Believe it or not - for the older guys here the more well known Linus is Linus Torvalds, the creater of Linux (after whom it is named), not a YouTuber. For those that freudian typo isn't quite that hilarious.
Typically desktop/workstation chips have higher clocks and less cores than the pure servers. But I'd be really surprised if the threadripper isn't significantly cheaper than the Epyc. Keep in mind a thread ripper has half the silicon dies, half the memory channels, and half the pci-e of the Epyc.
but where is room for the ryzen cpus then? when a 16 core server CPU cost only 899$ and TR is significantly cheaper. and if im not wrong there is at least a 12 core TR model too.
Plenty of room between $470 and $900 for Threadripper parts, and plenty of room below $470 for us regular joes who can't afford dropping a grand on just the CPU.
From my experience even highly threaded application get bottlenecked due to max single core performance, as the code path will have a single thread part in between the multithread path which becomes a huge bottle neck in these high core count cpus. Intel has a huge advantage here in terms of max turbo for the newer xeons all reaching 4.2 Ghz. But the 1p 32 core for 2000$ right away makes the v4 Xeons obsolete in terms of price to performance.
Dual socket SKUs 32-core CPUs starting at $3400, 24-core from $1850, 16-core from $650, 8-core from $475 for single socket the 32 cores at 2100$, 24 cores 1075$, 16 cores 750$.
Server margins are high so no reason for AMD to aim higher that that but they could be more aggressive in consumer.
All they need to be is cheaper per watt than intel for the same or better performance and they'll have massive sales to the cloud companies not even including taking any of the SB market. In other words they don't need to be half the price of Intel. But the hope is that Intel will lower prices on Xeon's and AMD will be forced to lower prices some more.
Frankly server part pricing is atrocious right now, a little competition from AMD could drive server part pricing down to something reasonable like the last time AMD competed with Opteron.
I'm really excited with Epyc. I remember Interlagos being released with performance well bellow existing Intel's offerings and now look at this! I can't wait for concrete benchmarks.
"Each CPU will support 128 PCIe 3.0 lanes, suitable for six GPUs with full bandwidth support (plus IO) or up to 32 NVMe drives for storage. " Shouldn't this be 8 GPUs or 32 NVMe drives? (Or 7/31 if a Southbridge is connected and eats 4 of the lanes.)
Well AMD is comparing a benchmark compiled with gcc-6.2 and running on Intel vs the same benchmark comiled with gcc-6.2 and running on AMD. For people who compile their own binaries with gcc this is quite fair. However intel's compiler is sometimes substantially faster than gcc, question is are the binaries you care about (Games? Handbrake? Something else?) compiled with intel's compiler or gcc?
you could be right but it reads on tomshardware as if they just take the numbers provided by intel and reduce them.. they actually don´t test on intel. the just take numbers from intel and reduce them by 46%.
From what I read they tested the flagship Xeon, found the performance multiplier (~0.57), then extrapolated them across the rest of the range. So not completely scummy, but still cause to wait for actual benchmarks.
Intel "cheats" in the icc compiler when compiling SPEC workloads. Libquantum is most notorious for such cheating but many others are also prone to this issue. In Libq, the icc compiler basically reorganize the datastructures and memory layout to get excellent vectorization, and I think 1/10th the bandwidth requirements as compared to gcc -O2. These transformations are there only for libquantum so it has little to no use for general workloads. Hence gcc (and llvm) reject such transformations. It's not unlike VW's emission defeat devices, actually.
Go to Ars Technica. They have the full dump of slides. AMD does benchmark Xeon 1 or 2 systems themselves. However for some of the graphs where Xeon data is presented, they use Intel's published numbers (on icc of course), and derate it by this factor to account for this cheating. You could argue that AMD should have benched all the systems themselves and that's fair enough. But I don't think Tom's hardware is exactly qualified to know or state that the derate is 20% and not 40% or whatever. They benchmark consumer hardware, and wouldn't know a thing about this. So any number coming from these sources are dubious.
And Intel's ability to optimise exceeding LLVM's and GCC's is certainly not cheating. Perhaps Google, Apple, Microsoft, and GNU could catch up instead of bitching.
"And Intel's ability to optimise exceeding LLVM's and GCC's" Complete and utter strawman. Nobody's questioning that icc is a good compiler. Plenty in HPC pay good money for icc. It's not nearly as good as SPEC suggests though because of stuff that is pretty much if (signature of libq, mcf, etc detected) { special codepath that nobody else uses/needs }. At issue is not whether icc is a better compiler but whether icc is also a fair compiler to use on benchmarks like SPEC. It's not, thanks to Intel's shenanigans with the compiler. The reason llvm and gcc don't have this is because they're not interested in selling chips (which intel is); they're interested in improving the average workload. Special casing libquantum doesn't align with those priorities.
So they've only been cheating in the decade or so before, and they've been saints for 3 years, except they cheat at libquantum to inflate their SPEC scores, purely on merit. Yeah.
This is a server chip, likely not as nice to use as a HEDT targeted chip like threadripper. The threadripper is likely to have fewer cores, faster clocks, and half the memory bandwidth... but more cores, and double the bandwidth of the Ryzen.
If you have video work, CAD, or MATLAB related things to do then the extra cores, memory bandwidth, and depending on how much you're dumping on co processors, even the PCIe lanes would be helpful.
AMD provided some basic benchmarks, seen in the slides above, that compare its processors to the nearest Intel comparables. The price and performance breakdown chart is perhaps the most interesting, as it indicates much higher performance (as measured by SPECint_rate_base2006), at every price point. It bears mentioning that Intel publicly posts its SPEC benchmark data, and AMD's endnotes indicates that it reduced the scores used for these calculations by 46%. AMD justified this adjustment because they feel the Intel C++ compiler provides an unfair advantage in the benchmark. There is a notable advantage to the compiler, but most predict it is in the 20% range, so AMD's adjustments appear aggressive. We should take these price and performance comparisons with the necessary skepticism and instead rely upon third-party data as it emerges ......
