AMD Gives Itself Massive Cost-cutting Headroom with the Chiplet Design

[btarunr]

At its 2020 IEEE ISSCC keynote, AMD presented two slides that detail the extent of cost savings yielded by its bold decision to embrace the MCM (multi-chip module) approach to not just its enterprise and HEDT processors, but also its mainstream desktop ones. By confining only those components that tangibly benefit from cutting-edge silicon fabrication processes, namely the CPU cores, while letting other components sit on relatively inexpensive 12 nm, AMD is able to maximize its 7 nm foundry allocation, by making it produce small 8-core CCDs (CPU complex dies), which add up to AMD's target core-counts. With this approach, AMD is able to cram up to 16 cores onto its AM4 desktop socket using two chiplets, and up to 64 cores using eight chiplets on its SP3r3 and sTRX4 sockets.

In the slides below, AMD compares the cost of its current 7 nm + 12 nm MCM approach to a hypothetical monolithic die it would have had to build on 7 nm (including the I/O components). The slides suggest that the cost of a single-chiplet "Matisse" MCM (eg: Ryzen 7 3700X) is about 40% less than that of the double-chiplet "Matisse" (eg: Ryzen 9 3950X). Had AMD opted to build a monolithic 7 nm die that had 8 cores and all the I/O components of the I/O die, such a die would cost roughly 50% more than the current 1x CCD + IOD solution. On the other hand, a monolithic 7 nm die with 16 cores and I/O components would cost 125% more. AMD hence enjoys a massive headroom for cost-cutting. Prices of the flagship 3950X can be close to halved (from its current $749 MSRP), and AMD can turn up the heat on Intel's upcoming Core i9-10900K by significantly lowering price of its 12-core 3900X from its current $499 MSRP. The company will also enjoy more price-cutting headroom for its 6-core Ryzen 5 SKUs than it did with previous-generation Ryzen 5 parts based on monolithic dies.



View at TechPowerUp Main Site

 

[HwGeek]

So this shows how much more expensive the mobile Ryzen 4000 parts are compared to desktop.

 

[Flanker]

Are there any info on how graphics component of APU fit into this chiplet design?

 

[_Flare]

The halved L3$ lets the cost drop because of the area savings and the symbiotic effects putting the GPU onboard the same die makes the decision even more logical for me they went monolithic.

 

[londiste]

From the graph, each CCD (75mm^2 at 7nm) costs about 1/3 of the price which means IOD (124mm^2 at 12/14nm) also costs 1/3. Cost is same but CCD is ~65% smaller than IOD.
I wonder how this compares to manufacturing cost of Intel mainstream CPUs? For example 8-core like 9900K with die size of 175mm^2.

Are there any info on how graphics component of APU fit into this chiplet design?

It doesn't. Renoir is monolithic.

The halved L3$ lets the cost drop because of the area savings and the symbiotic effects putting the GPU onboard the same die makes the decision even more logical for me they went monolithic.

Not halved. Renoir has 1/4 L3$ of Matisse. 3700X has 32MB L3$ (and 3900/3950X have 64MB), 4800H has 8MB.
I am willing to bet this is thanks to Integrated Memory Controller.

Edit:
What strikes me as odd is how comparatively small the manufacturing cost saving is for EPYCs. I mean, half the cost is nothing to sneeze at but with everything that is said I expected this to be bigger.
On the other hand, EPYC's massive IOD probably has a large part to play here.

 

[Raendor]

So, does this mean 4000 series will be priced cheaper than current 3000 prices already or they will also discount the 3000 series further down? I was thinking about getting 3600 and moving my trusty 6700k to htpc, but if the prices will get even better, I don’t mind to hold on a bit.

 

[londiste]

Edit:
What strikes me as odd is how comparatively small the manufacturing cost saving is for EPYCs. I mean, half the cost is nothing to sneeze at but with everything that is said I expected this to be bigger.
On the other hand, EPYC's massive IOD probably has a large part to play here.

Looking at the EPYC graph CCD cost is small. Adding 6 CCDs moves cost from 0.65 to 1 - 35% of 64-core EPYC manufacturing cost is 6 CCDs which means all 8 CCDs are only about 47% of the cost. More than half is IOD (and packaging).

Edit:
By the way, IOD on EPYC is 430mm^2.
For comparison, Intel Skylake-SP HCC (18 cores) is 484mm^2.

Prices of the flagship 3950X can be close to halved (from its current $749 MSRP), and AMD can turn up the heat on Intel's upcoming Core i9-10900K by significantly lowering price of its 12-core 3900X from its current $499 MSRP. The company will also enjoy more price-cutting headroom for its 6-core Ryzen 5 SKUs than it did with previous-generation Ryzen 5 parts based on monolithic dies.

This is based on what? I get that this is an assumption but there is nothing in this story to support it. Sure, AMD can drop pices as they have a very good profit margin. So can Intel. Manufacturing costs do not play as much into pricing equation on desktop.

 

[notb]

Nothing new here. And it's all very cute, but the simple fact is: this doesn't translate into excellent gross margins.

So what AMD is really saying: we decided to base our competitive advantage on a very expensive manufacturing process and going with MCM was the only way we could afford it.

And yeah... it makes you wonder how much AMD asks for (and earns from) these 8-core mobile SoCs.

