Anandtech：Intel's Skylake-SP Xeon VS AMD's EPYC 7000

[Technotronic]

IMHO Anandtech's testing isn't all that valid. Who doesn't virtualize their servers now? Would it have been difficult to install Hyper-V?

Agreed. Using DDR4-2400 for the EPYC system when it is validated for DDR4-2666. If they didn't have the memory supplied then the revew should have been halted. They were even supplied DDR4-2666 RAM! They literally didn't even offer a clue as to why. Only this.

Notice that the DDR4 DRAM in the EPYC system ran at 2400 GT/s (8 channels), while the Intel system ran its DRAM at 2666 GT/s (6 channels).

Also.. they used an extremely odd dataset size to test "database" performance. They used a set that is literally able to fit in the Xeon's L3. Just a few dozen megabytes. Most databases that actually require fast server chips like this are in the gigabyte or more range. My Fortune 500 company uses a primary database that is 45 TBs! Not to mention Anandtech even admitted

According to AMD, if you enable Memory Interleaving, performance should rise a bit (+10-15%?). Unfortunately, a few days before our deadline our connection to the BMC failed, so we could not try it out

and then

A small database that can be mostly cached in the L3-cache is the worst case scenario for EPYC.

I mean.. WHAT?! They knowingly published benchmarks without a plug and play 15% performance option then even more mind blowing they use a data set that is <50mb for actual database benchmarking?! Enterprise databases are NEVER that small.. why would they use a database that fits perfectly inside a Xeon's L3?

 

[tamz_msc]

I want SPECfp numbers, especially 2017. That's all I need to know if it's going to be any good for HPC.

 

[Technotronic]

Some of the language they decided to forward from Intel is laughable. Why would they include this in a review?

Intel believes buyers are willing to pay a bit more for the vendor with the better track record.

A few other choice tidbits..

Just... Wow. Stunning. In the worst case INT heavy workload EPYC consumes exactly 7% more power. In the worst case FP workload the Xeon consumes..... 50% more power. Even at idle the Xeon consumes almost 25% more power!

AMD is withint their TDP in all instances and Intel blows by its TDP in the POV-Ray benchmark.

Wow. What just happened? It looks like a landslide victory for the raw power of the four FP pipes of Zen: the EPYC chip is no less than 50% faster than the competition.

EPYC CPU performance is pretty stunning: about 16% faster than Intel's Xeon 8176.

Again, the EPYC 7601 simply crushes the competition with 41% better performance than Intel's 28-core.

 

[Mopetar]

IMHO Anandtech's testing isn't all that valid. Who doesn't virtualize their servers now? Would it have been difficult to install Hyper-V?

No testing is ever going to be perfect and they do mention they had somewhat limited time to do their testing in order to get the article out. I don't think any of the other reviews that came out yesterday (or at least the ones that I read) would be valid under your constraints, so there isn't much to go on then.

 

[Atari2600]

I think Johan and the guys did a great job to turn around what they have based on essentially a week of access.

But, that said, it's probably more a preview, or 1st edition of a review for the professional marketplace.

I'm sure they'd love to have more time to investigate many things further and polish it up.

 

[naukkis]

. In the worst case FP workload the Xeon consumes..... 50% more power.

Worst part for Intel isn't that Xeon consumes 50% more power, for that workload it's also 16% slower than epyc. Massive disaster for Intel and SIMD-512 capable FPU, Intel need instant turn-around to switch back to sane SIMD- width or they have almost Netburst size disaster in their hands.

 

[Osjur]

Aah c'moon, we need Epyc vs Xeon review from ServeTheHome but it seems they needed to postpone it a bit.

 

[Markfw]

WOW, now I see that they crippled the memory, used a database perfect for the Xeons, and didn;t set the bios correctly, and it spanked the Xeons in most tests except the database. Thats impressive.

Edit: Oh, and at 66% of the power and less expensive.

 

[Technotronic]

WOW, now I see that they crippled the memory, used a database perfect for the Xeons, and didn;t set the bios correctly, and it spanked the Xeons in most tests except the database. Thats impressive.

Edit: Oh, and at 66% of the power and less expensive.

Well @The Stilt can maybe shed more light. He offered this response. It seems Anandtech may not have been able to get them running at DDR-2666 but it still bugs me that they didn't address it at all.

