Info [Toms, Anand] AMD EPYC Benchmarks

[sdifox]

But then there is this note:

* 200W CPUs are supported under certain configurations. Please contact Supermicro Technical Support for additional information about specialized system optimization

So I assume that 225 watt are NOT supported. Maybe not 200 watt if all of them are that much.

Edit: found this also:

Support CPU TDP / cTDP up to 200W*

Booo... on the other hand 8x Epyc 7702 will cost US$51.6k lol

 

[nicalandia]

Crushing Cinebench V5 AMD EPYC 7742 World Record Edition
https://www.servethehome.com/crushing-cinebench-v5-amd-epyc-7742-world-record-edition/

The system was all stock settings, I can't wait to see the kind of Numbers that German Overclocker can achieve with OC and tweaking

 

[Abwx]

We should have some info from Lisa Su s conference at Hot Chips, with a lake of Intel sponsoring in between :


7:30 AM – 8:45 AM: Registration / Breakfast (sponsored by Intel)
9:00 AM – 10:30 AM: General Purpose Compute
10:30 AM – 11:00 AM: Break (sponsored by Intel)
11:00 AM – 12:30 PM: Memory
12:30 PM – 1:45 PM: Lunch (sponsored by Intel)
1:45 PM – 2:45 PM: Keynote 1: “Delivering the Future of High-Performance Computing with System, Software and Silicon Co-Optimization” by Dr. Lisa Su, CEO, AMD
2:45 PM – 4:15 PM: Methodology and ML Systems
4:15 PM – 4:45 PM: Break (sponsored by Intel)
4:45 PM – 6:45 PM: ML Training
6:45 PM – 7:45 PM: Reception (Wine & Snacks sponsored by Intel)

https://www.hotchips.org/program/

 

[zinfamous]

We should have some info from Lisa Su s conference at Hot Chips, with a lake of Intel sponsoring in between :


7:30 AM – 8:45 AM: Registration / Breakfast (sponsored by Intel)
9:00 AM – 10:30 AM: General Purpose Compute
10:30 AM – 11:00 AM: Break (sponsored by Intel)
11:00 AM – 12:30 PM: Memory
12:30 PM – 1:45 PM: Lunch (sponsored by Intel)
1:45 PM – 2:45 PM: Keynote 1: “Delivering the Future of High-Performance Computing with System, Software and Silicon Co-Optimization” by Dr. Lisa Su, CEO, AMD
2:45 PM – 4:15 PM: Methodology and ML Systems
4:15 PM – 4:45 PM: Break (sponsored by Intel)
4:45 PM – 6:45 PM: ML Training
6:45 PM – 7:45 PM: Reception (Wine & Snacks sponsored by Intel)

https://www.hotchips.org/program/

So would it be fair to say that AMD and TSMC (look at day 2) are literally eating Intel's lunch right now?

 

[DrMrLordX]

So would it be fair to say that AMD and TSMC (look at day 2) are literally eating Intel's lunch right now?

. . . and breakfast. Plus Intel is bringing cheese to go with their whine wine. AMD is eating that too!

 

[nicalandia]

It's live now

 

[nicalandia]

It's live now

Well, not much info was given about Threadripper perhaps a confirmation that the IO chip is 12nm

But hey more Epyc Benchmarks:

POWER9 & ARM Performance Against Intel Xeon Cascadelake + AMD EPYC Rome
https://www.phoronix.com/scan.php?page=article&item=rome-power9-arm&num=1

 

[Abwx]

.....................................

