Apparently Ryzen is sick for servers, those Nipples will be mighty fine.
servethehome.com/amd-ryzen-7-1700x-linux-benchmarks/ ^ Some production benches even by dual Xeons. >From a pure performance perspective, the AMD Ryzen is more than just competitive, it is downright awesome. After having hands-on time at the AMD Tech Day, I hit publish on our Ryzen pre-order story then promptly did so for a personal system and a system for our DemoEval lab (ETA on that one late March 2017 as it has some benchmarks it will be running in the meantime.) In fact, I even pre-ordered a complete set of components for STH’s William Harmon since I thought he would love the platform. We are going to have full power numbers once we get them in our standard test beds and into the data center.
>The true power of the Ryzen 7 1700X platform for STH readers, aside from gaming hobbies, will be replacing non-mission critical workstations with better hardware. If you are one of the Silicon Valley startups with a cluster of cheap Core i7 systems in the office or a set of AWS build instances constantly chugging, Ryzen is going to be worth a serious look. Likewise, there is a large segment of the dedicated server hosting business that today utilizes Core i7 chips. Given the choice between a Core i7 quad-core model and a Ryzen eight-core model, we would go with Ryzen. And look at that voltage at 3.3GHz, a frequency a lot higher than what 20+ core chips use nowadays, the perf/watt is phenomenal.
8c/16t at 3.3ghz fitting inside of 60w when all threads are hammered to the max. The Naples Opterons will likely be clocked somewhere around 2.5-2.7ghz which is remarkably high considering the core count. The process clearly favors this range thats right in the sweetspot for enterprise SKUs.
Christian Rogers
Yeah, I've seen the 1700 using around 70w running povray, with great performance to boot, that much cheap and high FPU performance at such a low power is something I would have never expected, I honestly expected AMD to really give us a "okay" FPU, but this is completely outside of my expectations.
Joshua Rogers
Some more power charts to help ya.
Angel Harris
You think performance will increase in production stuff with the new BIOS updates? The CCX scheduling and core parking issues don't seem to affect something running on most of the cores all that much as they do gaming and some ST stuff. I read that some new MSI BIOS update is actually decreasing power as well.
Thomas Green
>In regards to gaming ASUS in particular, and MSI to some extent. It explains why reviewers such as Joker, Crit, UFDiciple, and TechDeals had far better gaming performance.
>Golem.de in Germany had this to say in regards to their MSI motherboard.
>The MSI board was delivered with BIOS version 113, until last Friday a new one appeared.
>Version 117, which is still up-to-date, improved speed and stability. If we were still able to count on sporadic Bluescreens with the older UEFI, the board is currently stable. Much more important, however, is the drastically higher performance in games and the real pack with 7-Zip. The release notes include, among other things, a fixed problem with the memory act and its timing as well as the voltage.
>Compared to the original bios, the new UEFI increases the image rate in our game course between plus 4 and plus 26 percent, on the average even plus 17 percent!
This looks phenomenal, this should further increase ST perf, hopefully that Windows update isn't far off, but MS prioritizes Nvidia and Intel
Luis Fisher
Its a pretty impressive arch all around, especially considering its significantly smaller than Skylake. A single Zen core is 5.5mm2, and a single Skylake core is 6.575mm2. The caveat is that intel's 14nm Trigate process is nominally 15%~ denser. If they were both made on the same process it would show that Zen is over 30% smaller in die area. Getting that performance out of something so small is quite the feat.
A few sites have already posted stuff about BIOS fixes. ComputerBase.de and another German outlet posted articles claiming they saw performance uplifts up to 25% in some metrics on ASUS boards. Though performance on ASUS boards still seems to be a bit behind some of the MSI and Gigabyte ones, and who knows if those are totally sorted out either. There could be a little bit left to squeeze out. I think right now all boards are giving the chips way more voltage than they really need for a given clock, way more than any safety margin for stability. That'll probably get sorted out too, but its not actually a huge deal because Zen has hardware measures for addressing that. In regards to the Windows situation there is definitely something for Microsoft to fix there. Its been demonstrated that if the software scheduler in Windows shifts a workload from one core to another on the separate CCX that the latency penalty it incurs will harm performance. Recognizing the two CCXs and preventing needless shifts in software threads like that should be an easy fix. Who knows when Microsoft will get around to releasing such a patch though.
Oliver Turner
Simple.
