Ryzen vs i7 (Mainstream); Threadripper vs i9 (HEDT); X299 vs X399/TRX40; Xeon vs Epyc

To tell you the truth; I glossed over the actual testing (you know me and 'BM' 'scores'...). Doesn't mean anything to me or my workflows.

What I found 'balanced' in both articles is the pointing out of strengths and weaknesses in both platforms, which they both did very well, I thought.

Sure; it may still seem biased to some. But the points made is how buyers will view them though... not how we here on this forum may judge...

As ajc9988 has already stated; purchasers don't rely (solely) on these type of reviews; just like me, they want sample hardware for themselves to see how applicable it is to their workflows/workloads in a real world settings.

 

This sector or class of CPU previously was extremely limited and I doubt there is going to be a rush to go higher end to begin with. Eventually we will see the systems get further penetration in the market. At the risk of repeating myself OEM's need to get the systems available for purchase not just Alienware.

My only real concern is the 1920x and 1950x use stepping 2 of the core. We were told his would be true of the Epyc and TR as there are some micro code changes not includable in the updates.

 

There is nothing to worry about. I did the math and if we are looking at 90% scaling due to IF, 100% would be 3402 on the cinebench scores. Half that and you get 1701. Most have numbers from release of Ryzen 7 using 2400 ram at around 1600. If you do 3200 as 100% scaling, that means it is getting about 95-96% scaling. So, it is right in-line with where it should be atm. With that being priced against Intel's 10 core, if you do heavy multi-threading, which software is now adapting to, you get about 30% performance boost for the same price. With single thread, you still get BW-E performance levels. So you aren't taking much hit in single performance to get extra pcie lanes and good multithreaded performance! We still need to wait for reviews and MB reviews to make sure things are on the up and up. As to TDP, it is 180W. Ryzen is 95, but we've seen it overclocked pull 150W under load before (although still better than Intel Skylake pull). So let's say for the overclocker 300-350W. It should be noted, as you increase ram speed, as seen on Epyc, you can see 15W per 266mhz on the ram. It may be less with only 2 dies, but that is why we leave the leeway. It is less than the wattage on the 7900x though, even at that with 16 cores, if I'm correct. That is a HUGE achievement, especially with the wattage on the Skylake-SP chip and the overclocked wattage of the 10-core as references.

Edit: I forgot the scaling on the 12 core, which showed 400 points on cb 15 over the 10-core Intel, which was about 400 points over the 8 core Ryzen. It is also about 50% higher than the 1800x. This means they are doing well even with the IF and extra die.

As to memory access, Epyc has two jumps max to access any memory on a 2P system. This has 1 jump max to access the memory on the other controller. So you should see slightly better memory scaling than Epyc, but less than with a single controller like Ryzen 7. But the memory analysis will come soon enough.

With the server segment, I let some bias slide, more than any other, because we are talking 99% of the market. Everything has been optimized for Intel for years and specific tweaks between icores isn't much or hard. It's nailing down firmware and feature set utilization that takes the most work. I'd be dishonest not to recognize this advantage. It's also why STH is still trying to optimize their Epyc server for a good comparison (even if there is bias, I do note where it is fair and that most are given proper disclosure and coverage, at least in anandtech, admitting I haven't reviewed Tom's). There may be some shenanigans in slim numbers on Epyc, but it also means when the comparisons come in more in the weeks to come, Epyc will have had months with experts having hands on tweaking and sharing information to get better performance, some firmware updates, etc. So, when the wash comes out, we'll see and the numbers are good enough for test beds.

It also will help small and mid-side businesses and researchers with limited funds get access, as they don't always get as deep a discount as big businesses.

Sent from my SM-G900P using Tapatalk

 

Wow, getting desperate;
https://megagames.com/comment/1385061
https://www.techpowerup.com/235092/...cessors-glued-together-in-official-slide-deck

 

Yeah

 

I love this guy.

