Speculation: Ryzen 4000 series/Zen 3

[lobz]

Intel Xeon Gold 6248R Benchmarks and Review Big Refresh Gains

In our Intel Xeon Gold 6248R benchmarks and review we see the massive impact that the new "Refresh" Xeon line has on the competitive landscape

www.servethehome.com

A good review of relative performance vs 32C and 24C AMD options. Perf per dollar included as well. Intel is in no way in lead, but not getting massacred in lower range of deployments either, which is unexpected considering all Intel process debacle.

Question to those commenters pushing 64C AMD options - why does AMD even bother to offer other SKUs, like 24-32-48? And once you answer this question, does Intel offer comparable systems in same price range?

The perf per dollar metric does not include TCO and by no means could even include features the Intel part just does not offer.
(edit: and vice versa of course)

 

[Hitman928]

Intel Xeon Gold 6248R Benchmarks and Review Big Refresh Gains

In our Intel Xeon Gold 6248R benchmarks and review we see the massive impact that the new "Refresh" Xeon line has on the competitive landscape

www.servethehome.com

A good review of relative performance vs 32C and 24C AMD options. Perf per dollar included as well. Intel is in no way in lead, but not getting massacred in lower range of deployments either, which is unexpected considering all Intel process debacle.

Question to those commenters pushing 64C AMD options - why does AMD even bother to offer other SKUs, like 24-32-48? And once you answer this question, does Intel offer comparable systems in same price range?

I'm not pushing 64C AMD options but I'll answer the obvious, because not everybody needs 64C CPUs, even in a server. As to your second question, comparable is up for debate. If my load is dependent on memory bandwidth, then no, Intel doesn't offer anything comparable. If my load requires the most amount of IO, then no, Intel doesn't offer anything comparable. If I'm only looking at CPU performance then you could start to use the word comparable. However, I would then turn around and ask, instead of asking if they are comparable, why not look at actual performance and ask if the Intel systems offer a performance advantage or even a performance match at the same price range? The answer seems to be no outside of some niche cases, so why would you pay the same or more for less performance with fewer features?

Edit: from your own link,

Having ~50% higher perf/$ than the Intel option is pretty hard to overlook. This doesn't even get into TCO or the other benefits of the Epyc platform at this price range. Again, maybe massacre is a bit hyperbolic, but it certainly isn't pretty for Intel.

 

[Markfw]

Intel Xeon Gold 6248R Benchmarks and Review Big Refresh Gains

In our Intel Xeon Gold 6248R benchmarks and review we see the massive impact that the new "Refresh" Xeon line has on the competitive landscape

www.servethehome.com

A good review of relative performance vs 32C and 24C AMD options. Perf per dollar included as well. Intel is in no way in lead, but not getting massacred in lower range of deployments either, which is unexpected considering all Intel process debacle.

Question to those commenters pushing 64C AMD options - why does AMD even bother to offer other SKUs, like 24-32-48? And once you answer this question, does Intel offer comparable systems in same price range?

This whole discussion started with $4k 28 core DUAL Intel (total $8k) vs one $4k EPYC 64 core. Everypne keeps changing the parameters to make Intel look better. In the 32 core world (or 24 core for Intel) benchmarked here, the 7502 beats the Intel at every benchmark by quite a bit except AVX512. And they cost the same, but Intel lacks a bunch of other things that EPYC has, memory channels, memory size, IO bandwidth. Oh, and not to mention power usage... Anyone that thinks that is not big deal have never worked with a data center and budget. When you save 25% on power, its really 50% or so DUE TO AC costs. EPYC wins there also.

As one with 600 cores and 16 HIGH end video cards at home, my $800 electric bill would be $`1600 if they were all Intel. Thats a BIG consideration.

Edit: and before you go on about why do I need that much horsepower at home...
1) I lost to cancer, my bladder, prostate, hearing and balance, so I have a vested interest in trying to help find cures for it.
2) I am contributor number 54 to Rosetta@home, which based on this research has produced a drug that KILLS Covid-19 (in the lab only so far)

 

[Kenmitch]

The answer seems to be no outside of some niche cases, so why would you pay the same or more for less performance with fewer features?