So the numbers are clarified, 15% was for ST and MT is 23%, no wonder that Ryzen 3000 does so well despite roughly 10% lower MT frequency than CFL..

https://www.anandtech.com/show/1476...ay-1-dr-lisa-su-ceo-of-amd-live-blog-145pm-pt

 

[nicalandia]

So the numbers are clarified, 15% was for ST and MT is 23%

I was under the impression that we were all aware that the 15% IPC improvement were per ST? the 23% on MT performance(Naples vs EPYC) is due to better topology(NUMA vs UMA in 1P and 8 vs 2 NUMA nodes in 2P)

I extrapolated the same(about 22% performance IPC uplift using 12 core TR1 vs 3900X clock for clock)

 

[nicalandia]

I would also like to add that due to Topology Xeons only gain about 20% performance boost with HyperThreading on and Epyc Naples 28%(in SMT Integer Performance With SPEC CPU2006) This means that Future Xeons not only have to deal with the ST IPC boost of Rome/Milan but also an even higher boost in IPC with SMT On which in DataCenters and Server world is about 95% of the tasks


Desktop Intel HT boost is about 25% and Desktop Zen2 SMT boost is about 30% due to simpler topology

 

[DrMrLordX]

Well, not much info was given about Threadripper perhaps a confirmation that the IO chip is 12nm

But hey more Epyc Benchmarks:

POWER9 & ARM Performance Against Intel Xeon Cascadelake + AMD EPYC Rome
https://www.phoronix.com/scan.php?page=article&item=rome-power9-arm&num=1

I'm disappointed that they didn't have ThunderX2 in there.

Also WRT your commentary about SMT/HT, how many Intel servers out there are being run with HT off?

 

[moinmoin]

Also WRT your commentary about SMT/HT, how many Intel servers out there are being run with HT off?

As soon as virtualization is involved HT is usually off (for more predictable resource partitioning).

 

[DrMrLordX]

Yeah that's what I thought. Disabling HT is also a security measure for some shops.

 

[Topweasel]

Yeah that's what I thought. Disabling HT is also a security measure for some shops.

Or at least it has to be now.

 

[nicalandia]

Desktop Intel HT boost is about 25% and Desktop Zen2 SMT boost is about 30% due to simpler topology

I would like to point out that I was off by a large margin about Intel HT gains and Amd SMT gains, Intel will gain about 28% with HT On and Amd will gain 40% with SMT On, at least based on Cineben R15 and going by Anandtech recent 3900X Review, Zen 2 is up to 42% increase in performance

Stock i7 6900K HT On and HT off

8C/8T: 1231
8C/16T: 1580(28% increase over 8C/8T)


Stock Ryzen 1700X(Zen1) SMT OFF And SMT On

8C/8T: 1104
8C/16T: 1548(40% increase over 8C/8T)

 

[StefanR5R]

I would like to point out that I was off by a large margin about Intel HT gains and Amd SMT gains, Intel will gain about 28% with HT On and Amd will gain 40% with SMT On, at least based on Cineben R15

It depends a lot on the workload.
Code which saturates the FMA units for example typically does not gain anything at all from HT (on Haswell and Skylake and their variations). Ditto for Zen and its SMT; I don't know how it is with Zen 2.

 

[nicalandia]

It depends a lot on the workload.
Code which saturates the FMA units for example typically does not gain anything at all from HT (on Haswell and Skylake and their variations). Ditto for Zen and its SMT; I don't know how it is with Zen 2.

I agree, but I was pointing pointing about about apps that do take advantage of HT/SMT where you will see gains,


As per Anandtech Review: https://www.anandtech.com/show/14605/the-and-ryzen-3700x-3900x-review-raising-the-bar/19

Zen 2(3900X) 4.3 OC
1C/1T : 204
12/24T: 3375
Since 100% CPU Scaling is really not possible(specially in 12 core CPUs with CCX and IO) but giving it the same % scaling of Zen/Zen+(97%) Zen 2 shows a 42% boost in performance with SMT On.

 

[nicalandia]

Dual AMD EPYC 7742 Crushes Quad Intel Xeon 8180M's In Geekbench 4
https://www.tomshardware.com/news/dual-amd-epyc-7742-vs-quad-intel-xeon-platinum-8180m,40288.html

"If we do the math, each EPYC 7742 costs $6,950 each Xeon Platinum 8180M goes for $13,011. So two EPYC 7742 cost you $13,900 and four Xeon Platinum 8180M sets you back $52,044. You're getting 24.83% more performance while costing 73.29% less."