 

It seems the response from AMD is due to 7900x hitting the shelves today. I doubt Intel was too happy with the 1920x showing up the stock 7900x especially being as it should also be competitive with the 7920x. Right there is a no brainer let alone a 1950x that may in some respects beat a 7980XE.

This along with Epyc I am sure being a major thorn in their side. Yes Intel's new higher end chips outperform the Epyc but probably constitute less than 5% of server sales. AMD is placed for eating into the meat of the server market and Intel is not happy.

 

You missed a larger point: That the 10-core overclocked to 4.8 only gets about the same as the 12-core at stock. http://www.guru3d.com/articles_pages/intel_core_i9_7900x_processor_review,21.html ([email protected])
https://www.bit-tech.net/reviews/tech/intel-core-i9-7900x-and-x299-chipset-revie/5/ ([email protected])
https://www.techspot.com/review/1437-overclocking-core-i9/ ([email protected])

Some came in higher than the 12-core at stock when OCed to 4.6-4.8, but not by much and were destroyed by the equivalent priced 1950X 16-core beast.

 

I know about the OC but did not mention it for two reasons.
1.) we do not know yet with TR how it will over clock both on the core and memory and it's respective impact on performance.
2.) most people will purchase an OEM system eventually that will not be overclocked. Since they will be primarily used as constructed I think for the majority how it performs out of the box is more important.

 

https://forums.servethehome.com/ind...roadwell-ep-qpi-architecture-explained.15618/
A bit on the wait for the AMD Epyc system from STH.
"Looks like the box we have is getting an upgrade later this week. The good news is that will push us to DDR4-2666 which is important for IF as well.

Hopefully the new configuration we can publish benchmarks on. I have two sets of charts for the Skylake launch. One with 2x EPYC 7601 + DDR4 2400 and one without.

AMD is sensitive to us publishing numbers on de-tuned platforms, which I understand and agree with given it is a new platform and we are still a few weeks from folks being able to buy them publicly.

It is one of the hard bits about the work and a lesson learned from Ubuntu 14.04 with ThunderX that we are carrying forward. I do not think it helps anyone publishing pre-release system benchmarks if there is an expectation that those numbers can change.

Also, STH is not using an AMD reference platform tuned by AMD engineers so we are going more for real-world performance."

 

I don't see anything wrong with those statements? The tone could use some work, of course.

But that is what they are. The statements Intel points out are true (for server buyers).

AnandTech, Tomshardware and many other online mags basically said the same thing; optimizations are necessary on the AMD platform to get the most from it (no news to us, of course). When you're betting $1M and more for the next half decade with your employer's $$$$$$$$$$$$$; you don't impulse buy what happens to be cheaper - because cheaper does not get the work done.

 

That depends.

If you are out purchasing for $1M, and you buy from AMD, you could double the workload ;-)

 

That has to be proven on a case to case basis (double the workload for the same $$$$).

You could also be throwing a wrench in the operations that may end up with looking for another job (if anyone will hire you at that point).

 

You both have points. First, my siding with tiller, you have limited floor space. Buying double the amount of racks, etc., may not be possible. You have to look at the CPU density per cubic meter, which means you may have the budget to buy more AMD units, but you could run out of room before then, thereby making no sense. Second, I side with jaug, that you will likely not be out of a job. This is why we pointed out test beds and having all of that squared away before full deployment, that way you avoid the wrenches. I think @hmscott could speak more on that. So, the question is whether it frees the budget for other purposes, like getting those tasty TV100s if you need them, and the density on that, or upgrading to faster drives for the storage, or more memory on the budget, if those are needed.
(Edit: I read in a bit on where else your exact arguments would lead and did some assumptions on what you guys would say, but tried to stay true to the lines of thoughts while providing my own here. I do apologize for assuming. Here, feel free to let me know if you would disagree with my logic deductions on the lines of thought)


But, ultimately, it will vary by deployment. So you both have great points. But, neither is the clear winner and, if anything, we have seen people working to optimize the servers for accurate reporting or using an AMD optimized server for the numbers currently out. We have seen certain areas where the Epyc chips are outright better than Intel, but Intel, overall, and in many sub-categories, does provide the better results (albeit at double the cost). Now, on budget limits or if you are using these to drive certain other components, we may see a change on what is preferred. Both systems still need tested more. So let's chill until that information is available.