That seems like a stupid reason to reduce anything. At the end of the day, what matters is how fast shit runs, if software is more optimized for one platform that is a valid point of data to include into any conclusion.
I'm waiting to see consumer sites benchmark a server CPU against retail CPUs and then crow about clocks, etc. ;) it'll be done, it'll be vicious, and people will take it as the gospel truth. heck, let's just get the Cinebench testing done ASAP and call it day ;)
I am interested in Ian's take but I test this hardware on my own using the toolsets available to me. While I appreciate Johan's insights, I find most of the consumer sites (as Anandtech is one of them) to be reaching when they try to provide realistic workload testing. StorageReview.com does a decent job (mostly), but I'm finding that most reviews, sadly, are hit and miss against test benches and their applicability is...dubious at best.
and like LurkingSince97 infers, spec-int is a fanboy benchmark ;)
Plenty of people are interested in it, and those people do their own benchmarking. They don't visit Anandtech or Tom's hardware to get this information.
Except that nobody runs spec-int on their servers.
AMD made two mistakes:
1. spec-int should not be used to compare servers across architectures. Instead, run real software that people use on servers. Virtualization benchmarks, JVM stuff, databases, whatever. Real world things.
2. Trying to modify spec-int results (I guess, by using GCC instead of intel's compiler, and compensating for the stuff their compiler does). Yeah, a lot of the tricks that some compilers use on spec-int are absolutely garbage and would not make real-world applications faster -- just spec-int. But there is no objective way to disentangle that. So stay away from it.
That's because intel cheats on SPEC in icc by doing transformations that are specifically targetted at making SPEC faster and nothing else. Libquantum is a particularly egregious example where you nearly double the performance by doing tricks that help nothing else. But this is generally true across the suite. It's not unlike VW's emission defeat devices: do something special when you're being tested.
As for Tom's hardware, they're not authorities on anything server. What they know something about is gaming benchmarking, and that's pretty much it. I don't expect he'd know a thing about it, and whether 20% is correct vs 40%. It's a feeble attempt at sounding clever. The people buying this stuff know what they're doing, and aren't going to be influenced by some online reviewer.
I think you missed the point of the eight-core processor. That's for GPU compute servers, where you want the cheapest processor possible with the most PCIe lanes. It's probably going to be the one that sells the most, because Intel has nothing comparable.
probably not. Miners want the most GPU (hashes) per Watt (combined with total price). If they can do that with 5 smaller, cheaper machines vs 1 larger one, they will. Mining does not need coordination across multiple GPUs.
The enterprisey compute stuff -- machine learning being a huge one -- often _does_ need to coordinate across GPUs in one big data set and will run in datacenters where consolidation into performance/$ and performance/Watt will often like servers with few CPU and many GPU, with a ton of I/O and connectivity to other servers.
Mining doesn't care about I/O, just total # of ports. People even use tools to split up a x16 bus of a normal consumer motherboard into may smaller PCIe ports each with a GPU on it. The GPU will compute hashes just as well with a x2 port as a x16 one.
EPYC has 128 PCI-e lanes on both 1 socket and 2 socket systems, so if AMD had intended the EPYC 7251 to be used for GPU compute servers, they would have made it part of the single socket lineup. That doesn't mean that the chip won't be used in GPU compute servers; it just means that GPU compute servers are not the market that AMD intended to target with the chip.
All these 2P Epyc CPUs should be just fine in 1P as well. Obviously nobody will buy most of them to run in 1P, as 1P are cheaper. It is just 1P that is limited - it can be *only* used in 1P. And given 500-ish price of the chip, I can see why AMD didn't bother to give additional 400 chip for 1P - it wouldn't change anything. And limiting this chip to just 1P would be pointless, as the other segments such as "need tons of memory" would be hurting for no good reason.
you can flatten the NUMA domains via BIOS hooks to just 1 or 2 per processor but there is a relative hit to latency. this is the "downside" to a MCM-based processor vs. ring/monolithic mesh. however, the upside IS agility based on chip packaging and time to market so....there's always a cost/benefit ratio to consider.
Lol what a shady move from AMD to reduce Intel CPUs' benchmark numbers in order to make Epyc appear better than it actually is, never change AMD never change.
First there's a big difference between straight-out misleading customers and making backside deals with OEMs, and second that compiler crippling stuff is still unsubstantiated and Intel has no obligation towards AMD with regards to Intel's own compiler. AMD should make their own compiler that offers better or at least equal to Intel's own optimizations instead of using disgraceful tactics like that.
Um.... Where to begin? The fine doesn't go to AMD. It goes to the European Commission.... So far Intel's lawyers have held the EC at bay all these years. Which technically means Intel got away with it....
What I don't understand from the slide with the prices: it looks like the 1P cpu is priced higher ($750 versus $650) than the 2P counterpart? I assume that any 2P cpu could be used in a 1P motherboard, but not the other way around.
Well, the corresponding 2P part is >1.1k, so 1P is cheaper. No idea why there isn't 7301P instead and slightly cheaper than the bottom 2P, but I guess that 7351P looks better on the 2P vs 1P.
Could just be that the diagram is nonsense marketing bling, but that sure looks like external lanes are connected to only two of the 4 cores, with the remaining two getting 'passthrough' lanes.
The word 'epyc' (If it is a word) makes me want to cry when thinking about something as serious as servers. Gaming, yes it would work, but not servers.
