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

So, the fact Intel keeps missing deadlines and becoming less reliable, where AMD is on track, already taped out and sampling 7nm chips isn't worth mentioning, even though they have released roughly the same time for two years in a row now and have shown the market they are on track to do it next year. Really? That sounds more like a cop-out.

Edit: Also GF CTO saying that the chips should be over 5GHz is an industry executive telling the market information. That isn't rumor, that was an interview.

Edit 2: Also, you keep referencing Intel's 8-core. Right now, that is a paper tiger and not released. Funny how you hold that contrary view.

 

You mention only worrying about what you can actually purchase and not paper tigers, where is the 9900?

 

You guys should already know me by now. I'm not from Missouri, but I live in my own show me state. I will reserve any comments and opinions until actual results are available from the talented overclockers that will be pushing the new Intel and AMD chips to their functional limits. None of this matters to me at this point. I do not derive any pleasure from of speculation, best educated guesses or theoretical assumptions. At the end of the day I only care about the final outcome and what it means for me in the context of my overclocking hobby. Once we have reached that point and something is available for purchase, I will choose my overclocking winner accordingly. For now, I'm fine with what I have and not eager to spend money I don't have.

I'm seeing a lot of drama in other forums after all of the silly press releases over the leaks showing the i9-9900K being what the media muppets are calling a Ryzen killer and I'm not impressed by some of the responses from the either the red or blue peanut gallery fanbois. Same applies to the blathering going on about the new GeFarts GPUs. I'll take a wait and see attitude on all of it and wait for the end user results to surface at places like HWBOT and 3DMark.

 

I bought my 8700K for my Clevo in middle May. I will see when it come. I don’t use too much energy on it. But fun if it performs as the rumors say and can upgrade a 3rd time. Aka 4th K processor in same laptop chassis.

 

Which seems to give more backing to my previous statements that a 60% uplift in frequencies might be needed in order to achieve 40% performance uplift (which was stated in an interview with AMD that 7nm should allow).
The 60% frequency uplift is my own personal projection based on little experimentation on overclocking Ryzen 1.

Ryzen 1 technically needs 15% increase in frequency to result in 10% performance increase... meaning that if you want to achieve 40% performance improvement from frequencies alone, base clocks would have to go up by about 60% (and given what we know of 7nm - that its a high performing process designed for high clocks and efficiency, it doesn't seem far fetched).

Now, granted, 60% frequency increase may seem a bit high, so its possible that AMD meant 40% performance improvement will come from a combo of frequency and IPC improvements... which would drop the TDP requirements as well... and would allow them to increase the core counts (but we have no concrete data on whether they want to increase the core count just yet - all we have is a rumour).

This is why in my subsequent replies I mentioned that we could technically see a 45% frequency increase (maybe 50% increase), with rest coming from IPC improvements (if AMD aims for 40% performance uplift coming from both IPC and frequencies - and it would put most of their parts with base clocks of 3.6GhZ at 5.4GhZ... but the slower parts would come at about 4.8GhZ (base clocks - not all core boost and certainly not singe core boost)... but nothing is set in stone.
So, even if say 1700 is uplifted by 50% in frequencies alone, that would give it 4.8GhZ base clocks... and who knows what kind of all core boost (it would likely be higher than 200MhZ vs Ryzen 1 given that 7nm would allow much lower voltages at very high frequencies - and given what AMD is doing with 12nmLP node designed for low clocks and mobile parts already.... its a bit of a doozie to predict what Zen 2 will look like exactly given that the process differences will affect the boost speeds too).

 

You are missing the point. The 10-15% IPC is THE SAME CLOCK! That means you get 10-15% more performance WITH THE SAME CLOCK.

You then are talking about a 39% frequency increase to go from a 3.6GHz all core boost to 5GHz. At your 15% to go 10%, you then have about 27% performance from frequency, which needs multiplied by 1.1 or 1.15 to account for the added IPC. That gets you closer to 40-46% increase in performance. And that is talking over Zen1, not 1+. And that is just working within what we know.