Isaac Williams
No skylake xeons or 7nm will save Intel from this.
Under a heavy all core torture test the stock 1800X pulls 110w. In a 4K gaming bench however the CPU has an average power draw under 60w.
Samuel Bell
I honestly think AMD should capitalize on its amazing perf/watt and put the desktop market on the backburner, too much brainlet gamer nonsense, not enough appreciation or general knowledge from tech sites, anything positive gets nitpicked without context.
These aren't a problem in Big Iron where Cray, IBM, Cisco, Amazon and friends shit from their noses for 10higher FPS while ignoring half the lower power draw, they know what they want, and that's a low TCO, no hidden agenda bullshit(mostly) and pure business. This isn't a kneejerk market, and AMD will do very nicely with its efficient arch here.
William Phillips
This translates to embeded and mobile wins, not to mention while actually having a proper GPU arch with working 3D drivers, once those 15-25W laptop GPUs hit I'll probably start pissing myself.
I think it's kinda late but I don't think this chip would fare bad at all in the phone/tablet market, if Intel can get their core M down to 6W TDP then so could AMD, I'm only worried about the voltage cutoff part at 0.7v, seems to be a fab issue and for phones it needs to go lower. Maybe a Zen+ design with more optimized lower voltage process?
The only real issue you could point to is the voltage and clock scaling. You have to give it to intel that they can design a process to handle an incredibly high range of clocks. Having a process do moderately well at low power, and still be stable enough to have a desktop chip hit 5ghz is very impressive. Though we've already seen what they can do in super low power per core designs. The Broadwell Xeons have to hit super low clocks to fit within their given power envelopes. The E7 8890v4 has a base clock of just 2.1ghz, the newest SKU in the line up is just 2.2ghz, and its significantly more expensive. The Broadwell Xeon-D parts range from 1.4ghz to 2.4ghz tops.
It appears to be that AMD's arch on the 14nm LPP process scales down better. They're extrapolating significantly more frequency at a given power level in these lower envelopes, and that'll equate to big wins for enterprise. If they had a 32c/64t Opteron clocked at 2.7ghz they'd have over a 20% clock advantage, more than enough to make up for lower IPC. They still have their trump cards have higher memory bandwidth, AES and hashing throughput, and SME and SEV encryption.
Brayden Taylor
I'd like an AMD Zen powered laptop with ECC RAM, and energy efficient iGPU
Jeremiah Lee
These power figures seem to be AMD's trump card, I wonder why practically no review sites actually mentions these things? Not as interesting as 3000FPS in CSGO?
I remember when Nvidia released Maxwell, 980 IIRC, which was 5% than a 780Ti but used 30% less power everyone shat the bad, now that AMD leapfrogged fucking Intel of all things in perf/watt everyone's silent like a dead fish.
Lincoln Brown
They have a whole range of products planned for the enterprise and server space. I think they're using the consumer market as a proving ground to shake out all the serious problems before they bring it to the big leagues.
Andrew Turner
Good point, big players already have these for testing, if nobody was planning to use them AMD wouldn't be so pumped for Naples. Those racks presented at the Instinct release were sick.
Jack Torres
Intel is certainly impressive, their fabs that is. What I find impressive is these perf/watt figures from a company that's 1/10th the R&D without its own fabs anymore, on a butt naked fresh design, with obvious compromises taken to release on time. The next 2 itterations of the Zen core will be fantastic, if Intel only has Skylake Xeons and 10nm to compete, I don't think they can even close this current gap even if AMD sat on their asses.
AMD's lacking in support and software though, they'll lose no small number of contracts there.
Jordan Flores
Also, Intel's clearly shaken.
Server first now instead of mobile, they know. I mean if they didn't know I'd be questioning what the fuck were they doing not knowing about your competitors new product that's gonna decimate you in your largest revenue market.
I got a wish, I want AMD to grow from its server offerings as a company while the brainlets on Sup Forums and other 'enthusiast' sites scratch their head how that is happening without a competitive gamer CPU.
Charles Myers
I'm doing some numbers in my head and I can see a 4 core CPU running some 3.3/3.4 static clocks using some 30W running Povray, that's nothing short of insane, these laptop chips will be unseen before.