I guess it might be a hard sell for bosses, when they read AMD EPYC on the invoice, but at the end of the day, if the performance is there and the price is right, the CPU could be named "Momas Big Belly Filling Pie" and I would not care.
"All 2P E5 scores were derived from the following ICC compiler-based test results per spec.org, multiplied by 0.575 to convert them from the ICC compiler to GCC -O2 v6.1..."?!?
precisely! try another behavioural pattern: great claims, feeble results :) to compare E5-2698 v4 20 cores with EPYC 7551 32 cores - 32/20 = 1.6, the extrapolation is claimed to be +44% or 1.44. therefore 1.6/1.44=1.1(1). now we can to conclude that for a 32 core xeon (it's easy for intel) the result will be +11% for intel, and no more advantage for amd at all!!! such mathematical practices are ridiculous!
Anyone buying these is going to want to see some large independent reviews / studies. Some of the larger companies looking at these say like Amazon's data center may get a few and do a study. I think that type of info is what everyone should wait on. These numbers look like they came from the marketing side of the house.
"Anyone buying these is going to want to see some large independent reviews / studies. Some of the larger companies looking at these say like Amazon's data center may get a few and do a study."
Those companies that shared the stage with AMD for this presentation were included in the 5,000 Epyc chips that were given to OEMs/ODMs to test and validate over the last six months.
All the big boys know what EPYC is capable of and it seems that most are quite impressed.
Yeah well the story is that they obtained GCC 6.1 scores (by benching the CPU) for the top CPU and found out those are ~40% lower than the official ones. Some of that is ICC cheating on tests, some of that is optimization level. So, they reduced all official scores by the same ~40% for this comparison (with the top part being actually benched and achieved that result as on slides).
I can't fault them for using the middle ground GCC for those benchmarks, and the normalization step is reasonable enough as well. The only real issue here is use of -O2 instead of more optimized code. Sure, Ryzen/Epyc is new and GCC likely cannot optimize as good for it (which is probably they reason they used default O2), but this is not a valid excuse, should have used O3 at least if not advanced flags.
Erm. I use GCC with -O3 in a library I build at my job. With profile feedback it is easily 20% faster than -O2. So I don't know where these "clue" people are. I've only been coding for 20 years or so. I may not have a clue yet.
I believe that people were not aware how overpriced Intel products were. We keep seeing that now from the day Ryzen 7 launched, which is enough for me to never again buy an Intel CPU as long as AMD is at least a bit competitive. It's one thing to make a good CPU and price it accordingly but it's completely different thing to rip off the ones who help you earn money and pay for your employees just because you can.
"PCIe can be bifurcated all the way down to sets of x1, although there will be a limit of 8 PCIe devices per x16 link, but this becomes a mix and match game: x8 + x4 + x2 + x1 + x1 + x1 + x1 becomes entirely valid for a single x16 link"
Why are the clock speeds so low? Even the eight-core model only runs at 2.1 GHz, whereas the desktop eight-core Ryzen model runs at 3.6 GHz. Intel Xeons are the same way. I don't understand why x86 server CPUs always run at a lower clock rate than their desktop counterparts.
> "... One element that was not discussed is live VM migration across encrypted environments. We fully suspect that an AMD-to-AMD live migration be feasible, although an ...".
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
131 Comments
Back to Article
yuhong - Tuesday, June 20, 2017 - link
"Linus kernel"IanHagen - Tuesday, June 20, 2017 - link
Linux Tech Tips?Krysto - Tuesday, June 20, 2017 - link
That Linux guy makes some great videos.thomasg - Tuesday, June 27, 2017 - link
Believe it or not - for the older guys here the more well known Linus is Linus Torvalds, the creater of Linux (after whom it is named), not a YouTuber.For those that freudian typo isn't quite that hilarious.
Guspaz - Tuesday, June 20, 2017 - link
sudo modprobe socks_sandalsHollyJordan - Wednesday, June 21, 2017 - link
Do you have a pay_pal ? because you can generate an extra 1000 /week in your earnings only working at home for five hours a day... check.philehidiot - Tuesday, June 27, 2017 - link
Ooooh, I have Pay Pal and I'd love to earn 1000 Tasmanian Pesos a week for 5 hours a day from the comfort of my own home.I can imagine that would work out as all of $4/month.
OddFriendship8989 - Wednesday, June 28, 2017 - link
How about 1000 Bitcoins?satai - Tuesday, June 20, 2017 - link
1 socket 16 / 32 @ 2.9GHz max for $700+... it looks like 16 core Threadripper with reasonable frequencies for less then $999 looks reasonable.spikebike - Tuesday, June 20, 2017 - link
Typically desktop/workstation chips have higher clocks and less cores than the pure servers. But I'd be really surprised if the threadripper isn't significantly cheaper than the Epyc. Keep in mind a thread ripper has half the silicon dies, half the memory channels, and half the pci-e of the Epyc.Gothmoth - Tuesday, June 20, 2017 - link
but where is room for the ryzen cpus then? when a 16 core server CPU cost only 899$ and TR is significantly cheaper. and if im not wrong there is at least a 12 core TR model too.Bateluer - Tuesday, June 20, 2017 - link
Plenty of room between $470 and $900 for Threadripper parts, and plenty of room below $470 for us regular joes who can't afford dropping a grand on just the CPU.Gothmoth - Tuesday, June 20, 2017 - link
"significantly lower".and the cheapest epyc server is already a lot less than 900$ ... when we look at the information above.
Zingam - Wednesday, June 21, 2017 - link
I'm pretty sure MoBos will be more expensive without consumer features and especially no "Gaming" on the cardbox and no RGB lighting.Jimster480 - Monday, June 26, 2017 - link
I agree but there is alot of room even between $200 and $50 for CPU's lol.Most people don't need an 1800x, nevermind anything above that.