Edit: We know that the GF process allows for a 40% increase in normalized frequency, or a 55% reduction in energy, or there abouts. TSMC has 35% increase in normalized frequency, or a 55% reduction in power. We know TSMC's process on 16/12nm is more efficient on frequency than GF14nm, so that can explain part of the difference in frequencies, which means it should be similar regardless of fab used.

Edit 2: Overall, I'm expecting good things, but am unwilling to say in absolute terms AMD will hit the performance mentioned primarily because that assumes perfect scaling and we have no hard data in leaks to examine yet to come up with better, more accurate numbers. But, generally, am willing to say it can close the gap on Intel if not give a slight advantage to AMD.

 

I am aware that IPC improvements mean that Zen 2 would be faster at same clocks.
I was merely adding percentages primarily from clock increases and then added IPC boost on top of that, so I was off on that).

If you lift the clocks by 27% from 3.6GhZ, that gives you 4.5GhZ... adding another 10-15% from IPC results in 40% overall uplift yes.... but this doesn't jibe with what you posted about the interview.

The interview you mentioned says we should see clocks at ABOVE 5 GhZ... that could suggest 45% frequency uplifts over Ryzen 1 (just frequencies, not IPC gains).... but, again, we don't know if this is for base clocks, all core or single core boosts - again, it is very unclear.

So, if the interview is correct, then 40% performance uplift could indicate just frequency increases - not IPC and frequency together (that is, if it remains that about 15% freq boost is needed to get 10% performance).

Btw... its 1am here, and I'm half asleep... so I could be easily mistaken. Night

 

My only care is at the end of the day for my workload which CPU has the most power at a reasonable cost. To me it does not matter if overclocked or stock, again within reason (IE sans LN2 or secondary active chillers).

 

Do the math. Divide 5GHz by 3.6GHz. You come to 1.39 (39%). So 5GHz is lifting the clock by 39%. Now, 40% increase in frequency, divided by 15%, like you said, gives 2.6666667. You multiply that by the 10% uplift you talked about to get the 27% performance increase. Now, that performance is just from frequency, but the IPC improvement of 10-15% ALSO applies to the performance increase. So you multiply the 127% by either 110% or 115% to get the range I quoted.

Edit: also, what I posted shows normalized frequency increase, not overall performance increase. Intel talked about 25% as performance increase, not frequency. So, what I am telling you is that if all of the SoC increase goes into normalized frequency increase of 40%, then you count that as the boost clock, trying to remain on the conservative side, and considering 3.6GHz boost to 5GHz is 39%, it is possible, you could still get the 40% performance as well due to the IPC increase with the frequency increase.

 

Just in case it was not clear enough before, let me make it clear again. I do not care one way or the other about half of the misinformation being posted on the net. For starters. Anyone under NDA is not going to be posting any facts what so ever until they are authorized to do so. And these same chips that keep getting tossed around are already in beta testing. If you were under NDA one would know this already. And who is already the king or new king in that class.

If AMD comes with a good chip or two, cool. Good for them. About damn time, but if they don't. I do not want to hear the other 200 or so complaints/ reasons/ mistakes/ cheap parts/ cheap material and blah blah blah on why they failed again. Would love to put up some more AMD numbers on my profile, but they need to be "be able to compete" chips for the long run. Not for a month then fall off.

 

Half of what I posted came from the companies involved, such as Intel's slides from 2017 at their presentation, or the 5GHz coming from the CTO of GF, or GF stating that they changed transistor pitch for AMD so AMD can move from one fab to the other, or Intel saying directly in their investor meeting that they will not have 10nm until the holiday of 2019 and server until 2020. Believe it or not, if you looked through what I posted, this is presentations for corporations for a large part of it.