 

[beginner99]

Dual AMD EPYC 7742 Crushes Quad Intel Xeon 8180M's In Geekbench 4
https://www.tomshardware.com/news/dual-amd-epyc-7742-vs-quad-intel-xeon-platinum-8180m,40288.html

"If we do the math, each EPYC 7742 costs $6,950 each Xeon Platinum 8180M goes for $13,011. So two EPYC 7742 cost you $13,900 and four Xeon Platinum 8180M sets you back $52,044. You're getting 24.83% more performance while costing 73.29% less."

Nobody really pays list prices for xeons. Also in the TCO of the whole server incl. software the cpu plays a minor role. Of course yeah you fare better with epyc this time but in general CPU price isn't that important in the big picture.

 

[gorobei]

the Moore'sLawIsDead channel did a podcast with a finance industry IT guy, they go over the economics and purchase decision process in re intel security flaws and performance.

according to the IT guy: price is functionally meaningless in the big rack world. the big change in purchasing wont be until zen3 or 4 comes out demonstrating reliability in amd execution.
but intels security is so bad he was running a r7 2700x for non critical server stuff and getting results that shamed xeons.
he also indicates smt4 is worth something to the server side.

 

[DarthKyrie]

the Moore'sLawIsDead channel did a podcast with a finance industry IT guy, they go over the economics and purchase decision process in re intel security flaws and performance.

according to the IT guy: price is functionally meaningless in the big rack world. the big change in purchasing wont be until zen3 or 4 comes out demonstrating reliability in amd execution.
but intels security is so bad he was running a r7 2700x for non critical server stuff and getting results that shamed xeons.
he also indicates smt4 is worth something to the server side.

There was a reason that IBM developed SMT3/SMT4 (Lisa Su had a hand in that) for their POWER line.

 

[darkswordsman17]

the Moore'sLawIsDead channel did a podcast with a finance industry IT guy, they go over the economics and purchase decision process in re intel security flaws and performance.

according to the IT guy: price is functionally meaningless in the big rack world. the big change in purchasing wont be until zen3 or 4 comes out demonstrating reliability in amd execution.
but intels security is so bad he was running a r7 2700x for non critical server stuff and getting results that shamed xeons.
he also indicates smt4 is worth something to the server side.

That's interesting. Doesn't surprise me much. The big players need sustained capability to make such big switches (especially if they're going to consider transitioning and optimizing software for it). Zen 1 EPYC I think piqued some interest. Zen 2 is showing AMD is not messing around and can build on Zen 1, but Zen 3 is where AMD really shows if they're for real (after all, arguably the Zen 1 to Zen 2 could say was really in line with, perhaps more impressive in some ways, less so in others, than the K8/Opteron/Athlon 64 to X2 era (where AMD doubled core counts not unlike Zen 1 to 2, which they brought big efficiency and IPC improvements compared to Pentium 4 but then that was essential for that time, but they also brought the AMD64 extensions, whereas Zen was more just righting AMD's CPU design, but K7 was more competitive than the construction cores were so AMD making up that much ground isn't anything to sneeze at, let alone considering the other things they did).

If Zen 3 brings Zen 2 like improvement (combination of IPC improvement and more multi-threading either via SMT or more cores maybe both to end up with roughly double the threading capability, or even just 50% since I don't think they could probably double it with the move from 7 to 7+), I could see a lot of companies look at how they'll be needing to replace systems or overhaul for PCIe 5.0 and DDR5 as the point they would look at buying in big. And if AMD can get their GPU division going in the meantime, I think it could also payoff big where AMD can leverage fully integrated systems.

Another possible factor is any customization will take some time, and I think Zen 2 could be when it gets those companies to say AMD's platform is worth investing in, and then it'd take some time for AMD to show they can tailor to the customer. I think AMD is sitting well to leverage both ways (the inherent improvements they're making, and their willingness to tailor for the customer; I'm getting a strong hunch they're already in deep talks about that as they talked about integrating various AI IP versus developing their own right now and then using what they learn from that on how to possibly go about developing AI focused hardware of their own to then iterate on).