The average person only needs 4 cores, and most gamers will do fine with a 6C ryzen or the entry level 8C.
sharath.naik - Monday, July 3, 2017 - link
From my experience even highly threaded application get bottlenecked due to max single core performance, as the code path will have a single thread part in between the multithread path which becomes a huge bottle neck in these high core count cpus. Intel has a huge advantage here in terms of max turbo for the newer xeons all reaching 4.2 Ghz. But the 1p 32 core for 2000$ right away makes the v4 Xeons obsolete in terms of price to performance.jjj - Tuesday, June 20, 2017 - link
Dual socket SKUs 32-core CPUs starting at $3400, 24-core from $1850, 16-core from $650, 8-core from $475for single socket the 32 cores at 2100$, 24 cores 1075$, 16 cores 750$.
Server margins are high so no reason for AMD to aim higher that that but they could be more aggressive in consumer.
rahvin - Wednesday, June 21, 2017 - link
All they need to be is cheaper per watt than intel for the same or better performance and they'll have massive sales to the cloud companies not even including taking any of the SB market. In other words they don't need to be half the price of Intel. But the hope is that Intel will lower prices on Xeon's and AMD will be forced to lower prices some more.Frankly server part pricing is atrocious right now, a little competition from AMD could drive server part pricing down to something reasonable like the last time AMD competed with Opteron.
IanHagen - Tuesday, June 20, 2017 - link
I'm really excited with Epyc. I remember Interlagos being released with performance well bellow existing Intel's offerings and now look at this! I can't wait for concrete benchmarks.DanNeely - Tuesday, June 20, 2017 - link
"Each CPU will support 128 PCIe 3.0 lanes, suitable for six GPUs with full bandwidth support (plus IO) or up to 32 NVMe drives for storage. " Shouldn't this be 8 GPUs or 32 NVMe drives? (Or 7/31 if a Southbridge is connected and eats 4 of the lanes.)Gothmoth - Tuesday, June 20, 2017 - link
i read AMD reduced the benchmark numbers for intel by 46% because of compiler benefits for intel...can someone look at the fineprint and confirm or rebunk this???
spikebike - Tuesday, June 20, 2017 - link
Well AMD is comparing a benchmark compiled with gcc-6.2 and running on Intel vs the same benchmark comiled with gcc-6.2 and running on AMD. For people who compile their own binaries with gcc this is quite fair. However intel's compiler is sometimes substantially faster than gcc, question is are the binaries you care about (Games? Handbrake? Something else?) compiled with intel's compiler or gcc?Gothmoth - Tuesday, June 20, 2017 - link
you could be right but it reads on tomshardware as if they just take the numbers provided by intel and reduce them.. they actually don´t test on intel. the just take numbers from intel and reduce them by 46%.hamoboy - Tuesday, June 20, 2017 - link
From what I read they tested the flagship Xeon, found the performance multiplier (~0.57), then extrapolated them across the rest of the range. So not completely scummy, but still cause to wait for actual benchmarks.TC2 - Wednesday, June 21, 2017 - link
according to the numbersE5-2698 v4 / EPYC 7551 ~~ 1.11
all this looks quite misleading! but this is amd :)
TC2 - Wednesday, June 21, 2017 - link
1.11 per core i mean to sayRyan Smith - Tuesday, June 20, 2017 - link
We still need to independently confirm the multiplier, but yes, AMD is reducing Intel's official SPEC scores."Scores for these E5 processors extrapolated from test results published at www.spec.org, applying a conversion multiplier to each published score"
davegraham - Tuesday, June 20, 2017 - link
Jeff @ Techreport has the multiplier officially listed.deltaFx2 - Wednesday, June 21, 2017 - link
Intel "cheats" in the icc compiler when compiling SPEC workloads. Libquantum is most notorious for such cheating but many others are also prone to this issue. In Libq, the icc compiler basically reorganize the datastructures and memory layout to get excellent vectorization, and I think 1/10th the bandwidth requirements as compared to gcc -O2. These transformations are there only for libquantum so it has little to no use for general workloads. Hence gcc (and llvm) reject such transformations. It's not unlike VW's emission defeat devices, actually.Go to Ars Technica. They have the full dump of slides. AMD does benchmark Xeon 1 or 2 systems themselves. However for some of the graphs where Xeon data is presented, they use Intel's published numbers (on icc of course), and derate it by this factor to account for this cheating. You could argue that AMD should have benched all the systems themselves and that's fair enough. But I don't think Tom's hardware is exactly qualified to know or state that the derate is 20% and not 40% or whatever. They benchmark consumer hardware, and wouldn't know a thing about this. So any number coming from these sources are dubious.
patrickjp93 - Wednesday, June 21, 2017 - link
Wrong. See the CPPCon 2015/2016 "Compiler Switches" talks. ICC does not CHEAT at all. It hasn't since 2014.Intel wins purely on merit.
patrickjp93 - Wednesday, June 21, 2017 - link
And Intel's ability to optimise exceeding LLVM's and GCC's is certainly not cheating. Perhaps Google, Apple, Microsoft, and GNU could catch up instead of bitching.deltaFx2 - Thursday, June 22, 2017 - link
"And Intel's ability to optimise exceeding LLVM's and GCC's" Complete and utter strawman. Nobody's questioning that icc is a good compiler. Plenty in HPC pay good money for icc. It's not nearly as good as SPEC suggests though because of stuff that is pretty much if (signature of libq, mcf, etc detected) { special codepath that nobody else uses/needs }. At issue is not whether icc is a better compiler but whether icc is also a fair compiler to use on benchmarks like SPEC. It's not, thanks to Intel's shenanigans with the compiler. The reason llvm and gcc don't have this is because they're not interested in selling chips (which intel is); they're interested in improving the average workload. Special casing libquantum doesn't align with those priorities.Luckz - Wednesday, June 21, 2017 - link
So they've only been cheating in the decade or so before, and they've been saints for 3 years, except they cheat at libquantum to inflate their SPEC scores, purely on merit. Yeah.willis936 - Tuesday, June 20, 2017 - link
Just wow. Finally a chip worthy of the "HEDT" moniker.nevcairiel - Tuesday, June 20, 2017 - link
Except its not anything like that. Its a server chip, not any end of the desktop/workstation spectrum.Gothmoth - Tuesday, June 20, 2017 - link
well at work i have a two socket system under my desk that uses server CPU´s.vanilla_gorilla - Tuesday, June 20, 2017 - link
>well at workYou mean like a work ... station with a workstation class CPU?
