AMD releases new server CPU, EPYC 7H12; stronger than the EPYC 7742

[Sakkura]

9 minutes ago, porina said:

Where do they say that? In a quick look I found the following, where there is a minus sign in front of the boost value.

 

At the start, in the second paragraph.

 

https://www.anandtech.com/show/14882/amds-new-280w-64core-rome-cpu-the-epyc-7h12

 

Quote

The new processor features a higher base frequency and a higher boost frequency that the previous top-of-the-line processor, the EPYC 7742. The new EPYC 7H12 has a rated TDP of 280W, and as a result the chip is being marketed for server environments that offer liquid cooled solutions only. AMD is very specific about this, especially in the market for which this CPU is aimed at. One of AMD’s main partners, Atos, is set to offer an 1U solution featuring eight of these CPUs, all liquid cooled.

 

[leadeater]

5 hours ago, VegetableStu said:

8 in the same infinity-fabric network, or multiple discrete systems in same 1U space? o_o

I don't remember EPYC having more than two CPUs per system

2P nodes in a 1U chassis by the looks of it.

 

Quote

Specifications:

• 1344 compute nodes:
  • 2 x 64core, next-generation AMD EPYC processors, code name ‘Rome
  • Total number of cores: 172032
  • Peak performance: 5.9 Petaflops
• Interconnect: Mellanox HDR InfiniBand
• Storage:
• DDN Storage
• 2.5 Petabyte Lustre parallel file system
• Footprint: 14.78m2.
• Power consumptuion: 952kW
• Liquid Cooling: 95% of the heat will be captured to water

https://insidehpc.com/2019/06/atos-to-deploy-most-powerful-supercomputer-in-norway/

 

Edit:

[GOTSpectrum]

5 hours ago, cj09beira said:

back on topic it seems amd isn't taking its time to make the more specific skus, i wonder then if the high clock speed versions (with lower core count) will be coming soon

32-48 core 3.6ghz when?

[GOTSpectrum]

5 hours ago, NunoLava1998 said:

Eight of these in a 1U. That's 512 cores, 1024 threads.

@Ithanul @Gorgon imagine one of these folding, just got the luls

[leadeater]

2 minutes ago, GOTSpectrum said:

@Ithanul @Gorgon imagine one of these folding, just got the luls

Sadly* Anandtech can't count, it's only 3 2P or 6 total per 1U

 

*As sad as 6 of these are lol

[GOTSpectrum]

4 minutes ago, leadeater said:

Sadly* Anandtech can't count, it's only 3 2P or 6 total per 1U

 

*As sad as 6 of these are lol

Yeah but... How many per rack...

[GOTSpectrum]

6 minutes ago, GOTSpectrum said:

Yeah but... How many per rack...

P.S. that was rhetorical as I know racks come in different sizes... 

[Ithanul]

2 hours ago, GOTSpectrum said:

@Ithanul @Gorgon imagine one of these folding, just got the luls

Probably not worth it since bigadv is long dead in folding.  O how I wish it come back.

[GOTSpectrum]

20 minutes ago, Ithanul said:

Probably not worth it since bigadv is long dead in folding.  O how I wish it come back.

I'm aware, it would be interesting to see though

[Ithanul]

13 minutes ago, GOTSpectrum said:

I'm aware, it would be interesting to see though

Probably just spit out like 300-400K or so, maybe.  But, the power usage would be nuts.

 

:I  I did a run in F@H with my 1950X.  Not a whole lot of points and all that heat and power usage was not worth it (I aim it at BOINC instead).

 

And, I don't joke about 1950X being a power hog.

A fellow 1950X BOINC cruncher and his full system power draw when their chip crunches:

Spoiler

I haven't even dare to take a power meter to my main rig...

[- A -]

Whats concerning about this CPU is the TDP/Power consumption of this thing. You can easily load your electrical bill with just a week of use in this thing. But, enterprise wise, it looks like a decent CPU with specs that are really for server utilization. Though I'm quite curious on the cost-to-performance ratio of this thing. We might see Linus actually reviewing this thing and start another build for PC gaming using this.

 

[Drak3]

55 minutes ago, Ithanul said:

I don't joke about 1950X being a power hog.

It's almost like it has 16 cores and a healthy clockspeed.

[Ithanul]

1 minute ago, Drak3 said:

It's almost like it has 16 cores and a healthy clockspeed.

The individual had it tune to 3.7GHz.  So, no crazy 4.0 - 4.2GHz in their setup.

Personally, I keep mine at stock or tune lower to keep power usage in check.

[ouroesa]

Think most people have no regard for the strength of a CPU, more about speed and effeciency. 

[- A -]

5 hours ago, ouroesa said:

Think most people have no regard for the strength of a CPU, more about speed and effeciency. 

Well with regards on its performance, it can be inferred that it has enough power for server utilization. But for modern consumers that introduces an enterprise utilization, it might be tricky for them to settle on this new server CPU and might settle for the previous generations instead.

Moreover, I think it's reasonable to settle for this CPU. It could be good for the next couple of years.