Intel inside? Unable to swallow the truth? Brand loyalty? Prefer the color blue instead of red? Fear of change? Rooting for the underdog?....Rooting for the underdog?

 

[Hitman928]

Prefer the color blue instead of red?

Epyc branding is actually blue so it's more like prefer brighter blue versus more muted blue, lol.

 

[Gideon]

Here is a better comparison for the Xeon part:

AMD EPYC 7452 Benchmarks and Review 32 Inexpensive Cores

In our AMD EPYC 7452 Benchmarks and Review we see how these 32 inexpensive and low power cores can even consolidate Xeon Silver sockets two to one

www.servethehome.com

The Intel Xeon Gold 6238R puts a 165W (similar TDP albeit 10W higher) into the same range as the EPYC 7452. Based on what we saw with the Intel Xeon Platinum 8276L, a similarly clocked and TDP SKU to the Gold 6238R but designed for 4-socket deployments, we think that AMD still has a competitive offering at a lower price.

The consolidation case is extremely interesting here. With 32 cores and a $63/ core list price it is actually in-line with the Intel Xeon Silver 4216 in terms of pricing. There is a case to be made that one can consolidate two Xeon Silver 4216 sockets to an EPYC 7452 socket and effectively need half as many servers. That can offer enormous TCO savings.

hen we first saw the SKU stack, we likely had an oversight. We thought this SKU was simply a lower-power 32-core “Rome” offering. Instead, the AMD EPYC 7452 offers a truly unique consolidation case. While most processors we see in the AMD EPYC line focus on consolidating two higher-dollar per core processors such as the Xeon Gold or Platinum series, the EPYC 7452 is actually priced to consolidate Xeon Silver sockets by 2:1 or more.

 

[eek2121]

The only reply I am going to make is that while I personally buy AMD these days, my company has given AMD a fair shake. The Intel chips performed much better for our use case and the TCO is similar, so we ended up upgrading 500 of our servers with Intel chips. AMD still has a lot of work to do in enterprise, whether fanboys like it or not. Hopefully Vermeer will make inroads.

 

[Makaveli]

The only reply I am going to make is that while I personally buy AMD these days, my company has given AMD a fair shake. The Intel chips performed much better for our use case and the TCO is similar, so we ended up upgrading 500 of our servers with Intel chips. AMD still has a lot of work to do in enterprise, whether fanboys like it or not. Hopefully Vermeer will make inroads.

Vermeer is a desktop product so you are referring to Milan if we are talking enterprise

And would love for you to detail the use case if you can share.

 

[Markfw]

The only reply I am going to make is that while I personally buy AMD these days, my company has given AMD a fair shake. The Intel chips performed much better for our use case and the TCO is similar, so we ended up upgrading 500 of our servers with Intel chips. AMD still has a lot of work to do in enterprise, whether fanboys like it or not. Hopefully Vermeer will make inroads.

I don't see how they could do that, unless you have a very unusual use case. Rome is better perf/$ and lower power, and lower price. Higher IO capability also. They win all the benchmarks except AVX512 on a $/$ basis, and at least equal, usually better on the core/core but much less expensive basis.

I would like to know where AMD is failing in enterprise. Every review I have seen, and posted here has backed up these facts, nothing to do with who I like better. It also would be good to know what chips you are talking about.

One more note, and on topic (hopefully it will stay that way). Milan should just shut the door on Intel, until they get 10nm/7nm out. Then there might be a close battle, but right now, I can't see how anybody can justify any Intel server ship.

 

[moinmoin]

why does AMD even bother to offer other SKUs, like 24-32-48?

Because they change the balance of I/O capability (which stays the same per package regardless of amount of cores) as well as the amount of cache available per core. For instance Cloudflare chose 48 cores Epyc parts since they have 33% more cache available per core compared to the 64 cores parts.

