TSMC 7nm production is now HIGHLY mature. The defect rate is low. In volume, the actual cost per die is not excessive.

There is also another consideration: it is highly likely that Renoir desktop is a product of binning. What I mean is that the individual dies used for Desktop are likely dies that, for likely power/performance reasons, did not make the cut for mobile. I have no proof of that, but with the shortage of top end mobile APUs being reported in other threads, it is reasonable to conclude that there is possibly a power/performance goal yield issue at play.

So, I contend that, for AMD, aside from continuing to harvest CCDs that don't meet spec for use in Epyc machines, they will likely shift focus on low end desktop to Renoir dies going forward.

 

[Kenmitch]

So if the R&D has been heavily subsidized by the console makers, should we then expect even lower prices than would normally be the case? Have cake = no eat it.

I prefer pie.

 

[Shivansps]

I don't think that people understood my point. I believe that Renoir should be cheaper than Matisse at the 8, 6 and 4 core segments because:
The individual dies have to be assembled onto an MCP. At least one set of those dies must be shipped a good distance to get there.
The package assembly process is not 100% perfect. There are errors that can, and do, occur there, and processors are lost.
Matisse requires more time and energy, per completed processor package, than Renoir, due to the extra assembly steps in its production.

All of that adds cost to Matisse. Cost that Renoir either does not have, or has less of.

I will grant you that the individual 8 core CCDs that Matisse uses cost less to produce than each larger Renoir die. I grant you that the smaller 8 core Matisse CCD has a higher yield as a percentage of dies per wafer than Renoir. This is a fact of life in silicon production. However, TSMC 7nm production is now HIGHLY mature. The defect rate is low. In volume, the actual cost per die is not excessive.

There is also another consideration: it is highly likely that Renoir desktop is a product of binning. What I mean is that the individual dies used for Desktop are likely dies that, for likely power/performance reasons, did not make the cut for mobile. I have no proof of that, but with the shortage of top end mobile APUs being reported in other threads, it is reasonable to conclude that there is possibly a power/performance goal yield issue at play.

So, I contend that, for AMD, aside from continuing to harvest CCDs that don't meet spec for use in Epyc machines, they will likely shift focus on low end desktop to Renoir dies going forward.

The only problem is that you are only taking production cost in consideration, it is a common misconception. Renoir being easier or harder to produce than Matisse is just one of many elements to consider.

7nm supply demand being way too high due to console chips is likely to have a bigger impact, but it is just another element.

This whole idea of trying to get an estimated "price per mm2" on 7nm to determine how much it should cost vs the same mm2 on 12nm is wrong, yes it affects costs, but a product price is not determined by cost alone.

 

[LightningZ71]

There is nothing wrong with my argument. You cant keep using the same argument over and over that "just because" 7nm is more expensive it is justifed to have higher prices. PRICING DOES NOT WORK THAT WAY. Go and learn how prices are formed, production costs is just one element.

Is even worse when you try to argue that ordering "a lot of chips" should result in a lower price when there is something called "supply and demand" and 7nm demand is very high at this point. You need to see the whole picture.

You are partially right, AMD certainly doesn't price these things in a vacuum. AMD prices them with what they believe that the market FOR THEIR CHIPS will bear. Their 4300 and 4600 APUS are bellpark competitive with Intel processors that are priced similarly. YEs, they loose a little on the CPU side and win a little on the gpu side. Its called a trade off. Where they loose on the CPU side is not a big loss anywhere save for AVX2 loads and one or two other Intel favoring tests. Where they win in GPU loads, they win significantly.

For your continuing ax grinding vs. the 3400g, I remind you, the 4300 walks all over it in most every CPU bound test. Where it looses is in GPU items, and, those losses are rarely enough to make a noticeable effect in game playability in all but a very few cases. If you are THAT focused on gaming performance, why are you buying an APU? If you're stuck in a uSFF ITX case and can't use a GPU, why aren't you getting the 4600 or 4700? If you are THAT sensitive in price, the 3400g is still out there.