That's not the case.
Anandtech used 2DPC dual rank configuration (16x 32GB), which is rated for max. 1866MHz (RDIMM) or 2133MHz (LRDIMM).

EDIT: I completely ignored the fact that it was 2P
For 1DPC DR = 2400MHz (RDIMM), 2667MHz (LRDIMM).

@The Stilt do you care to elaborate?

 

[nwr]

I don't get it. EPYC is no Sledgehammer, not by a long stretch - but Intel makes it look like one. What the hell are they doing? Do they intend to crash the share price to prepare a massive buyback?

 

[The Stilt]

Well @The Stilt can maybe shed more light. He offered this response. It seems Anandtech may not have been able to get them running at DDR-2666 but it still bugs me that they didn't address it at all.

@The Stilt do you care to elaborate?

I originally quoted the officially supported speeds for 2DPC (2 dimms per channel) configuration.
Then I realized that AT was using 2P (two processors) configration, which means that the memory configuration was 1DPC instead of 2DPC (two EPYCs = 16 channels).
2400MHz is the maximum supported MEMCLK frequency for 1DPC registered DIMMs, while load-reduced DIMMs are supported up to 2667MHz.

So depending on which type of DIMMs was used for the review, the 2400MHz AT used might or might not be correct thing to do.

 

[Markfw]

Well, regardless of memory, the fact that the entire tiny database was sized to fit in the L3 cache of the Xeons, almost makes that test invalid. NOBODY has a 50 meg database or smaller. The smallest one where I used to work was 15 gig, and that would fit in the main memory (SGA, Oracle) of the server that was running it. And that was almost the only test it lost in !

 

[scannall]

So what I'm seeing, is Intel's response to Epyc is more segmentation and higher prices. Unless I'm missing something?

 

[moinmoin]

So what I'm seeing, is Intel's response to Epyc is more segmentation and higher prices. Unless I'm missing something?

This wasn't Intel's response, that product lineup was drawn up with Intel still in monopoly mode. "4 desktop chips glued together" is Intel's response for the time being...

 

[Topweasel]

Well, regardless of memory, the fact that the entire tiny database was sized to fit in the L3 cache of the Xeons, almost makes that test invalid. NOBODY has a 50 meg database or smaller. The smallest one where I used to work was 15 gig, and that would fit in the main memory (SGA, Oracle) of the server that was running it. And that was almost the only test it lost in !

Well its like this http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/12 memory pinned portion of the review. I get what Anandtech was trying to do, but realistically if it's not part of the competitive review enviroment (I checked the BW server review and it wasn't), then it should have been saved to be part of the actual CPU review instead of a head to head like that. It was used specifically to show how the architecture of EPYC affects memory bandwidth in regards to the separate dies. Which is great and kudos to them for gathering that info. But using it in a head to head like that as a new test made to show specifically one CPU's deficiencies isn't a great testing method. It only gets worse when you are specifically targeting a deficiency in the database test (seriously a database that is larger than 8mb but less than 50mb to fit into Intel's sweet spot who tests like that?). Other silly things like not testing larger memory tests because older hardware couldn't run it.

Why did they think AMD ran that two pc demo where the Intel server collapsed on itself. There was an active strength that they could showcase in way that would grab people's attention. That OK for Intel and AMD. That isn't ok for independent reviewer specially on a head to head. That's about comparing products to help people figure out what solution works in use cases that are similar. You shouldn't be framing the review around also being an architectural review where you have data analysis that has little to no information on actual live impact, none of those tell us how the CPU's actual perform in jobs, just how the architecture works, and worse needlessly makes something look worse for no reason.

 

[DrMrLordX]

That review is BS IMO, its only benchmarked against older Xeon processors, how lame is that !

Tom's may just not have an Epyc for review. Or they are taking more time to do the review to get it right. I hope?

They'll probably do an epyc article later. Double the articles, double the clicks!

Yeah that too.

Luckily companies like Microsoft, Google and Facebook are large enough that this tactic doesn't work so well anymore. Google has engineers who would accept EPYC configured in a sub-par way and they also will not be swayed by rebates from Intel since they are not reselling them processors.