Joseph Jackson
They abandoned their mobile efforts, with the PC market dying their only growing market or servers
Levi Young
I don't want to touch on any blind guesses or conspiracy talk, but I wouldn't doubt for a second if an intel rep told sites not to focus on it. Some really major facts and features about Zen have been totally skirted over by major outlets. Everyone has noted an apparent latency issue with system memory, barely mentioning the fact that Zen's IMC is producing more bandwidth with equal speed DIMMs compared to intel systems. More so all of them have practically omitted the fact that Zen's memory controller loves higher speed memory. In fact Corsair, Gskill, and Patriot are releasing high speed kits specifically for Ryzen. fudzilla.com/news/memory/43022-g-skill-unveils-its-amd-ryzen-ready-ddr4-memory
The timings on these kits are tight, and they're t1 command rate, not T2 like most high speed kits you'll find today. The IMC has absolutely no problem with both high speeds and tight timings, in fact it favors them.
And despite AMD shipping review kits with 3000mhz DDR4, most reviews I've seen around the web were using 2400mhz or 2667mhz kits. Different listed SKUs even. They didn't just lower the timings themselves, they simply tested with different kits. That is beyond suspicious.
Intel is a much larger company in terms of employees and budget, but that doesn't tell the whole story. The more employees a company has the more HR staff they'll have, they'll have more pointless PR people, they'll have more legal team. Intel's staff is spread out into a lot of divisions, and a big part of that is because they're invested in markets that AMD simply isn't competing in. They make their own cellular modems for example. They have teams developing flash and non volatile memory. And of course right off the bat a significant chunk of their total man power are fab workers, and their fab business constitutes a huge portion of their R&D expenditures.
When it comes strictly to X86 core development intel doesn't have that much more than AMD.
Adrian Martin
What the enterprise market shares in common with the mobile/ultrabook market is that they demand chips with the same power per core. A ultrabook chip has its CPU cores pulling power about equal to what you'd see in a big server chip. So naturally if your process is yielding an arch that is good at one thing, it'll also be well suited for the other.
The APUs should be fairly impressive at low TDPs. I wouldn't be surprised if the 35w mobile part ended up having near 90% of the performance of the 95w desktop part. It also stands to reason that Global Foundries will further refine the process as time goes on, and they may be able to squeeze out a tiny bit more frequency at a given power level. The company absorbed a lot of IBM's process engineers, and I doubt they're sitting idle. Though they'll be working to get the 7nm SOI lines up and ready for risk production in later 2018 some of them had to have gotten assigned to working out kinks with 14nm production.
Lucas Perry
7nm SOI on whatever the Zen iteration will be using it will be lunatic, not only a will we be doing a major leap in density and power, but also a switch from bulk to SOI, which by itself should be close to the benefit of half node jump.
I just hope GloFo doesn't fuck it up, they did good with Samsung's process, I hope they do the same with IBMs
Justin Flores
You think AMD will get rid of the GMI links for infinity fabric in their mobile dies? I don't see much use for them there. Those GMI links are pretty big, should lower die sizes by a good chunk.
Aiden Rivera
I wonder if AMD is going to release K12 this year after the other stuff lands or if it's on indefinite hold. I hope they manage to release it, I wanna see Zen in ARM form. It would be amazing if there were K12 APUs in smartphones
Brayden Johnson
IBM demonstrated test chips being run on their 7nm node quite some time ago. Though they were being pretty fucking smug in calling them test chips, they were fully functional POWER chips of some variety, but wouldn't go into further detail when pressed. The only issue I could foresee with Global Foundries implementing it for mass production would be low yields due to the usage of quad patterning, though IBM may have developed some IP specifically for this. Still they are hoping for production to include EUV in 2019, I think thats still very ambitious, but TSMC is claiming the same thing for their own 7nm bulk process. They're claiming EUV in 2019.
Ryzen has those links because the die was designed to fit the enterprise market, its a consequence of being multipurpose. I don't think the Raven Ridge APU will be used in the same MCM configuration, so I wouldn't think they'd be present. Though Infinity Fabric is sort of a catch all name for a lot of different things. The internal buses that connect the CPU, memory, and IGP in Raven Ridge are all technically Infinity Fabric.
Jack Mitchell
Trust me, K12 is still there, but clearly AMD is focusing on x86 first and foremost since that's where the money is, ARM in dense servers is still a few years away.
Logan Edwards
It gets even better, AMD can cherrypick 4 efficient dies to put in the most expensive Naples chips. And you end up with a SKU line literally made out of legos, TDP and clocks all over the place.