You are beating a drum for AMD to offer to you a CPU that has better CPU performance and better GPU performance for the same or less money than an existing product that they are still selling! You want them to deliberately land mine the ASP of their existing product. HOW IS THAT A SMART BUSINESS DECISION?!?! In what world should a company intentionally make a current product less appealing to their customers? Go and learn how companies make money, ASP and their ability to get those prices is also an important element!!!

 

[Hitman928]

There is nothing wrong with my argument. You cant keep using the same argument over and over that "just because" 7nm is more expensive it is justifed to have higher prices. PRICING DOES NOT WORK THAT WAY. Go and learn how prices are formed, production costs is just one element.

Is even worse when you try to argue that ordering "a lot of chips" should result in a lower price when there is something called "supply and demand" and 7nm demand is very high at this point. You need to see the whole picture.

WRITING IN ALL CAPS DOESN'T MAKE YOUR POINT ANY MORE VALID.

Your view of supply and demand and pricing is overly simplistic and completely ignores the business relationship AMD has with Sony, Microsoft, and pretty much every major OEM. Sony and Microsoft pay a small portion of what their chips would cost at retail or even to retail suppliers, that is a fact. OEMs pay significantly less than what a retailer would pay for their supply. That's the way things work and have worked for decades across multiple business segments. Even as a small business ordering COTS components, you get price breaks for large orders and that's just in the hundreds to thousands of units range, let alone millions or tens of millions agreed to by these giant corporations.

If you really think that AMD is screwing the pooch here, then contact your AMD sales rep and show them the numbers of how many orders and at what value they are missing out on. If your orders are at all significant enough in the marketplace, they'll listen. I'm guessing though that they aren't which leaves you at the mercy of the design decisions AMD is making for their large customers who (given the 7 nm shortage) would probably gobble up your yearly sales in a week. You don't have to like it, but you also don't have the right to tell people how things should be and keep posting trying to shut people up who criticize your opinion.

 

[maddie]

The only problem is that you are only taking production cost in consideration, it is a common misconception. Renoir being easier or harder to produce than Matisse is just one of many elements to consider.

7nm supply demand being way too high due to console chips is likely to have a bigger impact, but it is just another element.

This whole idea of trying to get an estimated "price per mm2" on 7nm to determine how much it should cost vs the same mm2 on 12nm is wrong, yes it affects costs, but a product price is not determined by cost alone.

He was backing your argument to some degree.

 

[Shivansps]

You are partially right, AMD certainly doesn't price these things in a vacuum. AMD prices them with what they believe that the market FOR THEIR CHIPS will bear. Their 4300 and 4600 APUS are bellpark competitive with Intel processors that are priced similarly. YEs, they loose a little on the CPU side and win a little on the gpu side. Its called a trade off. Where they loose on the CPU side is not a big loss anywhere save for AVX2 loads and one or two other Intel favoring tests. Where they win in GPU loads, they win significantly.

Lets drop the rest shall we?

AMD prices these things to what they belive the market will pay for. We agree there.

But remember that AMD can ask $1000 for the 4300G if they want, its up to the market to validate that price.

So AMD comes forward and prices the 7nm GPUs similar to Nvidia ones, and the market validated that. People here justifing that "it is because 7nm is more expensive". No, it is because the Nvidia options at the same performance are even more expensive. Thats the reason, it has nothing to do with 7nm being more expensive.

What is happening with Renoir? Price increase again? ok whats the other option if you want IGP that can run more than notepad.exe? There is none, other than keep buying Picassos. I have no doubt that the market will validate those prices again, and again, they have nothing to do with 7nm being more expensive.

There is a really long and well documented history of how a price is formed, not sure why people belives 7nm is somehow, a exception, they are not making the chips out of pure gold. So you can see why whole thing of trying to figure out how much a die cost is pointless to me.

Renoir, if those OEM prices are confirmed, is going to be in a really bad place, because the 10100 is cheaper than the 4300G, so does the 10400 vs the 4600G, not sure about the 10700 vs 4700G. So is not only about "ok is better on GPU (the part that got cutdown), but worse on CPU". But again if you need a good IGP there is no choice.