Also bear in mind that the big boys often spec out custom configs with ODMs, down to custom mobo designs. It wouldn't surprise me if they even meddle with firmware. Hell Intel sells them custom instruction sets baked right into the silicon now. I wouldn't be surprised if AMD did the same.

WOW, now I see that they crippled the memory, used a database perfect for the Xeons, and didn;t set the bios correctly, and it spanked the Xeons in most tests except the database. Thats impressive.

Edit: Oh, and at 66% of the power and less expensive.

They sure grew beyond the Opteron brand. Epyc is . . . epic.

This wasn't Intel's response, that product lineup was drawn up with Intel still in monopoly mode. "4 desktop chips glued together" is Intel's response for the time being...

When in doubt, spread the FUD.

 

[thecoolnessrune]

I would love if a site out there still had an older server based on a Quad Opteron 6386 SE platform just to show how far AMD has come in the last 5 years. 2x 32 core / 64 thread 180W TDP CPUs vs. 4x 16 core 140W TDP CPUs. I think it would be amusing to see the benchmarks.

 

[LightningZ71]

Judging just on the notion that it has been well established that inter CCX communications are significantly affected by lower DRAM clocks, it will be very interesting to see what effect getting an EPYc 2P system with DDR4-2666 memory will have on the performance metrics. Also, it seems that AMD's EPYc may be better able to take advantage of rising memory clocks as time goes on. This means that, as ECC LP-DDR4 DRAM gets its steady speed bumps to 2933 and eventually 3200, the EPYC platform will see improving performance numbers all the way through, even if the core clocks don't scale a whole lot.

 

[Ryan Smith]

Agreed. Using DDR4-2400 for the EPYC system when it is validated for DDR4-2666. If they didn't have the memory supplied then the revew should have been halted. They were even supplied DDR4-2666 RAM!

The server was tested as-configured from AMD.

 

[DrMrLordX]

The server was tested as-configured from AMD.

Odd. Why would they use that configuration?

 

[Ryan Smith]

Odd. Why would they use that configuration?

I honestly don't know. That's something Johan will have to answer when he's back from holiday.

 

[nix_zero]

Judging just on the notion that it has been well established that inter CCX communications are significantly affected by lower DRAM clocks, it will be very interesting to see what effect getting an EPYc 2P system with DDR4-2666 memory will have on the performance metrics.

not just inter ccx, its still unknown if the GMI controller is on the same clock domain afaik so it could impact even inter die communications - an indeep analysis would be nice.

 

[lobz]

The server was tested as-configured from AMD.

the sql database size was also suggedted by AMD, I suppose? Oh wait a minute....

Sent from my VTR-L09 using Tapatalk

 

[DrMrLordX]

I honestly don't know. That's something Johan will have to answer when he's back from holiday.

Okay. When will that be? Oh, and why the um . . . small database size? The review basically acknowledges that such a small database is going to lead to skewed results.

 

[beginner99]

Looks like Intel spent quite some time to disparage AMD's approach in its Xeon presentation (around 20 slides focus on that), so the spokesperson previously dissing Epyc as "stitching together 4 desktop dies" is apparently company PR policy now.
http://www.barrons.com/articles/amd-reveals-epyc-details-intel-vows-to-top-it-1497997334

Thanks for the laughs, Intel. They don't seem to be as confident in their own products anymore to have to go this low.

Yeah those slides are just sad. It's also dumb. Remember, there is no bad publicity. Most people not already knowing better by looking at the slides will be like "Oh, there is a competitor now, interesting. Lets look at their offering and pricing because this intel stuff is so expensive." Any marketing guy here? Isn't it plain stupid to mention your competitor on your slide? That's free commercial for your only competitor, Intel. If a product sucks, it discredits itself so such slides just show intel is in panic stupidity mode or the people in power are some trolls just their for well enjoying to troll.

No wonder Intel is crashing given this type of people and politics seem to have the power within the company.

I mean.. WHAT?! They knowingly published benchmarks without a plug and play 15% performance option then even more mind blowing they use a data set that is <50mb for actual database benchmarking?! Enterprise databases are NEVER that small.. why would they use a database that fits perfectly inside a Xeon's L3?

Yeah. If that DB is that small it's retarded. Maybe such DBs do exist but they will run just fine on the controller on your smartwatch so no need to benchmark it with server CPUs.