And once you answer this question, does Intel offer comparable systems in same price range?

Since you now know above answer, you should see this answer coming too: Intel doesn't offer comparable systems period, regardless of the price. 128 PCIe 4 lanes versus 48 PCIe 3 lanes, 256MB cache versus 39MB cache. Maybe you know a fitting word describing that discrepancy?

 

[tamz_msc]

I don't see how they could do that, unless you have a very unusual use case. Rome is better perf/$ and lower power, and lower price. Higher IO capability also. They win all the benchmarks except AVX512 on a $/$ basis, and at least equal, usually better on the core/core but much less expensive basis.

I would like to know where AMD is failing in enterprise. Every review I have seen, and posted here has backed up these facts, nothing to do with who I like better. It also would be good to know what chips you are talking about.

One more note, and on topic (hopefully it will stay that way). Milan should just shut the door on Intel, until they get 10nm/7nm out. Then there might be a close battle, but right now, I can't see how anybody can justify any Intel server ship.

One important use case where Rome doesn't offer much over Cascade Lake-SP(performance gain relative to the number of cores) is database performance, because of its cache layout. Then the per-core licensing of database software makes the 64-core Rome parts less attractive in terms of TCO.

 

[Markfw]

One important use case where Rome doesn't offer much over Cascade Lake-SP(performance gain relative to the number of cores) is database performance, because of its cache layout. Then the per-core licensing of database software makes the 64-core Rome parts less attractive in terms of TCO.

All of the benchmarks were not of cascade lake-SP ?? When I look it up it says "products formerly known as cascade lake" Regardless, what we are discussing as of late are the 32 core 7502, 7402 and 7452.

Edit its late, and I did not spend a lot of time researching htis, but here:
"AMD has released a game-changing processor family for the server market. AMD EPYC 7002 Series processors have more memory capacity (on standard SKUs), more memory bandwidth, and more general-purpose integer and floating-point performance per socket than current Intel Cascade Lake-SP processors. They also have PCIe 4.0 support and a higher number of PCIe lanes than the latest Intel server processors. AMD also gives you more performance per dollar and per watt with the EPYC 7002 Series.

Because of this, I think they will be a great choice for virtualization hosts and for Storage Spaces Direct nodes. I also think they will be a great choice for SQL Server DW/Reporting workloads because of the memory and I/O capacity/bandwidth advantages compared to Intel.
"

Googling for database performance I found this here https://sqlperformance.com/2019/08/sql-performance/amd-epyc-7002-sql-server

 

[tamz_msc]

All of the benchmarks were not of cascade lake-SP ?? When I look it up it says "products formerly known as cascade lake" Regardless, what we are discussing as of late are the 32 core 7502, 7402 and 7452.

Edit its late, and I did not spend a lot of time researching htis, but here:
"AMD has released a game-changing processor family for the server market. AMD EPYC 7002 Series processors have more memory capacity (on standard SKUs), more memory bandwidth, and more general-purpose integer and floating-point performance per socket than current Intel Cascade Lake-SP processors. They also have PCIe 4.0 support and a higher number of PCIe lanes than the latest Intel server processors. AMD also gives you more performance per dollar and per watt with the EPYC 7002 Series.

Because of this, I think they will be a great choice for virtualization hosts and for Storage Spaces Direct nodes. I also think they will be a great choice for SQL Server DW/Reporting workloads because of the memory and I/O capacity/bandwidth advantages compared to Intel.
"

Googling for database performance I found this here https://sqlperformance.com/2019/08/sql-performance/amd-epyc-7002-sql-server

If you look at the test results, in terms of the "per core" scores, Xeon is ahead by 19% in one test and Epyc is ahead by 28% in another test, so the performance gap isn't quite clear-cut. Then when you consider that the 6258R exists, comparisons that involve the 8280 become irrelevant.

 

[JoeRambo]

Then when you consider that the 6258R exists, comparisons that involve the 8280 become irrelevant.

