Ryzen 3900X

[MikeB]

I think previous Epycs were only 64 cores. I didn't know they were making a 128 core version now. What does it mean for computer chess?

The AMD 7742 has 64 cores, and with certain models, they can be run in a "dual socket" server or workstation, meaning two AMD 7742 processors with 64 cores each on one motherboard - 128 cores in total. A "2x" in the description of a computer next to the procesor description usually signifies two processors on one motherboard.

[Leo]

I think previous Epycs were only 64 cores. I didn't know they were making a 128 core version now. What does it mean for computer chess?

The AMD 7742 has 64 cores, and with certain models, they can be run in a "dual socket" server or workstation, meaning two AMD 7742 processors with 64 cores each on one motherboard - 128 cores in total. A "2x" in the description of a computer next to the procesor description usually signifies two processors on one motherboard.

OK. Thanks.

[Joost Buijs]

...

And the bottom graphic here shows that the very expensive AMD 7742 smokes the AMD 7601 and utterly destroys the Intel 8280 in terms of nodes per dollar with SF and AF.

...

Indeed,

If you compare the dual Xeon Platinum 8280 system at 125M nps to the projected 106M nps of 3rd gen TR 32-core system one wonders why anyone would ever spend an extra $18,000 on CPU's for the Intel system.

Regards,

Zenmastur

The same reason why some people drive Lamborghini and others drive Austin cars.

That's not the point. The point is how much faster is the fastest car than a cheap car. In this case the difference in speed from a cheap car to the fastest car has been significantly reduced because the cheap car is now much faster.

Regards,

Zenmastur

https://wccftech.com/intel-cascade-lake ... hedt-cpus/
https://newsroom.intel.com/wp-content/u ... l-RWPE.pdf

[dragontamer5788]

...

And the bottom graphic here shows that the very expensive AMD 7742 smokes the AMD 7601 and utterly destroys the Intel 8280 in terms of nodes per dollar with SF and AF.

...

Indeed,

If you compare the dual Xeon Platinum 8280 system at 125M nps to the projected 106M nps of 3rd gen TR 32-core system one wonders why anyone would ever spend an extra $18,000 on CPU's for the Intel system.

Regards,

Zenmastur

Probably Databases.

While EPYC and Threadripper are great for most workloads, consider how their L3 caches are split between "CCX" (core complex). You can only access 16MB of L3 cache per thread.

"External" L3 cache (between CCX) has been measured to be slower than accessing DDR4 RAM directly (!!). So any one thread can only access, at best, 16MB within 10-nanoseconds (the 16MB that fits in the "local" CCX L3 cache). In contrast, a single thread on a Xeon 8280 will be able to access all 38.5MB of L3 cache within 10-nanoseconds. In an L3 constrained environment (such as databases), this leads to better performance.

AMD EPYC / Threadripper can have 256MB of aggregate L3 cache, but its individual threads max out at 16MB of L3 cache. Now, this is a good thing for workloads such as Lazy SMP (only the "relevant" parts of the shared hash table get loaded into the local L3 caches of each thread). But its a bad thing for databases.

[Dann Corbit]

https://www.businesswire.com/news/home/ ... ld-Records
"In a similar fashion for database virtualization, the HPE ProLiant DL325 delivered an astounding 321 percent performance boost over the previous record holder, providing a significant agility advantage in database provisioning and maintenance for cloud, big data and IoT environments."

[dragontamer5788]

https://www.businesswire.com/news/home/ ... ld-Records
"In a similar fashion for database virtualization, the HPE ProLiant DL325 delivered an astounding 321 percent performance boost over the previous record holder, providing a significant agility advantage in database provisioning and maintenance for cloud, big data and IoT environments."

Database virtualization: yes. Because you split up your server into pieces and give it out to other customers. Which EPYC's "split" L3 cache design is outstanding, and its the ideal use.

However, if you run a bare-metal database, such as a single server with Redis, Memcached, or Postgres, the Xeon's "shared L3" cache becomes far more relevant. Larger, singular databases scale far better than trying to create replication clusters. Now granted: to "scale up" a server costs significantly more than to "scale out" with replecation / virtualization, but there's gains with regards to simplicity. Sometimes, spending $10,000 is cheaper than switching to a scale-out methodology (which may require programmers, aka $200,000+).