Drumsticks - Wednesday, June 21, 2017 - link
Probably, meaning exactly the kind of chip that nevcairel (who gothmoth replied to) implied that it wasn't.spikebike - Tuesday, June 20, 2017 - link
This is a server chip, likely not as nice to use as a HEDT targeted chip like threadripper. The threadripper is likely to have fewer cores, faster clocks, and half the memory bandwidth... but more cores, and double the bandwidth of the Ryzen.vision33r - Wednesday, June 21, 2017 - link
Typical novice user response that more is better.willis936 - Wednesday, June 21, 2017 - link
If you have video work, CAD, or MATLAB related things to do then the extra cores, memory bandwidth, and depending on how much you're dumping on co processors, even the PCIe lanes would be helpful.Drumsticks - Tuesday, June 20, 2017 - link
...Wow. THAT is an impressive comeback.msroadkill612 - Wednesday, June 21, 2017 - link
AMD have gone from love forty, to advantage server.willis936 - Wednesday, June 21, 2017 - link
How many wallets will be aced?Gothmoth - Tuesday, June 20, 2017 - link
can someone look at the endnotes and tell me if the intel benchmark reduction is true or not....Gothmoth - Tuesday, June 20, 2017 - link
i can´t do it myself i am on my phone with shitty connection....fanofanand - Tuesday, June 20, 2017 - link
I cannot find a single statement confirming your concerns. I'm not sure where you heard it, but it appears to be inaccurate.Gothmoth - Tuesday, June 20, 2017 - link
tomshardware:AMD provided some basic benchmarks, seen in the slides above, that compare its processors to the nearest Intel comparables. The price and performance breakdown chart is perhaps the most interesting, as it indicates much higher performance (as measured by SPECint_rate_base2006), at every price point. It bears mentioning that Intel publicly posts its SPEC benchmark data, and AMD's endnotes indicates that it reduced the scores used for these calculations by 46%. AMD justified this adjustment because they feel the Intel C++ compiler provides an unfair advantage in the benchmark. There is a notable advantage to the compiler, but most predict it is in the 20% range, so AMD's adjustments appear aggressive. We should take these price and performance comparisons with the necessary skepticism and instead rely upon third-party data as it emerges
......
nevcairiel - Tuesday, June 20, 2017 - link
That seems like a stupid reason to reduce anything. At the end of the day, what matters is how fast shit runs, if software is more optimized for one platform that is a valid point of data to include into any conclusion.Gothmoth - Tuesday, June 20, 2017 - link
except when you want to sell your product. :-)well i am waiting for third party benchmarks and real reviews.
davegraham - Tuesday, June 20, 2017 - link
I'm waiting to see consumer sites benchmark a server CPU against retail CPUs and then crow about clocks, etc. ;) it'll be done, it'll be vicious, and people will take it as the gospel truth. heck, let's just get the Cinebench testing done ASAP and call it day ;)Gothmoth - Tuesday, June 20, 2017 - link
yeah why is anandtech reporting about server hardware. nobody is interested in that.just let us take AMDs numbers as gospel.... a interpolate threadripper numbers until august.
SkiBum1207 - Tuesday, June 20, 2017 - link
Excuse me? There are us who use servers to make money - we definitely care about Anandtech's analysis of enterprise hardware.davegraham - Tuesday, June 20, 2017 - link
I am interested in Ian's take but I test this hardware on my own using the toolsets available to me. While I appreciate Johan's insights, I find most of the consumer sites (as Anandtech is one of them) to be reaching when they try to provide realistic workload testing. StorageReview.com does a decent job (mostly), but I'm finding that most reviews, sadly, are hit and miss against test benches and their applicability is...dubious at best.and like LurkingSince97 infers, spec-int is a fanboy benchmark ;)
D
at80eighty - Tuesday, June 20, 2017 - link
did you discover the site yesterday?davegraham - Tuesday, June 20, 2017 - link
lol. who, me? nope. ;)at80eighty - Wednesday, June 21, 2017 - link
sorry, the comment threading is not helpful - i directed that question to Gothmoth's absurd postdeltaFx2 - Wednesday, June 21, 2017 - link
Plenty of people are interested in it, and those people do their own benchmarking. They don't visit Anandtech or Tom's hardware to get this information.LurkingSince97 - Tuesday, June 20, 2017 - link
Except that nobody runs spec-int on their servers.AMD made two mistakes:
1. spec-int should not be used to compare servers across architectures. Instead, run real software that people use on servers. Virtualization benchmarks, JVM stuff, databases, whatever. Real world things.
2. Trying to modify spec-int results (I guess, by using GCC instead of intel's compiler, and compensating for the stuff their compiler does). Yeah, a lot of the tricks that some compilers use on spec-int are absolutely garbage and would not make real-world applications faster -- just spec-int. But there is no objective way to disentangle that. So stay away from it.