And 6248R also exists, making comparisons with 6258 irrelevant too. 95% of perf for 60% price advantage.

Because they change the balance of I/O capability (which stays the same per package regardless of amount of cores) as well as the amount of cache available per core. For instance Cloudflare chose 48 cores Epyc parts since they have 33% more cache available per core compared to the 64 cores parts.

Just because 2TB option exists, people don't put 2TB into every server, also just because some CPU has a truckload of IO, certain uses might need only fixed amount of I/O. Same with core counts, marketing 101 and making use of defective cores also factor in.
Epyc2 is certainly good in web "serving/caching" workload, cause that is mostly independent "static" data that loves cache, cores and storage.

More complex workloads like actual web apps (sometimes, but not always) include some sort of DB backend, generate "dynamic" data that crosses CCX boundaries and eat into internal chip latencies. Not all software is well written and there are definitely hidden gotchas and performance pitfalls due to AMD chip level architecture ( cores are great!).
There is always a mindshare problem too. For example on Intel i know what to do and what tools to run if I suspect workload mix is overloading QPI links and diagnose "hw level" performance problems. On AMD -> that know-how is not available in my company.

The only reply I am going to make is that while I personally buy AMD these days, my company has given AMD a fair shake. The Intel chips performed much better for our use case and the TCO is similar, so we ended up upgrading 500 of our servers with Intel chips. AMD still has a lot of work to do in enterprise, whether fanboys like it or not. Hopefully Vermeer will make inroads.

^This^ I own both 3950x (and 9980xe) and 10900K, but for web app servers we will continue to use Intel chips. Definitely looking forward to AMD moving to 8C CCX and hopefully removing some of pitfalls and concerns in our use case.

 

[teejee]

I don't see how they could do that, unless you have a very unusual use case. Rome is better perf/$ and lower power, and lower price. Higher IO capability also. They win all the benchmarks except AVX512 on a $/$ basis, and at least equal, usually better on the core/core but much less expensive basis.

I would like to know where AMD is failing in enterprise. Every review I have seen, and posted here has backed up these facts, nothing to do with who I like better. It also would be good to know what chips you are talking about.

One more note, and on topic (hopefully it will stay that way). Milan should just shut the door on Intel, until they get 10nm/7nm out. Then there might be a close battle, but right now, I can't see how anybody can justify any Intel server ship.

Why would it be strange with a use case where Intel is significantly faster? It is enough that the main use case is with a SW that is badly optimized for AMD.

For example Matlab was very slow on Ryzen until a recent update.
(used SIMD instead of AVX2...)

I'm sure there are many other similar cases among less known software.

 

[lobz]

Why would it be strange with a use case where Intel is significantly faster? It is enough that the main use case is with a SW that is badly optimized for AMD.

For example Matlab was very slow on Ryzen until a recent update.
(used SIMD instead of AVX2...)

I'm sure there are many other similar cases among less known software.

That's exactly the strangest thing ever, since the guy's talking about 500 systems. The cost difference there is most likely in the millions. He's also talking about a large web application. Like I said. There are cases where there's no other good option than to stick with Intel. However, I'm unlikely to be convinced that every time something is poorly optimized, the best solution is just to buy hardware that is a million dollars more expensive. That is just the most short-sighted business practice I can imagine.

 

[Zstream]

And 6248R also exists, making comparisons with 6258 irrelevant too. 95% of perf for 60% price advantage.



Just because 2TB option exists, people don't put 2TB into every server, also just because some CPU has a truckload of IO, certain uses might need only fixed amount of I/O. Same with core counts, marketing 101 and making use of defective cores also factor in.
Epyc2 is certainly good in web "serving/caching" workload, cause that is mostly independent "static" data that loves cache, cores and storage.

