[News] AMD Defends MCM Approach for Building 32-core EPYC: Huge Cost Savings of ~41%

[StyM]

https://www.techpowerup.com/236430/a...-savings-of-41

AMD, presenting at a HotChips talk, packed with VLSI gurus, defended its decision to make AMD EPYC enterprise processors in the socket SP3r2 package a multi-chip module (MCM) of four 8-core "Summit Ridge" dies, rather than building a monolithic 32-core die. For starters, it stressed on the benefits of a single kind of silicon, which it can use across its performance-desktop, high-end desktop, and enterprise product-stacks; which translates into higher yields. "Summit Ridge" is the only CPU silicon based on the "Zen" micro-architecture, with the company giving finishing touches to the second silicon, codenamed "Raven Ridge," which will power mobile and desktop APUs. AMD has to pick the best-performing dies out of a common bin. The top 5% dies go into powering the company's Ryzen Threadripper HEDT processors, and a higher percentile go into making EPYC.

The relatively smaller 8-core common die has an inherently higher yield than a larger chip due to the rule of inverse-exponential reduction in yield with increasing die-size. This, coupled with the R&D costs that would have gone into developing the hypothetical monolithic 32-core "Zen" based silicon, works out to a significant cost saving for the company. A 4-die/32-core MCM is 0.59X the cost of a hypothetical monolithic 32-core die, according to the company, which is a cost-saving that enables the company to aggressively price its products. The slide AMD used in its presentation also confirms that each 8-core "Summit Ridge" die features four external Infinity Fabric links, besides the one that connects the two CCX units with each other. On a 4-die EPYC MCM, three out of four of those external IF links wire out to the neighboring dies, and one link per die probably wires out to a neighboring socket on 2P machines.

[madmaxx]

AMD just can't win, always on the defensive. I applaud them for their efficiency & tennacity, and look to reward them with a future CPU purchase. People love to talk smack about AMD but I'd like to see those same persons take on BOTH Intel & nVidia, then come and talk.

[AliG]

I really don't see the issue. As long as people accept the increased L3 latency, there's really no disadvantage to the end customer.

Think of it this way, most server applications are so large that almost everything has to be addressed in RAM. Cache efficiency helps, but won't make that much of a difference over a 3 week long simulation.

[kromosto]

I really don't see the issue. As long as people accept the increased L3 latency, there's really no disadvantage to the end customer.

Think of it this way, most server applications are so large that almost everything has to be addressed in RAM. Cache efficiency helps, but won't make that much of a difference over a 3 week long simulation.

Increased latency is the only reason I dont buy this cpus. And this is not only for 8 core dies. 8 core dies are actually 4 core ccxs which also have latency in themselfs. Also I dont lile gambling. A game that uses 4 threads will run best when all threads are opened on same ccx and worst when 4 different ccxs. And of course different perfformance profiles for other thread ccx combinations. I am not saying this is a bad product but also not the best. For professionals it will depend on what they are using for standart home users they dont actually need more than 4 core. Gamers 6 maybe. For people like me using their system for half gaming half professional it becomes a choice that is made with feelings. I am using vms for development and testing where core and ram is important but 8 or 10 core is enough for me. greater will decrease price performance dramatically. On this range amd intel price gap is so small that I will choose an intel over amd which also makes sense for gaming. But if ccx thing is not there I believe I will prefer a 12 core amd over 8 core intel for sure.

[AliG]

Increased latency is the only reason I dont buy this cpus. And this is not only for 8 core dies. 8 core dies are actually 4 core ccxs which also have latency in themselfs. Also I dont lile gambling. A game that uses 4 threads will run best when all threads are opened on same ccx and worst when 4 different ccxs. And of course different perfformance profiles for other thread ccx combinations. I am not saying this is a bad product but also not the best. For professionals it will depend on what they are using for standart home users they dont actually need more than 4 core. Gamers 6 maybe. For people like me using their system for half gaming half professional it becomes a choice that is made with feelings. I am using vms for development and testing where core and ram is important but 8 or 10 core is enough for me. greater will decrease price performance dramatically. On this range amd intel price gap is so small that I will choose an intel over amd which also makes sense for gaming. But if ccx thing is not there I believe I will prefer a 12 core amd over 8 core intel for sure.

Lol buddy they aren't trying to sell Epyc to gamers. It's meant for corporate use - which as I pointed out above, is less reliant on cache latency.

If you want a gaming CPU just buy a 7700k. There's literally nothing faster on the market.

[kromosto]

Lol buddy they aren't trying to sell Epyc to gamers. It's meant for corporate use - which as I pointed out above, is less reliant on cache latency.

If you want a gaming CPU just buy a 7700k. There's literally nothing faster on the market.

What about my vms

[zanzabar]

I really don't see the issue. As long as people accept the increased L3 latency, there's really no disadvantage to the end customer.

Think of it this way, most server applications are so large that almost everything has to be addressed in RAM. Cache efficiency helps, but won't make that much of a difference over a 3 week long simulation.

you also get UMA memory mode so things that need to talk to memory across dies are much more efficient than going across sockets.

What about my vms

epyc does not support mounting hardware directly to a vm at launch, and you can get 2011 really cheap ATM. for multi socket epyc has no real downsides compared to 2p 2066/2011 and has huge advantages over 3647.

[kromosto]

you also get UMA memory mode so things that need to talk to memory across dies are much more efficient than going across sockets.



epyc does not support mounting hardware directly to a vm at launch, and you can get 2011 really cheap ATM. for multi socket epyc has no real downsides compared to 2p 2066/2011 and has huge advantages over 3647.

If am not wrong uma does not make talking across dies more efficient but pushes up performance by preventing cross die talking as much as possible. But if am not wrong uma wont work between ccx.

2011 is still very expensive here. I didnt know about the hardware mounting but this also narrows the corporate use you told. What I am saying is this is a choice and I dont say it is wrong actually cost reduction is awesome but what I am trying to say is result will bewordt than AMD expected.

[zanzabar]

If am not wrong uma does not make talking across dies more efficient but pushes up performance by preventing cross die talking as much as possible. But if am not wrong uma wont work between ccx.

2011 is still very expensive here. I didnt know about the hardware mounting but this also narrows the corporate use you told. What I am saying is this is a choice and I dont say it is wrong actually cost reduction is awesome but what I am trying to say is result will bewordt than AMD expected.

UMA mode allows full cross talk between cores and is much more efficient for neighboring cores than going multi socket would be. NUMA mode tries to stop cross die talk like it would be a classic 4p system per socket. TR just gets unified memory out of UMA for the most part, but epyc gets a system that would be like 2p 16 core box but has much less latency and unified memory for those 32 cores.

with desktop in the US you can get 6 cores 40 lanes and a nice x99 ddr4 board for under $500 new. it is about $50-75 more than a 7700k set up. if you needed vms on a consumer platform that would be the way to go ATM.