Techpowerup：Intel Says AMD EPYC Processors "Glued-together" in Official Slide Deck

[PliotronX]

Now that everyone's done laughing at the glue slides, we finally got some numbers for Epyc.

https://www.servethehome.com/amd-epyc-infinity-fabric-latency-ddr4-2400-v-2666-a-snapshot/


But these are not the numbers that will usher in a new era of general purpose MCM CPUs. For comparison, ~140ns is the remote memory latency of a 112-core 4P Skylake SP system.

AMD will need to increase the core/node or improve the IF latency dramatically in the next Zen generations to break out of the niche market they've crammed themselves into.



Nah, Intel NDA/vendor only slides always look like crap. You'd think with all the billions in revenue, there would be a professional PowerPoint guy whose only job is to make all of them look nice, but no.

If intrasocket latency is important for a role, restricting vCPU assignment to respective NUMA nodes is best practice anyway, even with Intel hardware.

 

[krumme]

Facebook, google and amazon is niche markets.
Always knew this cloud was overrated hot air.
What we want is databases that fit into a L3.

Best regards
The Flintstones

 

[sushukka]

Facebook, google and amazon is niche markets.
Always knew this cloud was overrated hot air.
What we want is databases that fit into a L3.

Best regards
The Flintstones

Like (L3+L4) on https://en.wikipedia.org/wiki/IBM_z14_(microprocessor)

 

[quikah]

Glue is an industry term that means there is not a direct connection between the cores. https://en.wikipedia.org/wiki/Glue_logic

I am not sure how Epyc works, so I have no idea if saying it is Glued is accurate.

 

[moinmoin]

Glue is an industry term that means there is not a direct connection between the cores. https://en.wikipedia.org/wiki/Glue_logic

I am not sure how Epyc works, so I have no idea if saying it is Glued is accurate.

Infinity Fabric is certainly AMD's glue logic for everything Zen so far, but I'm not sure it's commonly used as a verb with that technical meaning.

Anyway good (unofficial) fun:

 

[formulav8]

Wrong thread

 

[quikah]

Infinity Fabric is certainly AMD's glue logic for everything Zen so far, but I'm not sure it's commonly used as a verb with that technical meaning.

It absolutely is commonly used. I just attended a presentation by Dell about their R940 servers, one of the slides said it was a glueless solution. Previous presentations about other 4 socket servers I have attended will mention if they are glued or glueless solutions. It is very common in the 4+ socket server space.

 

[moinmoin]

It absolutely is commonly used. I just attended a presentation by Dell about their R940 servers, one of the slides said it was a glueless solution. Previous presentations about other 4 socket servers I have attended will mention if they are glued or glueless solutions. It is very common in the 4+ socket server space.

Out of curiosity, what is a glueless 4 socket server? Sounds unusable to me (4 disconnected sockets?). And "glueless solutions" sounds like a bad pun. ^^

 

[quikah]

Out of curiosity, what is a glueless 4 socket server? Sounds unusable to me (4 disconnected sockets?). And "glueless solutions" sounds like a bad pun. ^^

For intel solution, they connect the CPUs with the QPI links directly. A glued solution has an external controller that link the sockets, typically there will be 2 or 4 glueless sockets that will be glued together.

Here is a random explanation I found on google:
https://deinoscloud.wordpress.com/2012/10/19/two-main-scale-up-server-architectures-part-1/
https://deinoscloud.wordpress.com/2012/10/22/two-main-scale-up-server-architectures-part-2/

 

[moinmoin]

For intel solution, they connect the CPUs with the QPI links directly. A glued solution has an external controller that link the sockets, typically there will be 2 or 4 glueless sockets that will be glued together.

Here is a random explanation I found on google:
https://deinoscloud.wordpress.com/2012/10/19/two-main-scale-up-server-architectures-part-1/
https://deinoscloud.wordpress.com/2012/10/22/two-main-scale-up-server-architectures-part-2/

I see, thanks. Before you wrote this I had expected the QPI/UPI links themselves to be the glue.

 

[maddie]

For intel solution, they connect the CPUs with the QPI links directly. A glued solution has an external controller that link the sockets, typically there will be 2 or 4 glueless sockets that will be glued together.

