AMD EPYC "Turin" with 192 Cores and 384 Threads Delivers Almost 40% Higher Performance Than Intel Xeon 6

[kondamin]

Why is Zen 5 in server config so much better than Zen 5% in desktop config. We are talking up to 10x better performance improvements.

Zen 5 has A much wider pipeline than zen 4
it is capable of doing more work in one tick than zen 4 is but the workloads just aren’t there on the desktop.

buying a new truck that is just as fast but can carry 50% more load isn’t going to get the job done faster if the load remains the same as on the original truck.

amd Should follow intel and split server architecture from the desktop architecture and try to make it lean so it can better deal with higher frequencies

 

[AusWolf]

Zen 5 has A much wider pipeline than zen 4
it is capable of doing more work in one tick than zen 4 is but the workloads just aren’t there on the desktop.

buying a new truck that is just as fast but can carry 50% more load isn’t going to get the job done faster if the load remains the same as on the original truck.

amd Should follow intel and split server architecture from the desktop architecture and try to make it lean so it can better deal with higher frequencies

That means to me that Zen 5 is a server architecture, and the fact that we have desktop parts is a bonus.

I do not think splitting is a good idea, though. It requires double the R&D efforts, double the manufacturing allocation at TSMC, and results in an incoherent hardware/software stack, meaning you can only do certain things with one, but not the other. This is the same problem AMD has with RDNA and CDNA. AMD is trying to unify them within a few generations for a reason.

Edit: Splitting might be a reason for Intel's problems, imo. Too many projects going on at the same time, too much money spent on stuff that doesn't bear fruit, or doesn't have the time to bear fruit before they hop onto the next one. Not to mention, it makes their product stack a convoluted mess.

 

[igormp]

Zen 5 has A much wider pipeline than zen 4
it is capable of doing more work in one tick than zen 4 is but the workloads just aren’t there on the desktop.

buying a new truck that is just as fast but can carry 50% more load isn’t going to get the job done faster if the load remains the same as on the original truck.

amd Should follow intel and split server architecture from the desktop architecture and try to make it lean so it can better deal with higher frequencies

But Intel hasn't split architectures, the cores used in their xeon lineup are the same cores used in their consumer lineup.
Their newest granite rapids uses the same redwood cove found in meteor lake.
Emerald rapids uses raptor cove, same found in desktop 13th gen.
Sapphire rapids used the same golden cove as the desktop 12th gen.
Sierra forest uses the same crestmont found in mtl E-cores.

Zen 5's design already allows it to be used in different nodes (3 and 4nm), while also focusing in both dense and high clock variants.

AMD just hasn't focused in improving gaming performance with this new μarch, but it's an amazing architecture for most other tasks, and is a multithreading best (its 2x4-wide decoder can't be properly used without SMT, as an example).

 

[Tek-Check]

That means to me that Zen 5 is a server architecture, and the fact that we have desktop parts is a bonus.

No. There's Strix Point mobility chips too. There's nothing wrong with Zen 5 in laptops or on Linux in desktop. AMD simply needs to spend more time with Microsoft to roll out a few more updates on Windows.

Besides, if gamers do not benefit as much, it doesn't mean that productivity folks do not. 9950X is a desktop productivity powerhouse that happens to be good enough in gaming, but gaming would not be a good reason to buy it, as there are cheaper and more suitable gaming CPUs.

 

[kondamin]

But Intel hasn't split architectures, the cores used in their xeon lineup are the same cores used in their consumer lineup.
Their newest granite rapids uses the same redwood cove found in meteor lake.
Emerald rapids uses raptor cove, same found in desktop 13th gen.
Sapphire rapids used the same golden cove as the desktop 12th gen.
Sierra forest uses the same crestmont found in mtl E-cores.

Zen 5's design already allows it to be used in different nodes (3 and 4nm), while also focusing in both dense and high clock variants.

AMD just hasn't focused in improving gaming performance with this new μarch, but it's an amazing architecture for most other tasks, and is a multithreading best (its 2x4-wide decoder can't be properly used without SMT, as an example).

Intel stripped out avx512 from the consumer version which is present in the server version.

they also have the big little thing going on

 

[qcmadness]

That means to me that Zen 5 is a server architecture, and the fact that we have desktop parts is a bonus.

I do not think splitting is a good idea, though. It requires double the R&D efforts, double the manufacturing allocation at TSMC, and results in an incoherent hardware/software stack, meaning you can only do certain things with one, but not the other. This is the same problem AMD has with RDNA and CDNA. AMD is trying to unify them within a few generations for a reason.

Edit: Splitting might be a reason for Intel's problems, imo. Too many projects going on at the same time, too much money spent on stuff that doesn't bear fruit, or doesn't have the time to bear fruit before they hop onto the next one. Not to mention, it makes their product stack a convoluted mess.

Zen 5 is an architecture with different implementations.

There are at least 4 implementations:
1. Zen 5 (512-bit data path)
2. Zen 5c (512-bit data path)
3. Zen 5 Lite (256-bit data path)
4. Zen 5c Lite (256-bit data path)

 

[igormp]

Intel stripped out avx512 from the consumer version which is present in the server version.

they also have the big little thing going on

They made components modular, the underlying arch is still the same. Even the current arrow lake cores without HT have that as an option, and we will likely see lion cove on xeons with HT.
As mentioned above, Zen 5 also has different implementations, but it's still the same μarch in the end.