 

The ([email protected]) Cinebench R15 seems a bit low Sweclockers.com/test i9-7900x

 

With the other scores, I'm taking it that was a fluke. Also, that was before the VRM stuff came out, so it may have been power delivery without seeing CPU throttle. Either way, I'd say an outlier and that we will see better as the platform develops. But, it is in the ballpark of the stock 12-core before people get their hands on it. But if Ryzen showed us anything, it is to expect 200-250 on the score, not a 300-400 increase on this test. That may grow as it may get that per die, but that is speculation.

Sent from my SM-G900P using Tapatalk

 

There is a tipping point, where the application demands set the trigger for purchase, and it's rarely an unclear choice.

Once AMD hit those marks in the datacenter during their last heyday, Intel couldn't give away their hardware - they tried.

Even with some odd bugs in the AMD firmware cropping up during massive deployments, the bugs were mitigated until fixed, and build out's continued.

There were installations where AMD couldn't get in to be evaluated, to the detriment of the clients needs.

I don't see it being much different this time, and if any with those experiences are still calling the shots, AMD will get a chance to prove it's worth again.

I hope AMD has followon work in progress to keep pressure on. The new product starvation to continue in the datacenter is what killed them last time. Intel has to hope that's going to happen again.

Even so, all those AMD purchases paid for themselves well over and over such that there was no loss of face or position for those that chose them, and again, there was a clear choice as AMD dropped away and Intel came in to pick things up again.

The benefit was that Intel was much more willing to negotiate and that humility remained for quite a while.

I think there will be clear choices as customers make tests against their software mix, and I also think optimization will be pushed to fit AMD's exceptions and advantages.

Intel would do well to start dropping prices now

 

Let's see what the Infinity Fabric is all about

http://www.anandtech.com/show/11624...n-compatible-dimms-up-to-ddr43400-up-to-64-gb

 

Right now, 3400 is about tops with Ryzen 7 before playing with the bclk. So, this is good, but we need to wait for details on quad channel and see what the new stepping does.

Sent from my SM-G900P using Tapatalk

 

A breakdown of the Anandtech review.

 

https://www.pcper.com/reviews/Proce...nounced-Flagship-Performance-999?sf97938541=1

Sent from my SM-G900P using Tapatalk

 

How well ryzen has done is scheduled for release July 25. http://www.marketwatch.com/story/am...ep-for-volatile-post-earnings-move-2017-07-14

 

Well, I would hope a bit better but the real sales have not yet started. I am sure some money from OEM's is starting to come in but until there are some real sales of off the shelf box's it will be hard to say where AMD is.

 

True, but there is mixed evidences of the DIY crowd. This will give a clearer picture of that market.

Sent from my SM-G900P using Tapatalk

 

I found a great gem on reddit, about this very topic of just how Intel messed everything up for themselves.. and why.

AMD's response to "glued together" bash:



why Intel is failing:

 

https://www.fool.com/investing/2017/07/14/key-detail-of-intel-corporation-tiger-lake-process.aspx
So, Intel will be introducing new iGPUs with Tigerlake (and every generation in between abouts). That's all I got from this.

 

With respect, but I didn't find how those videos were relevant to the comments you made?

As I've mentioned previously; they are 'glued together' parts - AMD admits as much too - that doesn't make them bad; but Intel is hinting that they've already explored that option and are on the path to something better.

As for the SJ's video? An interesting aspect he highlights, for sure. But nothing in the form of hard proof that would make me believe it. (For Xerox or Intel, let alone any other company that is at that level today, or ever).

 

I got a bit more than that.

But... I do not agree with the above assessment at all. There is no 'bungled transition' that I'm aware of. There is only experience and learning. And; Intel does plan many years in advance...