Adam Adams
The issue with looking at the top end of 140w, is that this is isolated to floating point heavy operations where all modern CPU's have more floating point power at a core level than they need overall. If you look at power consumption of modern intel CPU's performing AVX instructions, it absolutely overdraws it's intended power range and can easily saturate the memory lanes in the highest end offerings. It's a misleading measurement to look at because 80% of float heavy enterprise applications offer GPU accel, and the other 20% are where the specific use cases of the Xeon Phi series are the only reasonable alternative. It's nice that Tom's Hardware offers these measurements but the final conclusion drawn shouldn't be "this processor will regularly operates at 140w", since the use case is very small.
Jonathan Baker
I'd love to see high performance variants of Zen and K12 go head to head in some benchmarks. I wonder if there would even be any notable differences? I guess there must be some but the two core designs would be quite similar you'd think.
Levi Rodriguez
That 140w line is a quick manual overclock, not what the chip draws at stock settings. It only pulls 110w~ stock, and thats the heaviest possible load they could place it under. Its drawing significantly less power than Broadwell-E.
Kevin Williams
Oh. That makes things even better. But, just wanted to point out that modern architectures all seem to feature overdedicated and underutilized floating point units on a per core basis.
Jace Cox
Yeah, but if you're offering similar perf/watt with x86 and ARM and similar sized cores, what's the point of ARM with the x86 library?
I think the K12 chips will have far smaller cores but bundled with a fuckload of I/O and memory bandwidth. There's a good amount of money in markets where connectivity, cost and low power come into play more than performance.
Jason Garcia
>Yeah, but if you're offering similar perf/watt with x86 and ARM and similar sized cores, what's the point of ARM with the x86 library?
I think AMD asked themselves this same question when they realized just how well Zen scaled in practice. If an X86 core with that level of performance is capable of attaining those low power envelopes while maintaining appreciable performance, what exactly is the point of an ARM core? Either K12 is significantly smaller and even lower power than Zen, or K12 is significantly wider, and contains the massive large vector throughput FPU that Zen lacks. I could see the latter happening more than the former, but I don't think thats likely to pan out. ARM themselves are pushing for it, ARM for HPC, but I don't see that getting sales to compete with X86. Performance would have to be high enough to draw clients away from their familiar software ecosystem.
Connor Stewart
If you have that wide cores only for FPU throughput, why not swallow the pill and go for GPUs? That's their only job in GPGPU.
Yeah I know that GPU's cant branch for shit, but aren't these specific AVX512/AVX workloads mainly not too branchy that you can get away with a little higher latency for massive bandwidth?
Jeremiah Reyes
>Ryzen singlehandedly should make you reconsider an AWS implementation Stopped reading right there
Juan Bennett
Honestly have you seen the idle power of this thing? Laptop are gonna become even thinner(FUCK) and more powerful.
Easton Rivera
For whatever reason a lot of people are seeing a lot of potential in larger vector instructions. Some of the most powerful super computers are using AVX crunching Xeon Phis. ARM themselves are trying to design arch to take marketshare away from intel here, Qualcomm wants to spearhead it. They seem to think that they can edge out GPUs for a lot of scientific workloads.
Evan Ortiz
All that while FP64 Tesla's continue to sell like candy.
It's almost like there's some kinda split in the HPC market.
Joshua Lee
But that's the thing. Why would they want GPU-like performance on the CPU that means compromises on the CPU rather than using a GPU?
Xeon Phi simply does not compare to the Tesla and it won't compare to Vega, no matter how much Intel's marketing on the Xeon Phi lies about it.
I could understand with an APU. Maybe you could do away with the FPU on the CPU almost entirely and use the GPU for vectorized SIMD float tasks.
Nathaniel Bennett
XeonPhi shat the bed because Intel promised "plug&play" instead of screwing around with GPUs and hiring opencl/CUDA dudes that want 6 digit salaries.
It didn't work out, turns out you're hiring similar people because the Phi is a pain to program for as well, but ends up being less efficient than GPUs.
Blake Evans
Isn't 6 digits putting it lightly? I imagine such engineers must make around 200k-600k
Sub 100k is "video game programmer" level.
Andrew Lewis
It's still 6 digits, but semantics aside these people are paid real well, since there's not as much as them and there's demand for them.