IanHagen - Tuesday, June 20, 2017 - link
Intel's compiler cripples code on AMD and VIA chipsAnti-competitive at the machine code level
https://www.theinquirer.net/inquirer/news/1567108/...
Intel finally agrees to pay $15 to Pentium 4 owners over AMD Athlon benchmarking shenanigans
https://www.extremetech.com/computing/193480-intel...
FTC Settles Charges of Anticompetitive Conduct Against Intel
https://www.ftc.gov/news-events/press-releases/201...
deltaFx2 - Wednesday, June 21, 2017 - link
That's because intel cheats on SPEC in icc by doing transformations that are specifically targetted at making SPEC faster and nothing else. Libquantum is a particularly egregious example where you nearly double the performance by doing tricks that help nothing else. But this is generally true across the suite. It's not unlike VW's emission defeat devices: do something special when you're being tested.As for Tom's hardware, they're not authorities on anything server. What they know something about is gaming benchmarking, and that's pretty much it. I don't expect he'd know a thing about it, and whether 20% is correct vs 40%. It's a feeble attempt at sounding clever. The people buying this stuff know what they're doing, and aren't going to be influenced by some online reviewer.
Ryan Smith - Tuesday, June 20, 2017 - link
I've posted galleries of the full slide decks. The slide you're interested in is: http://images.anandtech.com/galleries/5699/epyc_te..."Scores for these E5 processors extrapolated from test results published at www.spec.org, applying a conversion multiplier to each published score"
patrickjp93 - Wednesday, June 21, 2017 - link
No, it vastly underestimates and undermines Intel's real-world performance.lefty2 - Tuesday, June 20, 2017 - link
I think you missed the point of the eight-core processor. That's for GPU compute servers, where you want the cheapest processor possible with the most PCIe lanes. It's probably going to be the one that sells the most, because Intel has nothing comparable.Luckz - Tuesday, June 20, 2017 - link
Is this useful for the mining craze?LurkingSince97 - Tuesday, June 20, 2017 - link
probably not. Miners want the most GPU (hashes) per Watt (combined with total price). If they can do that with 5 smaller, cheaper machines vs 1 larger one, they will. Mining does not need coordination across multiple GPUs.The enterprisey compute stuff -- machine learning being a huge one -- often _does_ need to coordinate across GPUs in one big data set and will run in datacenters where consolidation into performance/$ and performance/Watt will often like servers with few CPU and many GPU, with a ton of I/O and connectivity to other servers.
Mining doesn't care about I/O, just total # of ports. People even use tools to split up a x16 bus of a normal consumer motherboard into may smaller PCIe ports each with a GPU on it. The GPU will compute hashes just as well with a x2 port as a x16 one.
KAlmquist - Wednesday, June 21, 2017 - link
EPYC has 128 PCI-e lanes on both 1 socket and 2 socket systems, so if AMD had intended the EPYC 7251 to be used for GPU compute servers, they would have made it part of the single socket lineup. That doesn't mean that the chip won't be used in GPU compute servers; it just means that GPU compute servers are not the market that AMD intended to target with the chip.Zizy - Thursday, June 22, 2017 - link
All these 2P Epyc CPUs should be just fine in 1P as well. Obviously nobody will buy most of them to run in 1P, as 1P are cheaper. It is just 1P that is limited - it can be *only* used in 1P.And given 500-ish price of the chip, I can see why AMD didn't bother to give additional 400 chip for 1P - it wouldn't change anything. And limiting this chip to just 1P would be pointless, as the other segments such as "need tons of memory" would be hurting for no good reason.
armtec - Tuesday, June 20, 2017 - link
NUMA NUMA IEI... I hadn't previously made this connection but now I will every time I need to think about server configs...Barilla - Tuesday, June 20, 2017 - link
I read that header and now that stupid song will be stuck in my head for days...SodaAnt - Tuesday, June 20, 2017 - link
pJ per bit is a very interesting measurement to use, is this the normal way to measure interconnect power consumption?Ian Cutress - Tuesday, June 20, 2017 - link
Yes :)MrSpadge - Tuesday, June 20, 2017 - link
Yes, unless you're looking at experimental optical interconnects which can reach the fJ/bit range :)ajc9988 - Tuesday, June 20, 2017 - link
https://www.pcper.com/reviews/Processors/Ryzen-5-R...The latency at 2666MHz should be around 120ns, not 140ns, which they got from 2133MHz memory. At 2400MHz, they found 129.62ns.
Gothmoth - Tuesday, June 20, 2017 - link
and you think you can translate that 1:1?to a cpu that has way more security features etc.
cheshirster - Sunday, June 25, 2017 - link
Yes, IF runs at memory speed.All the bandwith and latency numbers are scaling with memory.
ikeke1 - Tuesday, June 20, 2017 - link
https://www.servethehome.com/amd-epyc-7601-dual-so...Crazy numbers.
davegraham - Tuesday, June 20, 2017 - link
it's B2 stepping silicon running ~200-400MHz under the final clocks, fyi.Luckz - Tuesday, June 20, 2017 - link
So with Epyc they made the CCXes NUMA nodes in Windows - https://www.servethehome.com/wp-content/uploads/20...davegraham - Tuesday, June 20, 2017 - link
you can flatten the NUMA domains via BIOS hooks to just 1 or 2 per processor but there is a relative hit to latency. this is the "downside" to a MCM-based processor vs. ring/monolithic mesh. however, the upside IS agility based on chip packaging and time to market so....there's always a cost/benefit ratio to consider.vladx - Tuesday, June 20, 2017 - link
Lol what a shady move from AMD to reduce Intel CPUs' benchmark numbers in order to make Epyc appear better than it actually is, never change AMD never change.tamalero - Tuesday, June 20, 2017 - link
COUGH COUGH COUGH Yeah, because Intel never has done the same.. COUGH COUGH COUGH..https://www.extremetech.com/computing/193480-intel...
