AMD Eypc design in detail

[NumLock21]

At Hot Chips event, AMD has revealed the reason why they went with a MCM (Multi-Chip Module) vs the traditional monolithic approach, when designing their Eypc processors. On a monolithic design of Eypc, AMD did a estimate, that it would have a die area of777mm2 vs their MCM design with a die area coming in at 852mm2. Now a smaller die size should result in lower cost, so why did AMD went with a MCM instead of the traditional monolithic design? It turns out, while a monolithic design with its smaller die size, should have lower cost, it also has a high failure rate during manufacturing. So even though, MCM with its larger dies size cost more, it actually helped AMD reduce their production cost by 41%. AMD Eypc consist of 4 MCM, communicating with each other through AMD's Infinity Fabric. It's also able to provide up to 128 PCIe gen 3 lanes so users can installed as much devices as they want without ever need to worry about running out of PCIe lanes. Fail on Intel, for making fun of AMD, for using this "glue together" design. Cause AMD clearly shows that,

glue is good!

glue works!

glue helped AMD save lots and lots of money!

 

 

http://www.tomshardware.com/news/amd-threadripper-epyc-mcm-cost,35306.html

 

 

[TacoSenpai]

Pretty cool stuff.  Maybe just maybe someday we will figure out how the infinity stones work.

[NumLock21]

Just now, LordTaco42 said:

Pretty cool stuff.  Maybe just maybe someday we will figure out how the infinity stones work.

It travels over 9000x the speed of light.

[TacoSenpai]

Just now, NumLock21 said:

It travels over 9000x the speed of light.

Amazing...the possibilities are endless

[LienusLateTips]

Doesn't everyone already know that larger dies = more failure rate = more expensive processor?

[Jurrunio]

I hope AMD can earn enough on CPUs to support development on GPUs. Even Vega is just their old design built with new manufacturing process.

[Marinatall_Ironside]

7 minutes ago, JDE said:

Doesn't everyone already know that larger dies = more failure rate = more expensive processor?

Probably. If you ask me, I would say that I did kind of figured that out that a monolithic die, although has zero bandwidth issues, it is not very cost-effective when you factor in the higher failure rate of dies due to faulty transistor blocks.

 

I didn't know that a monolithic die would run you less, but again, higher failure rate = more expensive device to make up the cost because those dies would have to go to the dump (or have those bad blocks fused off and bin it as a lower-end part). It becomes even more expensive when you begin to test for performance of each block. If they're not up to snuff, then their either binned as a lower-end part with lower clocks, or just thrown away outright.

[Flavio hc 16]

2 minutes ago, JurunceNK said:

Probably. If you ask me, I would say that I did kind of figured that out that a monolithic die, although has zero bandwidth issues, it is not very cost-effective when you factor in the higher failure rate of dies due to faulty transistor blocks.

 

I didn't know that a monolithic die would run you less, but again, higher failure rate = more expensive device to make up the cost because those dies would have to go to the dump (or have those bad blocks fused off and bin it as a lower-end part). It becomes even more expensive when you begin to test for performance of each block. If they're not up to snuff, then their either binned as a lower-end part with lower clocks, or just thrown away outright.

if you consider that to make a single monolithic die for the v100 (850mm), Nvidia says that the can carve only 1 functioning die from an entire wafer on average 

[Marinatall_Ironside]

2 minutes ago, Flavio hc 16 said:

if you consider that to make a single monolithic die for the v100 (850mm), Nvidia says that the can carve only 1 functioning die from an entire wafer on average 

Maybe two or three out of what, maybe 15 possible working Tesla V100 dies from a 12-inch diameter silicon wafer? Still, the point of your post is that with a larger die size, yields go down real fast.

Edited August 25, 2017 by JurunceNK
Changed silicon wafer diameter.

[Trixanity]

59 minutes ago, Jurrunio said:

Even Vega is just their old design built with new manufacturing process.

Yeah, that's not true.

[Jurrunio]

7 minutes ago, Trixanity said:

Yeah, that's not true.

Compared to how they completely scrapped the old design on FX chips and created Ryzen, it is true. Vega's performance per watt is just too much behind Pascal competitors. Vega 56 uses 40% more power than 1070, Vega 64 uses 63% (295W TDP) / 44% (260W TDP) more power than 1080 while Vega 64 Liquid uses nearly twice as much power as a 1080 but only gives extra 10% performance (already over estimated this). I know AMD put a lot of effort on Vega, but Nvidia is always in a ready-to-fight state. Unlike Intel, which sleeps under the banana tree and count money all day.

[sazrocks]

GLUE YES!

[WereCat]

Finally, I can build a PC with 8 core CPU that supports 8 memory channels, has 8 graphics cards that feature 8GB of VRAM, 8 NVME SSDs, and 8 normal HDDs.

 

Simply epic.

[LienusLateTips]

2 minutes ago, WereCat said:

Finally, I can build a PC with 8 core CPU that supports 8 memory channels, has 8 graphics cards that feature 8GB of VRAM, 8 NVME SSDs, and 8 normal HDDs.

 

Simply epic.

Don't forget the 800W PSU, the fact that all the 8 normal HDDs are 8TB, all the RAM sticks are 8GB, and all the NVME SSDs are 64GB, which is 8x8.

 

don't forget about the 800W PSU, the be Quiet! Silent Base 800 case

[Drak3]

5 minutes ago, JDE said:

Don't forget the 800W PSU, the fact that all the 8 normal HDDs are 8TB, all the RAM sticks are 8GB, and all the NVME SSDs are 64GB, which is 8x8.