Here is a random explanation I found on google:
https://deinoscloud.wordpress.com/2012/10/19/two-main-scale-up-server-architectures-part-1/
https://deinoscloud.wordpress.com/2012/10/22/two-main-scale-up-server-architectures-part-2/

Based on that explanation, where is the AMD glue?

 

[imported_ats]

If intrasocket latency is important for a role, restricting vCPU assignment to respective NUMA nodes is best practice anyway, even with Intel hardware.

Problem is that with EPYC, you are limited to VMs of 4 CPUs max. Not just that, allocation boundaries matter a lot too. Basically, you need the workload to fix inside of a CCX, if it doesn't, it takes a significant performance hit.

 

[itsmydamnation]

Problem is that with EPYC, you are limited to VMs of 4 CPUs max. Not just that, allocation boundaries matter a lot too. Basically, you need the workload to fix inside of a CCX, if it doesn't, it takes a significant performance hit.

The larger the vcpu count per vm the less total Cpu over subscription you can run because of scheduling contention. In the farms I'm currently working on not a single guest has more then 4 vcpu's. We have Web front end apps with 4-8 hosts load balanced.

Running large vcpu counts isn't ideal, but let's be real here most applications are processing transactions and thus this ccx boundary is meaningless.

 

[quikah]

Based on that explanation, where is the AMD glue?

I don't know, ask Intel. I suppose they are implying that Infinity Fabric is a Glued solution. As I said, I am not that familiar Epyc, so I have no opinion on whether this is BS or not. I was just pointing out that "Glued together" means something specific in the CPU space, not that it was stuck together with literal glue.

 

[coercitiv]

I don't know, ask Intel. I suppose they are implying that Infinity Fabric is a Glued solution. As I said, I am not that familiar Epyc, so I have no opinion on whether this is BS or not. I was just pointing out that "Glued together" means something specific in the CPU space, not that it was stuck together with literal glue.

This was indeed pointed out by other people as well, including Ian Cutress at Anandtech. However, it was also pointed out that in the context of the Intel presentation slides "glued togheter" was used as a pejorative term.

Keep in mind this is the same presentation that portrayed Zen as a consumer desktop arch, not suited for the server space. The technical term Intel used on this instance was Repurposed Desktop Dies for Server.

 

[wahdangun]

Problem is that with EPYC, you are limited to VMs of 4 CPUs max. Not just that, allocation boundaries matter a lot too. Basically, you need the workload to fix inside of a CCX, if it doesn't, it takes a significant performance hit.

nah, IF between CCX in single die is quite good and the performance penalty is not significant and only on handful scenario, so the VM limit is likely 8 core max.

 

[thecoolnessrune]

The larger the vcpu count per vm the less total Cpu over subscription you can run because of scheduling contention. In the farms I'm currently working on not a single guest has more then 4 vcpu's. We have Web front end apps with 4-8 hosts load balanced.

Running large vcpu counts isn't ideal, but let's be real here most applications are processing transactions and thus this ccx boundary is meaningless.

From the last report I looked at, I think out of all of our clients, across thousands of VMs, only a couple percent of the VMs are using more than 4 vCPUs. Of those that are using more, the vast majority of them are using like 16 cores. Most have entire hosts dedicated to them, and the whole VM structure is simply for easy abstraction, backup, and DR. In other words, they would be seeing the same NUMA related penalties running in a VM as they would on the physical host, because they're already using everything the physical host has to offer.

 

[Det0x]

Interesting video for those who want a trip down memory lane.

Intel - Anti-Competitive, Anti-Consumer, Anti-Technology -by AdoredTV

Sadly this isn't the first time Intel use shaddy market tactics, can you can bet it won't be the last time either.

 

[dark zero]

Wow... So Intel is imitating what SEGA did to Nintendo

 

[formulav8]

Intel - Anti-Competitive, Anti-Consumer, Anti-Technology

That's been known at least since I was computering in the early/mid 90's. Wouldn't surprise me if they try to do their stuff again. Hopefully it won't be tolerated this go round if they try that trash again.