[Mesterial]

15 hours ago, GOTSpectrum said:

Yeah but... How many per rack...

30 blades, so 90 nodes, at max.

 

(I know it because I work there, and participated in the bench of these!)

 

[Doobeedoo]

Monster CPU seriously really a beast with that core count and TDP rating. 

[X-System]

22 hours ago, NunoLava1998 said:

AMD has released a new EPYC server CPU, the EPYC 7H12. This is an even stronger CPU than the EPYC 7742. At 64 cores, 128 threads and 2.6GHz base clock and 3.3GHz boost clock, this is absolutely insane, especially for servers.

It's great but the performance of the calculation is not enough face to Xeon 9000 series for the supercomputers.

 

This new AMD EPYC has always 256-bit AVX2 with two 256-bit FMA units. It makes 16 double precision FLOPS per cycle per core.

 

https://www.microway.com/knowledge-center-articles/detailed-specifications-of-the-amd-epyc-rome-cpus/

 

With 64 cores at 2.6 GHz, it gives 2.66 TeraFLOPS per CPU.

 

And with Xeon 9282 (56 cores with two AVX-512 FMA units per core). It makes 32 double precision FLOPS per cycle per core.

 

https://www.microway.com/knowledge-center-articles/detailed-specifications-of-the-cascade-lake-sp-intel-xeon-processor-scalable-family-cpus/ (I don't found for Cascade Lake AP information but I think it's the same)

https://ark.intel.com/content/www/fr/fr/ark/products/194146/intel-xeon-platinum-9282-processor-77m-cache-2-60-ghz.html

 

With 56 cores at 2.6 GHz, it gives 4.66 TeraFLOPS per CPU.

 

For the performance of the calculation, Intel wins.

 

For the power consumption, 9.5 GFLOPS per watt for AMD EPYC 7H12 and 11.65 GFLOPS per watt for Intel Xeon 9282. Intel wins.

 

And for the price, AMD wins for sure. But for the greenpeace, I think they choice Intel Xeon for the performance per watt.

 

But for the classic servers, AMD EPYC is a better choice.

[Mesterial]

30 minutes ago, X-System said:

With 64 cores at 2.6 GHz, it gives 2.66 TeraFLOPS per CPU.

 

With 56 cores at 2.6 GHz, it gives 4.66 TeraFLOPS per CPU.

 

These are only theoretical values! You have to introduce the HPL efficiency, which is a catastrophy for Intel at AVX512 frequency compared to base clock.

[porina]

In HPC, FLOPS are only part of the performance equation. It gets complicated fast with cache configuration, bandwidth between parts, and more. Feeding AVX-512 can be difficult.

 

y-cruncher doesn't seem to be a popular benchmark, but it is written to scale with hardware and IMO is more interesting than Cinebench. Still, see below.

 

I currently hold the hwbot y-cruncher 10b records for 2, 6, 8, 12 core CPUs. My best overall time is ~350s on a 7920X OC to 3.7 GHz. Need better cooling to push further on 2 (stock i3-6100), 8 (stock 3700X), 12 core times. I'm done on the 6 core though (7800X@4.3 GHz).

[Franck]

Is there any test or documents about stability on Hypervisor vs different Xeon. Specially Xeon Gold series. This is a very good news for having lots of VM at low cost and small factor. It's insane, i can replace literally 6 U1 with a single one. Could get easily and extra drive storage by replacing 6 U1.

[leadeater]

6 hours ago, Mesterial said:

These are only theoretical values! You have to introduce the HPL efficiency, which is a catastrophy for Intel at AVX512 frequency compared to base clock.

It also ignores cost and power, if you aren't floor area limited just deploy more EPYC2 nodes at a cheaper cost.

 

Also there is no way there is enough memory bandwidth to supply 56 dual AVX-512 cores at peak throughput, ain't happening .

 

Edit:

Also using base clocks for the EPYC2 which is wrong too.

[mr moose]

7 hours ago, X-System said:

 

And for the price, AMD wins for sure. But for the greenpeace, I think they choice Intel Xeon for the performance per watt.

 

 

the only thing i have to say on the topic is that greanpeace don't care about critical thinking or actual good tech, if the CPu isn't made form recycled bamboo and can be made form 100% wind or solar than the CPU is evil and must be boycotted to save the gay whales and hippy dolphins.

[Mesterial]

Power is going insane on the Intel part, if you ask me... 

The 9282 is 400w TDP, that's 800w only for the CPUs of a single node !!  

I might be an Intel fanboy, but sometimes I hardly understand the last years decisions... 

[spartaman64]

15 minutes ago, Mesterial said:

Power is going insane on the Intel part, if you ask me... 

The 9282 is 400w TDP, that's 800w only for the CPUs of a single node !!  

I might be an Intel fanboy, but sometimes I hardly understand the last years decisions... 

and 56 cores with a monolithic die i dont want to imagine the yield rate

edit: nvm the higher core count ones are multichip i guess glued together dies are not that bad after all