https://www.theinquirer.net/inquirer/news/1567108/...
vladx - Tuesday, June 20, 2017 - link
First there's a big difference between straight-out misleading customers and making backside deals with OEMs, and second that compiler crippling stuff is still unsubstantiated and Intel has no obligation towards AMD with regards to Intel's own compiler. AMD should make their own compiler that offers better or at least equal to Intel's own optimizations instead of using disgraceful tactics like that.galahad05 - Wednesday, June 21, 2017 - link
How's Intel doing fighting that enormous fine the EU levied against it for their underhanded tactics against AMD years ago?vladx - Wednesday, June 21, 2017 - link
Afaik they paid billions which AMD squandered like it was nothing.galahad05 - Wednesday, June 21, 2017 - link
Um.... Where to begin?The fine doesn't go to AMD. It goes to the European Commission....
So far Intel's lawyers have held the EC at bay all these years. Which technically means Intel got away with it....
Such is life.
Mugur - Wednesday, June 21, 2017 - link
What I don't understand from the slide with the prices: it looks like the 1P cpu is priced higher ($750 versus $650) than the 2P counterpart? I assume that any 2P cpu could be used in a 1P motherboard, but not the other way around.Zizy - Wednesday, June 21, 2017 - link
Well, the corresponding 2P part is >1.1k, so 1P is cheaper. No idea why there isn't 7301P instead and slightly cheaper than the bottom 2P, but I guess that 7351P looks better on the 2P vs 1P.1008anan - Wednesday, June 21, 2017 - link
Trying to calculate how many 32 bit floating point operations (FPO) a zen server completes per second:Assume a 2 socket Zen server with two 32 core chips; operating at 2.5 gigahertz:
512 bits wide vector, Fused Multiply Add, two FPO per clock = 64 FPO per clock = 512/32 * 2 * 2.
64 FPO/clock * 32 cores = 2048 FPO/clock
2048 FPO/clock * 2 sockets = 4096 FPO/clock
4096 FPO/clock * 2.5 gigahertz = 10 trillion FPO/second = 10 teraflops
Is this accurate? Is Zen approximately the same number of FLOPS as Skylake E5/E7?
edzieba - Wednesday, June 21, 2017 - link
An interesting diagram lurking on the corner of this slide: http://images.anandtech.com/doci/11551/epyc_tech_d...Could just be that the diagram is nonsense marketing bling, but that sure looks like external lanes are connected to only two of the 4 cores, with the remaining two getting 'passthrough' lanes.
ET - Wednesday, June 21, 2017 - link
Can someone explain to me why an 8/16 EPYC at much lower clocks than a Ryzen has a 120W TDP?PixyMisa - Wednesday, June 21, 2017 - link
8 memory channels and 128 PCIe lanes use a fair bit of power.damianrobertjones - Wednesday, June 21, 2017 - link
The word 'epyc' (If it is a word) makes me want to cry when thinking about something as serious as servers. Gaming, yes it would work, but not servers.FMinus - Wednesday, June 21, 2017 - link
I guess it might be a hard sell for bosses, when they read AMD EPYC on the invoice, but at the end of the day, if the performance is there and the price is right, the CPU could be named "Momas Big Belly Filling Pie" and I would not care.Zingam - Wednesday, June 21, 2017 - link
AMD loves the Y. Should call themselves AyMD!Holliday75 - Wednesday, June 21, 2017 - link
AMyDHollyDOL - Wednesday, June 21, 2017 - link
AT bench or it didn't happen ;-)Byrn - Wednesday, June 21, 2017 - link
Interesting stuff - any news on when availability is expected?Supermicro have some motherboard details here:
https://www.supermicro.com/products/nfo/AMD_SP3.cf...
msroadkill612 - Wednesday, June 21, 2017 - link
yuk, I would expect at least 3 x nvme ssd ports. a pittance to add to a bobo, but a decent add in controller card will cost hundreds.msroadkill612 - Wednesday, June 21, 2017 - link
As u were, i misread the specs, one mobo does offer 4x nvme.msroadkill612 - Wednesday, June 21, 2017 - link
Sounds a powerful feature for vid editors etc."Hot-swap NVMe/SAS3/SATA3
drive bays and M.2 slots"
Breit - Wednesday, June 21, 2017 - link
Really?: http://images.anandtech.com/galleries/5699/epyc_te..."All 2P E5 scores were derived from the following ICC compiler-based test results per spec.org, multiplied by 0.575 to convert them from the ICC compiler to GCC -O2 v6.1..."?!?
Not cool.
TC2 - Wednesday, June 21, 2017 - link
i smell a tragedy for amd with those "multiplications" :)HollyDOL - Wednesday, June 21, 2017 - link
Reminds me so called "Resultin's constant" joke...[Value you get] [any operator] [Resultin's constant] = [Value you wanted]
TC2 - Wednesday, June 21, 2017 - link
precisely!try another behavioural pattern: great claims, feeble results :)
to compare E5-2698 v4 20 cores with EPYC 7551 32 cores - 32/20 = 1.6, the extrapolation is claimed to be +44% or 1.44.
therefore 1.6/1.44=1.1(1).
now we can to conclude that for a 32 core xeon (it's easy for intel) the result will be +11% for intel, and no more advantage for amd at all!!!
such mathematical practices are ridiculous!
FreckledTrout - Wednesday, June 21, 2017 - link
Anyone buying these is going to want to see some large independent reviews / studies. Some of the larger companies looking at these say like Amazon's data center may get a few and do a study. I think that type of info is what everyone should wait on. These numbers look like they came from the marketing side of the house.Intel999 - Wednesday, June 21, 2017 - link
"Anyone buying these is going to want to see some large independent reviews / studies. Some of the larger companies looking at these say like Amazon's data center may get a few and do a study."Those companies that shared the stage with AMD for this presentation were included in the 5,000 Epyc chips that were given to OEMs/ODMs to test and validate over the last six months.