Intel gives us timelines and other hints of what and where it expects to be (usually; to the quarter), thereby exposing itself when those lofty goals cannot be met for one reason or another; I'm sure other companies that play their cards much closer to their chests had the same/similar issues yet went unnoticed because they don't communicate with their customers like Intel does.

What I got from this is that Intel seems to be on track again for the 10nm rollout. And, when I'm about ready to update my platforms again in a big way then; I'll be jumping multiple generations in graphics architectures on a lot of my systems.

For someone that can't stand the idea of a discrete gpu sucking down power, heating up other (more important) compute components, making additional noise/vibrations and annihilating battery life/power efficiency; this news is almost as important as new CPU's and platforms themselves.

This is the type of 'sharing' that Intel does (even inadvertently) that might get them in trouble in the near future (if/when they can't deliver on time). But for mere customers; this type of info shows that Intel is not just resting on it's current/past successes - they are still hard at work optimizing every aspect of their offerings (and I may add; as a cohesive whole, in which the whole package is greater than the sum of it's parts).

 

I have to ask, please place a reference where AMD admits a RYZEN is a"Glued together part"? As far as I can tel Ryzen is a highly integrated sect of CPU CCX complex's integrated by the IF. "Glued together" just infers two items adhered to one another by the glue (physical bonding material) binding them.

I understand Intel misinforming the public but now you have lost just about ALL credibility with me!

 

Hmmm, I thought I already posted this AMD video... I did a search and can't find it...

AMD Ryzen™ Threadripper™ and Ryzen™ 3 Product Updates

 

http://forum.notebookreview.com/thr...399-xeon-vs-epyc.805695/page-75#post-10563858

 

Look at the latest videos posted here (sorry; I can't index the videos I've watched).

Agreed that there is much more than just the 'glue'. And agreed that it is a smart way to do it (i.e. it is complex...).

But the specific video (if I'm remembering correctly) states that this was done on purpose because it gives higher yields...

Can't really remember if this was Ryzen or Epyc, but it was AMD themselves that state that...

This isn't a subject of debate or my credibility; these are simple facts known from day one. Intel isn't misinforming the public and neither am I.

IF makes the separate cores work together as seamless as possible ( but not seamless enough with certain workloads). But it is not as seamless as when a monolithic die is used (even if it helps with multi processor setups on a single board).

From the link in the paragraph above:

The above is just talking about Memory bandwidth, latency and L3 cache. It highly depends on what you're running (from only 42% of SL-SP's BW to 2.26x more).

Hope someone can remember the video I saw (seems like days ago now?) for you...

 

Weird, it's not finding that media id => J3pJ_--nf5E

I guess some of those special characters aren't allowed in a search... first time it's done that.

 

Again, I need a link. of the video's watched here the only one here by AMD claims "it is being said by some (we know who that is) that Ryzen is a glued together solution and this video is to dispel that thought". in other words they are disputing that. You attempting to make their refuting of the claims the validation of them again where is your thought process?

Yes they approached the present methodology to increase yields among other advantages, and? Ryzen is a whole integrated solution not just any one part.

 

I tried to search my history for you; but I don't even know which device I viewed it on.

I don't know why anyone would dispute this basic fact anyway? That is the whole design; when looked at a zoomed out level.

Like I said; this isn't something new: even this old article from March 2017 states the same thing:

See:
http://www.anandtech.com/show/11183...aples-cpus-for-1p-and-2p-servers-coming-in-q2

My thought process is always to view things as they actually are: not just how they are presented.

Right now, Intel is saying a monolithic design is better (and some of their numbers/'scores' support that - according to other links we've all read here in this thread). No misinformation being disseminated - just one company differentiating itself from it's competitor.

As usual; the future is where we'll see who was 'right'.

 

@tilleroftheearth Not once is the word glue or glued used in the article. now you have not just about lost all credibility, you have. I will not ever discuss another thing with you because even if you are being somewhat intelligent I can no longer trust a thing from you. I can not ignore you as I have to moderate but consider yourself ignored as to myself your points are now fruitless as well as meaningless!