 

[londiste]

Nothing new here. And it's all very cute, but the simple fact is: this doesn't translate into excellent gross margins.

Yes it does. AMD's profit margin in Q4 2019 was up 2% from Q3 2019 and up 7% from Q4 2018.

 

[notb]

Yes it does. AMD's profit margin in Q4 2019 was up 2% from Q3 2019 and up 7% from Q4 2018.

That's net profit.
Gross margin tells you how much it costs to make a product, i.e.:
(revenue - cost of goods sold) / revenue
In case of AMD majority of "cost of goods sold" is what they pay for outsourced manufacturing. It doesn't include things like R&D.
This ratio is growing, but it's still much lower than Intel's (and Nvidia's).

In other words: despite MCM and more dense node, it still costs them more to make chips than it costs Intel and Nvidia.
And combining that and what they say about MCM saving, it suggests that they're losing money on large 7nm monolithic chips (4800U will be as big as two Zen2 chiplets).

 

[ARF]

That's net profit.
Gross margin tells you how much it costs to make a product, i.e.:
(revenue - cost of goods sold) / revenue
In case of AMD majority of "cost of goods sold" is what they pay for outsourced manufacturing. It doesn't include things like R&D.
This ratio is growing, but it's still much lower than Intel's (and Nvidia's).

In other words: despite MCM and more dense node, it still costs them more to make chips than it costs Intel and Nvidia.
And combining that and what they say about MCM saving, it suggests that they're losing money on large 7nm monolithic chips (4800U will be as big as two Zen2 chiplets).

Not exactly. Intel sells for $20,000, AMD sells for $6,500.
As long as there are stupid consumers who invest in anything Intel, it will be the same story.

From the graph, each CCD (75mm^2 at 7nm) costs about 1/3 of the price which means IOD (124mm^2 at 12/14nm) also costs 1/3. Cost is same but CCD is ~65% smaller than IOD.
I wonder how this compares to manufacturing cost of Intel mainstream CPUs? For example 8-core like 9900K with die size of 175mm^2.

Intel's N14 process is already more than 5-year-old, hence the manufacturing cost should be very low, including the economy of scale.

 

[Tsukiyomi91]

I ain't even mad at them doing this. If it means cheaper processors to buy, I'm sold.

 

[mtcn77]

Are there any info on how graphics component of APU fit into this chiplet design?

They place it in its own mcm perhaps. Though I suppose 7 nm 12 nm don't make that much of a difference.

 

[Imsochobo]

Not exactly. Intel sells for $20,000, AMD sells for $6,500.
As long as there are stupid consumers who invest in anything Intel, it will be the same story.



Intel's N14 process is already more than 5-year-old, hence the manufacturing cost should be very low, including the economy of scale.

Just to point out.

No-one have to pay full price from intel, very far from it.
No-one gets serious rebate on amd cpu's.

 

[Vya Domus]

The main advantage isn't even the cost saving, it's the fact that they can build processors that would otherwise be physically impossible to make due to hard limits in the manufacturing process. That's "priceless".

 

[londiste]

@Vya Domus none of these processors possibly with the exception of 64-core EPYC are physically impossible to build. Financially not viable... probably

 

[mtcn77]

(4800U will be as big as two Zen2 chiplets).

There is a reference in the first slide, if this is a reply to that. It says 8-core monolithic costs 90% of 16-core mcm-io discrete dies. Maybe it will take 10% more to build than the two die solution(3900 series). I wouldn't say AMD cannot pay for it.

 

[londiste]

Renoir is 150mm^2. So yes, pretty much exactly the size of 2 CCD chiplets:

AMD Ryzen Mobile 4000: Measuring Renoir’s Die Size

www.anandtech.com

 

[ARF]

The main advantage isn't even the cost saving, it's the fact that they can build processors that would otherwise be physically impossible to make due to hard limits in the manufacturing process. That's "priceless".

It depends on the size of the individual core. Navi 21 will have 5120 cores in just 505 mm^2.

 

[bug]

I sense a change of tune here. Initially AMD claimed IO doesn't scale well with the process and that's why they left it alone. Now it's cost reasons.

 

[ppn]

Yes they are pocketing the massive cost cutting headroom for themselves. So the only good time to buy 8 core like that is when it drops to $99.

 

[kapone32]

Yes they are pocketing the massive cost cutting headroom for themselves. So the only good time to buy 8 core like that is when it drops to $99.

But then how would we help AMD force price corrections for Intel and Nvidia products? That money will go into R&D (They are a tech company afterall). It will be interesting to see where AMD does in both the GPU and CPU space now that they are positive in terms of cash flow and revenue.

 

[r.h.p]

Im no brainiaK so can some one explain what is wrong with chiplets … or is monolithic a better option if it is more expensive

 

[londiste]

I sense a change of tune here. Initially AMD claimed IO doesn't scale well with the process and that's why they left it alone. Now it's cost reasons.

Both are true at the same time. IO not scaling well means it will have to stay larger which wastes die space. Die space is expensive and moreso on 7nm.

 

[kapone32]

Im no brainiaK so can some one explain what is wrong with chiplets … or is monolithic a better option if it is more expensive

Monolithic is better for gaming at 1080P. You can usually get higher clocks with that and reduced latency.