More complex workloads like actual web apps (sometimes, but not always) include some sort of DB backend, generate "dynamic" data that crosses CCX boundaries and eat into internal chip latencies. Not all software is well written and there are definitely hidden gotchas and performance pitfalls due to AMD chip level architecture ( cores are great!).
There is always a mindshare problem too. For example on Intel i know what to do and what tools to run if I suspect workload mix is overloading QPI links and diagnose "hw level" performance problems. On AMD -> that know-how is not available in my company.




^This^ I own both 3950x (and 9980xe) and 10900K, but for web app servers we will continue to use Intel chips. Definitely looking forward to AMD moving to 8C CCX and hopefully removing some of pitfalls and concerns in our use case.

WOW! I won’t go into it, but based off this post, you guys have no business being in the web application space. Web applications use reddis, varnish, mongoDB and other caching mechanisms. If you’re building a web app to use a monolithic approach, you failed.. completely. Key value pairs are also synced in memory.

 

[JoeRambo]

WOW! I won’t go into it, but based off this post, you guys have no business being in the web application space. Web applications use reddis, varnish, mongoDB and other caching mechanisms. If you’re building a web app to use a monolithic approach, you failed.. completely. Key value pairs are also synced in memory.

Doesn't it depend on nature of web application? For example in finance sector, pretty much all content in those online banks is dynamic, you can't really cache much, while a news front or forum software can cache probably 99% of content? And "monolithic" comment is rather a joke, thanks for teaching a "lesson".

 

[Markfw]

Why would it be strange with a use case where Intel is significantly faster? It is enough that the main use case is with a SW that is badly optimized for AMD.

For example Matlab was very slow on Ryzen until a recent update.
(used SIMD instead of AVX2...)

I'm sure there are many other similar cases among less known software.

Since in all the benchmarks above, EPYC is faster in everything but AVX512

 

[Zstream]

Doesn't it depend on nature of web application? For example in finance sector, pretty much all content in those online banks is dynamic, you can't really cache much, while a news front or forum software can cache probably 99% of content? And "monolithic" comment is rather a joke, thanks for teaching a "lesson".

You said web apps. Finance software uses database calls and make unique requests for data to the database as real data such as your bank account can change every second. There is very little cached on the CPU for that reason, as the data has to be refreshed from the database every time a call is made. If you’re asking about vector, math, etc.. like crunching, folding at home, or repeat adobe functions than yes.

 

[Gideon]

You said web apps. Finance software uses database calls and make unique requests for data to the database as real data such as your bank account can change every second. There is very little cached on the CPU for that reason, as the data has to be refreshed from the database every time a call is made.

I really hope those run on distributed commit logs (e.g. Kafka & co) and microservices rather than some ancient transactional databases (god-forbid, if it's Oracle) and some "enterprise Java" monoliths that only scale vertically. If that's the fromer is the case, then some saner architectural decsision can be made. If the latter ...

 

[Markfw]

You said web apps. Finance software uses database calls and make unique requests for data to the database as real data such as your bank account can change every second. There is very little cached on the CPU for that reason, as the data has to be refreshed from the database every time a call is made. If you’re asking about vector, math, etc.. like crunching, folding at home, or repeat adobe functions than yes.

To you and @Gideon and anyone else who is wondering on web apps. How about we find some benchmarks to back up what you are all saying. My opinions were based a several sites benchmarks. The word of people that things are slower when "blah, blah, blah" are all opinions. Let back this up with a set of benchmarks that we can argue about. And please try to pick ones from sites that we all here recognize as valid popular sites.

 

[Pumice]

Any news yet on when the desktop Zen 3 processors are coming out ?

 

[thigobr]

The only word so far is before 2020 ends...

 

[TitusTroy]

I've been waiting for Zen 3 before starting my new build...the PSU and motherboard shortages over the past few weeks/months has made it easier but I was recently able to obtain the PSU I wanted (Corsair RM750x) as well as a good X570 motherboard (MSI X570 Tomahawk)...so now I have them sitting in my place un-opened and the wait now feels longer for Zen 3

I previously heard rumors about a September launch but now I'm hearing it might not be until November...I really don't think I can wait that long...if it doesn't get released next month I might just get a 3700X