A last word, admit it you have resorted to FUD to push your agenda and you are now trying to justify it which is what made it so very much worse. You were willing to give away your own intelligence as a human being just to satisfy your cause, I hope Intel appreciates it.

 

Intel Says AMD EPYC Processors "Glued-together" in Official Slide Deck-Techpowerup.com

 

Really, like Intel haven't done this before Remember Pentium D? Core 2 Quad? Atom 330?

 

If you go to either the Pentium D dual core, C2Q or in AMD's case the BullDozer you have not much more than exactly that, glued on extra cores to the die. little was done to make the CPU's either modular or scaleable. This is what Intel wants you to think of when thinking Ryzen. So Intel is relying on FUD and old memories.

Even the staunchest of the Intel people out there are questioning the marketing here. Even in the days with Opteron where AMD had the performance crown hands down FUD was not used as a marketing tool. but there are two major fronts they are facing when it comes to bottom line.

1.) the fat cow has gone on a severe diet. Prices have come down allot for the same performance chips. Even those choosing a new Intel system will be getting a much better deal than they had before. This will cut into the profit margins for these chips.

2.) Even the new skews now face competition. This in the 95% of servers that are actually constructed. that is the High performance 1P and 2P systems and some smaller 4P ones as well. Now this cuts into market share again cutting into profitability.

 

Just to address the connection between dies, the numbers suggest that you are getting 90%+ scaling. If you use 1600 cinebench score as a base, the 2400 score by the 12-core would be a 100% scaling and the 3000 score on the 16-core would be at 95.6% scaling, which is impressive (the disparity likely is from the extra cache per core on the 12-core, but it is unknown yet). If you use about 1700 as the 8-core representative score after optimizations, you get around 95-96% scaling on the 12 core (still impressive) and 90% scaling on the 16-core (also with the caveat of the cache per core ratio). So, this isn't just glued together, nor Intel's prior disaster of an implementation compared to what a single die could do for performance. This is why Nvidia is taking a similar approach moving forward and AMD may implement similar on its graphic cards. Now, with the extra cores, the question is whether Nvidia will take the artificial two-card SLI cap off their products, but that is too far out to know yet. Also, considering AMDs SMT implementation provides a couple percent better multi-threading than HT does, it works out in the wash. This means that AMD has a really competitive product on their hands. By the time the IF2 comes out (2019-20), we should see even better gains, although that is promises of tomorrow, not for today (but it should use the same socket and boards).

As to anything from March, that is before information was known on the data transfer rates of IF in relation to Epyc. Now, there is more information available that changes the points made in older articles. I agree to view things how they are, but you didn't review the numbers when you posted the Anand review article either, which shows where things are as far as public information, noting your ever-present hedge that BMs are not representative of real-world results (and which I note my comment discussing the wait for the STH article on the topic). But, that also doesn't mean one should accept Intel's marketing FUD either. This is why looking at the percentage scaled and the raw numbers on performance are important, as otherwise, it stands that AMD would have beat Intel but for its use of smaller dies interconnected. Is that what you are standing for? (I doubt it) What matters is that performance crosses a threshold and that it works with the ecosystem, or can be made to work with the ecosystem at relative cost and in a specified time period. All x86 processors can use mostly the same instruction sets, with some variance. But optimizations must be made for every chip, including Intel's chips.

If it was a simple differentiation without FUD, no one would be talking about Intel's faux pas. But we are. Why? Because it was disingenuous and was beneath the market leader in chips. It shows weakness and fear. In fact, it will likely cause more people to look at and test Epyc than before, because Intel doesn't behave like this, raising the awareness of what threat level Epyc is to Intel. It was a stupid move on their part!