The next big innovation I'm hoping to see from AMD is leveraging HBM, probably HBM3. I think it either right in the I/O chip or nearby, where it essentially is serving as a huge shared cache and even some system memory. It would reduce the dependency on memory channels (which means they could probably replace that with IF links in servers, meaning more sockets so they could increase effective cores per rack by having more CPUs instead of cramming just more cores/threads per CPU, although the HBM would enable that as well since it'd basically be shrinking that to the chip packaging level, but it'd massively reduce the length and thus latency of IF links). People are probably sick of me talking about how HBM could possibly even make sense outside of server (but HBM3 should up quantities that it would be a lot more feasible to possibly replace some DRAM in servers), so I won't espouse myself too much more there. From there, its about the NAND (higher quality stuff with lots of endurance could possibly make ditching the rest of the DRAM possible, where it'd be nearly as fast while offering larger quantity and non-volatility), then even large pool of slower NAND storage. That stuff will remove the overall system bottlenecks such that then there will be again extra emphasis on the processing (which I think will drive the co-processor development as we'll be wanting to accelerate the processing and make sure it doesn't turn into a bottleneck).

 

[tomatosummit]

The next big innovation I'm hoping to see from AMD is leveraging HBM, probably HBM3. People are probably sick of me talking about how HBM could possibly even make sense outside of server .

I think hbm memory makes the least sense on servers (for amd's current path anyway). For embedded and laptops is where I think it'll really shine, allowing the creation of really cmopact socs. Think intel compute stick sizes but with actual performance.
With zen2, amd improved their packaging techniques to stop die heights being an issue and increate contact pitch which was a problem with some hbm in vega10 packages and I suppose the zen2 ccd and IO dies had differences to overcome.

Did anything ever come of the cheap hbm2 that samsung announced? I think the reveal was along the lines of not requiring a silicon interposer. There must be a middle ground of reduced transfer rate hbm on regular organic substrate using the new packaging technologies to allow on package memory for more mainstream devices soon.

 

[Abwx]

according to the IT guy: price is functionally meaningless in the big rack world. the big change in purchasing wont be until zen3 or 4 comes out demonstrating reliability in amd execution.

So he s saying that they should demonstrate reliability and execution be it without the necessary sales to sustain the develloppement cycles..?..

And in the meantime does he advise to buy Intel s vastly inferior gear..?.

For me that sound like viral marketing and a smoked screen support to buy time for Intel, if he was really serious he would had stated that any firm concerned with such matters as security, perf/watt, TCO would be highly incompetent if it goes the Intel way currently, rest is bla bla that dont touch the important points, dunno what the 2700X has to do in such a "debate".

What s next from this "expert", that firms should wait for Zen 10...?

 

[joesiv]

I think the crux of what the IT guy was saying is that they already have intel, they know intel, flaws and all. They've vetted Intel's products for decades, they know their software stack works on it, and if it doesn't, it's a known quanity, and they have workarounds to make it work with their business. To throw that away and start with a new hardware vender takes time, a lot of time. more than months, years. They need to ensure the hardware is rock solid, and actually has compatibility with their software stack, and meets expectations in terms of performance and reliability.

I would say that ever since last year with the security issues intel has been hit with, and AMD offering somewhat competitive parts a lot of IT firms have probably started looking closer at Epic, perhaps some have them on their benches or in their racks as side projects, just to see if it's viable. Now that Rome is here, more of the hold outs are likely also starting to more seriously consider evaluating the platform.

But the evaluation will take time, you can't risk your multi-million/billion dollar business on a 2 year old line of products, so if they evaluate over the next few years, and everything works, as they should, and AMD keeps delivering on their roadmaps, more and more IT companies will start shifting over.

But obviously the Intel sales team will be aggressively trying to keep them as well, by offering deep discounts. So, if you are given a choice of adding more servers that are equivalently priced, but has 100% guaranteed compatability no risk to your farms, vs, a new system without that, the choice is not so clear cut.

I'd bet that IT companies are happy AMD is here, just to help them get cheaper prices on their already planned purchases, probably similar to a lot of the consumer market that buys Intel/Nvidia, "c'mon AMD be more competitive, not so I can buy your stuff, but so that the prices of your competition can stay in control"