About the big little thing, they're using the little core in some xeon SKU as well, as I pointed before.
They're just not mixing and matching those cores in single SKUs, in the same manner that AMD has no epyc with both normal and C cores (whereas that mix is common in their mobile/APU offerings).

 

[Minus Infinity]

AMD just hasn't focused in improving gaming performance with this new μarch, but it's an amazing architecture for most other tasks, and is a multithreading best (its 2x4-wide decoder can't be properly used without SMT, as an example).

Also maybe, just maybe it's time for the game engines to be better optimised. It's clear a lot of titles are using old engines not designed for modern cpu's. People expecting 30% better performance every gen are unrealistic. There's only so much the cpu can do, some things can't be done in parallel, unlike what gpus can do. No way a game would be able to take advantage of say new wider pipeline without needing work and recompiling.

The fact Intel and AMD are showing poor game improvements despite decent IPC uplifts shows it's not a simple problem.

 

[igormp]

Also maybe, just maybe it's time for the game engines to be better optimised. It's clear a lot of titles are using old engines not designed for modern cpu's. People expecting 30% better performance every gen are unrealistic. There's only so much the cpu can do, some things can't be done in parallel, unlike what gpus can do. No way a game would be able to take advantage of say new wider pipeline without needing work and recompiling.

The fact Intel and AMD are showing poor game improvements despite decent IPC uplifts shows it's not a simple problem.

That circles back to game producers focusing on shoving products out of the door ASAP, developers not having time to focus on such thing as performance, absurdly demanding gaming for no reason being somewhat praised as benchmarks, etc etc.
Whole market is a mess, I'm glad I'm not into games anymore so I honestly dgaf.

 

[Nhonho]

On the EPYC CPU specifications pages, AMD did the "favor" of not showing which type of core the processor has (whether ZEN5, ZEN5c, etc.), nor does it show which instruction sets the processor supports or how much cache memory each chiplet has:

https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9755.html

https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9575f.html

Is this performance difference between the EPYC 9xx5 and 9xx4 CPUs due to improvements in the AVX-512 unit of the ZEN5 architecture?

Why doesn't the EPYC 9755 (128 cores) perform nearly twice as well as the 9575F (64 cores) in this AV1 video encode test? Is this a limitation of the CPUs or of the software?

Source:

AMD EPYC ‘Turin’ 9005 Series - we benchmark 192-core Zen 5 chip with 500W TDP

Cranking up the power.

www.tomshardware.com

 

[Patriot]

On the EPYC CPU specifications pages, AMD did the "favor" of not showing which type of core the processor has (whether ZEN5, ZEN5c, etc.), nor does it show which instruction sets the processor supports or how much cache memory each chiplet has:

https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9755.html

https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9575f.html

Is this performance difference between the EPYC 9xx5 and 9xx4 CPUs due to improvements in the AVX-512 unit of the ZEN5 architecture?

Why doesn't the EPYC 9755 (128 cores) perform nearly twice as well as the 9575F (64 cores) in this AV1 video encode test? Is this a limitation of the CPUs or of the software?

Source:

AMD EPYC ‘Turin’ 9005 Series - we benchmark 192-core Zen 5 chip with 500W TDP

Cranking up the power.

www.tomshardware.com

AMD tells you, Toms doesn't ... The 9575F (F is for frequency) is a boosty boi, hits 5ghz. so on applications that dont scale well to high thread counts, this is going to be a dominant choice.
For more consumer stuff, GN, for server stuff... https://www.servethehome.com/amd-epyc-9005-turin-turns-transcendent-performance-solidigm-broadcom/

 

[semitope]

Crushing performance blow for new Xeons!
Xeons were one generation behind when they were released.
Now, Granite Rapids are two generations behind, yet again.

there's too much noise in the benchmarks. lots of meaningless tests when it comes to what the CPUs are for.

 

[igormp]

there's too much noise in the benchmarks. lots of meaningless tests when it comes to what the CPUs are for.

Those CPUs are meant for lots of different things, so I guess it's more of a matter of one looking at which benchmarks are relevant to their use case.

 

[Tek-Check]

there's too much noise in the benchmarks. lots of meaningless tests when it comes to what the CPUs are for.

It's an average score from hundreds of tests to give us an overall picture.
If you need a server CPU for very specific workload, you will look more into those tests that benefit such workload.

Why doesn't the EPYC 9755 (128 cores) perform nearly twice as well as the 9575F (64 cores) in this AV1 video encode test? Is this a limitation of the CPUs or of the software?

Performance never improves in linear way, with core count increase. CPU performance does not scale and work like that. 16 core CPU is never twice as faster in gaming than 8 core CPU.

Also, 64 core 9575F has almost double bandwidth throughout due to 8 GMI-Wide links to IOD, whereas 128 core Turin has single GMI links from each CCD to IOD. This makes 64 core EPYC very fast indeed in workloads requiring a lot of memory bandwidth.

See the article below.

AMD's Turin: 5th Gen EPYC Launched

5 Gigahertz Server CPUs

chipsandcheese.com