All the big boys know what EPYC is capable of and it seems that most are quite impressed.
Zizy - Wednesday, June 21, 2017 - link
Yeah well the story is that they obtained GCC 6.1 scores (by benching the CPU) for the top CPU and found out those are ~40% lower than the official ones. Some of that is ICC cheating on tests, some of that is optimization level.So, they reduced all official scores by the same ~40% for this comparison (with the top part being actually benched and achieved that result as on slides).
I can't fault them for using the middle ground GCC for those benchmarks, and the normalization step is reasonable enough as well. The only real issue here is use of -O2 instead of more optimized code. Sure, Ryzen/Epyc is new and GCC likely cannot optimize as good for it (which is probably they reason they used default O2), but this is not a valid excuse, should have used O3 at least if not advanced flags.
petteyg359 - Thursday, June 22, 2017 - link
Don't be a ricer. -O3 is often else than -O2 and EVERYBODY with a clue recommend against using it anywhere, ever.Zan Lynx - Monday, June 26, 2017 - link
Erm. I use GCC with -O3 in a library I build at my job. With profile feedback it is easily 20% faster than -O2. So I don't know where these "clue" people are. I've only been coding for 20 years or so. I may not have a clue yet.petteyg359 - Thursday, June 22, 2017 - link
s/else/worse/gNem35 - Wednesday, June 21, 2017 - link
I believe that people were not aware how overpriced Intel products were. We keep seeing that now from the day Ryzen 7 launched, which is enough for me to never again buy an Intel CPU as long as AMD is at least a bit competitive.It's one thing to make a good CPU and price it accordingly but it's completely different thing to rip off the ones who help you earn money and pay for your employees just because you can.
Very, very unethical.
Joseph_Crox - Wednesday, June 21, 2017 - link
Quite right. I'll do the same thing, I feel taken by Intel.Hurr Durr - Wednesday, June 21, 2017 - link
Welcome to the market.Anato - Wednesday, June 21, 2017 - link
"PCIe can be bifurcated all the way down to sets of x1, although there will be a limit of 8 PCIe devices per x16 link, but this becomes a mix and match game: x8 + x4 + x2 + x1 + x1 + x1 + x1 becomes entirely valid for a single x16 link"Two x1's too much -> x8+x4+x2+4*x1 = x18
jtd871 - Wednesday, June 21, 2017 - link
When you wrote "consummate", you probably meant "commensurate". Page 1.manuelblanca - Wednesday, June 21, 2017 - link
AMD has publiques the results for CINT2006 Rate Andrés CFP2006 Rate for 1 and 2 processorhttp://www.specbench.org/cpu2006/results/res2017q2...
http://www.specbench.org/cpu2006/results/res2017q2...
http://www.specbench.org/cpu2006/results/res2017q2...
http://www.specbench.org/cpu2006/results/res2017q2...
zamroni - Wednesday, June 21, 2017 - link
The AMD-Samsung cooperation is indeed the thing needed to curb Intel Xeon monopoly.mdriftmeyer - Wednesday, June 21, 2017 - link
You do realize that AMD uses Global Foundries and TSMC, right?PixyMisa - Wednesday, June 21, 2017 - link
Global Foundries is working with Samsung and IBM on new process nodes.stimudent - Wednesday, June 21, 2017 - link
Still measuring in inches. How sad.richmaxw - Thursday, June 22, 2017 - link
Why are the clock speeds so low? Even the eight-core model only runs at 2.1 GHz, whereas the desktop eight-core Ryzen model runs at 3.6 GHz. Intel Xeons are the same way. I don't understand why x86 server CPUs always run at a lower clock rate than their desktop counterparts._zenith - Thursday, June 22, 2017 - link
Efficiency. Datacenters primarily care about ops/watt as their metric of success, mostly.Dr.Neale - Thursday, June 22, 2017 - link
On the first page, the article incorrectly states that the rest of the stack will be made avaiable AT the end of July.AMD actually announced that the rest of the stack would be released in a "staggered" fashion over the coming weeks, FINISHING at the end of July.
I think this is a significant difference.
SanX - Friday, June 23, 2017 - link
$100 per core... How many orders larger then production cost?Threska - Sunday, June 25, 2017 - link
Encrypted virtualization, anything like Intel's SGX?agentd - Tuesday, June 27, 2017 - link
Do you think there will be Infinity socket to socket IP available? Previously there was IP for HyperTransport available from third parties.jjj - Saturday, July 1, 2017 - link
New proof of a new stepping for Epychttps://browser.primatelabs.com/v4/cpu/3284583
AuthenticAMD Family 23 Model 1 Stepping 2
rangerdavid - Thursday, July 6, 2017 - link
Yeah, but will it run Crysis???!?StitchExperimen - Sunday, October 15, 2017 - link
NUMA https://www.servethehome.com/amd-epyc-infinity-fab...Rοb - Friday, November 2, 2018 - link
> "... One element that was not discussed is live VM migration across encrypted environments. We fully suspect that an AMD-to-AMD live migration be feasible, although an ...".There's some source code available: "[RFC PATCH v2 21/32] crypto: ccp: Add Secure Encrypted Virtualization (SEV) interface support", by Brijesh Singh @ AMD.com, Thu, 02 Mar 2017 07:38:26 -0800 - linux-kernel@vger.kernel.org/msg1343947.html" target="_blank" rel="nofollow">https://www.mail-archive.com/linux-kernel@vger.ker...