Its a niche, but bare-metal databases are still a thing. On the other hand, that 18-core / 72-thread POWER9 chip with 90MB L3 cache (!!) for around $2000 is better than a $10,000 Xeon with only 38.5MB of L3 cache. It seems like EPYC will win for cloud-instances and many multi-threaded programs, while POWER9 is superior for the memory-constrained / L3 constrained workloads.

Intel Xeon stays in the x86 ecosystem though, and still is better "scale up" than EPYC. I'm not saying AMD EPYC is bad or anything, I'm purposely focusing on a niche where an IT engineer might decide to go Xeon still. In most deployments, EPYC looks like a superior choice. But remember, EPYC is a "Scale out" platform. Split L3 caches, CCX, etc. etc. Its great for the modern containerized / virtualized workload, but not everyone has switched to containers / scale out designs yet. (And besides: scale up is simpler to design and run: fewer servers often means less downtime)

[Dann Corbit]

https://www.technologyperformancepulse. ... erformance

[Zenmastur]

...

And the bottom graphic here shows that the very expensive AMD 7742 smokes the AMD 7601 and utterly destroys the Intel 8280 in terms of nodes per dollar with SF and AF.

...

Indeed,

If you compare the dual Xeon Platinum 8280 system at 125M nps to the projected 106M nps of 3rd gen TR 32-core system one wonders why anyone would ever spend an extra $18,000 on CPU's for the Intel system.

Regards,

Zenmastur

Probably Databases.

While EPYC and Threadripper are great for most workloads, consider how their L3 caches are split between "CCX" (core complex). You can only access 16MB of L3 cache per thread.

"External" L3 cache (between CCX) has been measured to be slower than accessing DDR4 RAM directly (!!). So any one thread can only access, at best, 16MB within 10-nanoseconds (the 16MB that fits in the "local" CCX L3 cache). In contrast, a single thread on a Xeon 8280 will be able to access all 38.5MB of L3 cache within 10-nanoseconds. In an L3 constrained environment (such as databases), this leads to better performance.

AMD EPYC / Threadripper can have 256MB of aggregate L3 cache, but its individual threads max out at 16MB of L3 cache. Now, this is a good thing for workloads such as Lazy SMP (only the "relevant" parts of the shared hash table get loaded into the local L3 caches of each thread). But its a bad thing for databases.

At first I thought you were just trolling. But...

In the end the only measure that makes a difference is how it performs on your specific application. I don't care if it takes an hour per cache transfer as long as my programs complete their workload in less time. In most cases, EPYC seems to do exactly that AND at less than half the cost. You can play with the numbers all you like, the only thing that really matters is does it get the job done cheaper and/or faster. For small business and cloud customers it seems like it's by far the best choice for most applications.

Regards,

Zenmastur

[dragontamer5788]

https://www.technologyperformancepulse. ... erformance

Hmm, I was going off of pgbench results from last gen and my expectations for Zen2. But if TPC has a set of results for the new generation, that overrides what I was going off of. Good to know.

I wish there was a 1x Xeon result. There's only a 2x Xeon result: http://www.tpc.org/tpch/results/tpch_re ... =117111701

2x28 Core setup outperforms the 1x32-core Zen2 result: http://www.tpc.org/tpch/results/tpch_re ... =119080501

There's not an apples-to-apples comparison, unfortunately. The other Zen2 results for 3000 GB are for "EXASOL" Database, not SQL Server 2017. So the above two results seem like the best apples-to-apples comparison. But 2x 28Core vs 1x32 core leaves a lot to be desired...

http://www.tpc.org/tpch/results/tpch_pr ... lttype=all

----------

EDIT: Also, those benchmarks are "bigger" than I was expecting. I was thinking more about RAM-sized databases. Once the dataset grows beyond the size of RAM, EPYC gains an advantage due to having more PCIe lanes than Xeon. Hmm, now that I think of it, EPYC seems to have a lot of advantages.

[Joost Buijs]

I wonder when Zen 2 will be readily available, 3900X is almost impossible to get, 3950x is still not announced, and the processors that are released have trouble to reach the specified turbo speed. As ususal it is drama with AMD, I'm not an Intel fan whatsoever, but if Intel delivers 'Cascade Lake X' for a price that compares favorably to AMD Zen 2, I will go that way.