 

@ajc9988 rofl adoreTV talks some tech, the guy is actually very good at observing little things and i remembered his video talked about how recent games are starting to do better with bulldozer than even 2500k because it had more cores, software are starting to utilize more cores now so AMD had that vision years ago but it was simply too early, now its finally doing better.

also, look at this video him talking about skylake-x and tomshardware


tomshardware biased towards intel? i think not. at least not for paul alcorn

edit: LMAO @ someone's comment Intel Skylake X - $100 per core, $10 per 1°C

 

Things can change, especially with time and being called out. But, the retreating of Ryzen was nice and I give kudos on the review done once available and more was known. Even I am gentler on optimizations, although when it comes to games, I agree. Also, the variance in cache on AMD only occurs with that linked per core, so a little less is needed as the L3 has no change between chips. I disagree with the mesh conclusion as Intel even said that the mesh adds latency to inter-core comms. But, it lowers latency to I/O and ram, as well as other CPU components. Also, notice the scaling potential with faster ram and the likely support of 4000+ ram, potentially even the upcoming 4800 from G.Skill. This will help with mesh latency just like faster ram makes IF latency lower (not saying these are the same, just that a similar effect is being seen on the two different solutions).

That edit made me lol! Also, sometimes you can find a good journalist among a sea of biased reporting. But point taken.

Sent from my SM-G900P using Tapatalk

 

from what jim@adoretv mentioned, theres maybe little chance for skylake-x to get optimization and that troubles me. my biggest needs arent gaming so losing 20-30fps isnt an issue but sure sucks. i just wished intel didnt apply those mesh changes or cache changes to 6-10 cores, only for 12-18 cores.

based on tom's anand pointed out, theres no IPC loss, its all software optimization and not optimized for the new cache design. i have a lot of stuff uses multi threading maybe i'll just stick to my original plan get 18c and have it cover the performance difference rofl.

 

At that point, the platform will be more developed, the cache utilization will be better, etc. But, by then TR will be much more developed, and will see many optimizations as well, considering the Ryzen optimizations may transfer to TR for part of that...

Sent from my SM-G900P using Tapatalk

 

Or, jump the fence and pick up an AMD ThreadRipper 12c/16c for much less, or an AMD Epyc 2P/1P with appropriate core count, and explore new worlds - for a lot less cash

 

Just an AMD Fanboy?? - i7-7820X Review Comment Responses



WHAT VRM HEAT ISSUES? - ROG Strix X299-E Gaming Review

 

i9-7900X Vs. Ryzen 7 1700 Vs. i7-7700K - Is Intel's Flagship Flawed or Failed?

 

I think the x299 and 7900x are getting much worse than they deserve. Like Ryzen it is way early in the deployment and yes there are growing pains. Just as I am expecting some with TR as well. But as always we push things well beyond their intended limits just to see what we can get from them.

 

I think some of that comes from how tough a time was given to Ryzen. It is the turn around while scrutiny is on every new piece of hardware now. This is why I do actually argue against too harsh a treatment, but also try to point to issues that exist. I know some improvements will come from cache optimization, but those will also help Ryzen in many cases. The added support for n-core scaling will benefit both companies moving forward (and ARM, but that is a future fight not yet worth talking about, just like increased GPU use for async parallel processing).

Intel will have many issues straightened out by end of August, into September, which is fair and the time to address many issues with X370. That also means that except for software optimizations, Intel needs to hit the ground running on the HCC CPUs in October. This should give X399 about the same time to mature (2 months for major issues, then 6+ months for full optimizations with regular updates until EOL). I originally planned on getting the Intel $1000 chip this round. Then I saw Ryzen and figured out they are coming to HEDT. I hit the brakes, but came to the realization that a 16-core at 4+ghz at the original price planned (12-core is what I still stand by for Intel on value on their lineup, but $1200 wasn't known until I switched). But therein lays the point, things change. We still need proper x399 execution. But with OEMs on board at launch, I think we see a change in prospective support versus Ryzen at launch. Intel rushed, so they get the growing pains. But that doesn't mean it isn't a good product, just that we have options. Options available in less than one month at extremely competitive pricing.

Sent from my SM-G900P using Tapatalk

 

Intel Core i7-7800X Overclocking, Temps & Power!


Again, more outrageous heat and thermals from huge power draw when OC'd...

He sums it up at the end saying that he got 5% better gaming performance from OC'ing, so he doesn't think it's worth the power and heat hassle.