 

don't forget about the 800W PSU, the be Quiet! Silent Base 800 case

You've forgotten about the other six 800W PSUs

[LienusLateTips]

Just now, Drak3 said:

You've forgotten about the other six 800W PSUs

oh yeah

[GDRRiley]

1 hour ago, Jurrunio said:

Compared to how they completely scrapped the old design on FX chips and created Ryzen, it is true. Vega's performance per watt is just too much behind Pascal competitors. Vega 56 uses 40% more power than 1070, Vega 64 uses 63% (295W TDP) / 44% (260W TDP) more power than 1080 while Vega 64 Liquid uses nearly twice as much power as a 1080 but only gives extra 10% performance (already over estimated this). I know AMD put a lot of effort on Vega, but Nvidia is always in a ready-to-fight state. Unlike Intel, which sleeps under the banana tree and count money all day.

not when vega is well underclocked. AMD runs to much power thrugh the chips and cutting it by a lot brings the power usage closer. AMD is just pushing the architect out of its most efficient area 

[Jurrunio]

12 minutes ago, GDRRiley said:

not when vega is well underclocked. AMD runs to much power thrugh the chips and cutting it by a lot brings the power usage closer. AMD is just pushing the architect out of its most efficient area 

If it's well underclocked then it will cut into 580's market. I do agree on AMD making a nerfed Vega to replace the 580 so there wont be power efficiency advantage on the 1060. The thing is, they want to grab money from the high end market, so they have to push Vega hard like this.

[GDRRiley]

11 minutes ago, Jurrunio said:

If it's well underclocked then it will cut into 580's market. I do agree on AMD making a nerfed Vega to replace the 580 so there wont be power efficiency advantage on the 1060. The thing is, they want to grab money from the high end market, so they have to push Vega hard like this.

I mean maybe 5% underclock and cutting the power by a bit. 

[Jurrunio]

8 minutes ago, GDRRiley said:

I mean maybe 5% underclock and cutting the power by a bit. 

5% underclock and it will lose its advantage on performance to the 1070. Can that result in 150W power draw? As for price... I dont know how much it will cost after prices get stable, so cant talk about that.

[Morgan MLGman]

In the server environment, scalability is the key and AMD has a better architecture for that part of the market.

[Flavio hc 16]

Guys have you  notice 1 thing? That number 41% cost reduction over the monolithic  die, if we reverse that 59:100=100:x  and do the proportion  100×100÷59= 169.6. It means, in other world, that building a monolithic  die that big would cost almost 70% more respect to the MCM design.

Doesnt it make it ring a bell?

Threadripper 16 core 1950x cost 1000$

I9 7960x 16 core cost : 1700$

Even though  we know that intel is money-grabbing, that's  interesting 

( and i also know that i'm talking about HEDT and not servers chip)

[Drak3]

18 minutes ago, Flavio hc 16 said:

Guys have you  notice 1 thing? That number 41% cost reduction over the monolithic  die, if we reverse that 59:100=100:x  and do the proportion  100×100÷59= 169.6. It means, in other world, that building a monolithic  die that big would cost almost 70% more respect to the MCM design.

Doesnt it make it ring a bell?

Threadripper 16 core 1950x cost 1000$

I9 7960x 16 core cost : 1700$

Even though  we know that intel is money-grabbing, that's  interesting 

( and i also know that i'm talking about HEDT and not servers chip)

Except that with monolithic dies, cost per working die increases exponentially. Your math doesn't account for that. We're also ignoring uncore RnD, particularly the IMC, where Intel is worlds better. Whereas the numbers you based your math on are theoretical 32c die EPYC vs 4 8c die EPYC.

[DXMember]

4 hours ago, Flavio hc 16 said:

Guys have you  notice 1 thing? That number 41% cost reduction over the monolithic  die, if we reverse that 59:100=100:x  and do the proportion  100×100÷59= 169.6. It means, in other world, that building a monolithic  die that big would cost almost 70% more respect to the MCM design.

Doesnt it make it ring a bell?

Threadripper 16 core 1950x cost 1000$

I9 7960x 16 core cost : 1700$

Even though  we know that intel is money-grabbing, that's  interesting 

( and i also know that i'm talking about HEDT and not servers chip)

retail price is not cost of manufacture....

[leadeater]

9 hours ago, Flavio hc 16 said:

Guys have you  notice 1 thing? That number 41% cost reduction over the monolithic  die, if we reverse that 59:100=100:x  and do the proportion  100×100÷59= 169.6. It means, in other world, that building a monolithic  die that big would cost almost 70% more respect to the MCM design.

Doesnt it make it ring a bell?

Threadripper 16 core 1950x cost 1000$

I9 7960x 16 core cost : 1700$

Even though  we know that intel is money-grabbing, that's  interesting 

( and i also know that i'm talking about HEDT and not servers chip)

 

5 hours ago, DXMember said:

retail price is not cost of manufacture....

This is also about a 4 active die product not a 2 active die product as well so the math is basically wrong due to that anyway. The cost saving would actually go down for a 2 die design as well since there is less difference between a monolithic die due to less IF interconnects etc, the production cost of a 2 active die product will be lower of course.

 

AMD is also comparing the cost from the same silicon fabrication supplier using the same fabrication process, we don't know if Intel's cost per mm2 is more or less than GlobalFoundries.

 

Threadripper effective size is 378mm2 and Skylake-X 18 core is 484mm2.