As to rumors, there are the leaked roadmaps which are rumors, the 10-15% IPC is a rumor, the 12 and 16 core on AM4 is a rumor, but came from MSI's own b450 advertisement saying 8 cores and up, which then AMD pulled it down instead of correcting it and did not say it was a mistake, which then would be a false or misleading statement subjecting them to liability for securities law purposes. Those things are rumors, just like the 9900K on Intel's side running 5GHz is a rumor, but also supported by the statements of the company at Computex. The difference is being able to look through the material.

As to density discussions, that is all published information from Intel and GF, TSMC, Samsung, etc. That isn't conjecture, it is what the companies have disclosed.

So tell me again how this is ALL rumors. There are indicia of truth to statements out there. I cited my sources.

So, saying you don't want to participate in speculation is fine. You saying it is all opinion without reviewing the material is not. Facts are independent of opinion. Public statements by companies are public statements, other than puffery.

 

What I'm saying is. I'd rather stay out of these overly opinionated debates.

 

7nm is not out yet so I will reserve opinions on it till then. Now the delay of Intel 10nm does not look good for them. A further delay of the 9900 is not good either. Again till 7nm is where we can get our hands on it, well there is nothing concrete to discuss or get real excited about.

 

I can totally get behind this.

 

Well said. No point in having debate over theories and speculation, no matter how sound the logic might be. Fanbois will vote based on brand. People that don't care will vote based on price. Winners will be determined based on results. If everybody gets what they want, then it's all good. If everything sucks, then we all lose. Wake me up when the train arrives at the station.

 

Trying to get back on the 3DMark Leaderboard!
JayzTwoCents
Published on Aug 4, 2018

 

Well he hyper scrolled past my name at 26. LOL

 

With the 2990wx, and soon 7nm, the leaderboards are most likely about to start seeing a lot of AMD systems.

 

That I do not know. And really makes no difference to me since i'm not the fan boy type. If they have a decent setup then i'll grab it and bench it and call it a day.

 

Before the announcement of the 2990 the general accepted thought was only 16c/32t for the new TR. Under that with 12nm there was little enhancement to give Intel any competition. I would never drop money on so little of a boost. Now with the new skews there is good hope.

Now for $1,799 I have to pause a bit. I can't nearly stress the 1950X as it is now. Getting it just to have it at that cost is a bit much. I was hemming and hawing at $1,500 as it was. I might just have to wait for the 7nm. However if thee is a cooling kit for the VRM's, even though not needed, I may just have to go for it.

 

https://www.guru3d.com/news-story/w...readripper-2990wx-cinebench-score-online.html

Looks like AMD France leaked internal stock scores for the 32c CPU. CB15 of 5099.

Sent from my SM-G900P using Tapatalk

 

I am disappointed somewhat. We do still have to wait as this does cooling solution etc..I'd like to see it with faster memory too.

Edit; This, from leaked info, looked to be the 28 core killer. While it will compete I do not think it is the killer of the 28 core. While Intel's chip will be power hungry and hot they should be able to compete at least at stock.

Using bing translate of the French pages foot notes;

 

5099cb What’s the stock boost clocks for all cores?

 

I would think it they were running stock. That is why I wonder with XFR and better cooling what it ends up being.

 

Strange that AMD France did not want to post best possible results. Remember the result was posted on their own website. And they compared the result with Intel's last years top dog. Or was it the delivered cooler who hit max cooling capacity and XFR can't be utilized?

 

XFR2 is always utilized, it provides stronger clocks with stronger cooling though. As said it is a wait and see issue. As far as last years top dog, well it still is for now.

Edit; Assuming same clocks at stock and 100% scaling on the 28 core;

 

Hummm, either they disabled the other 16C or 3dmark is not compatible?
https://www.3dmark.com/fs/16048341

And 5099 in CB for a 64 thread cpu vs a 36 thread cpu from last year....

And I sure as hell am not spending 10K on a cpu. AMD, here I come.

Although not really sure how people missed all the insulation inside the case.
https://www.tomshardware.com/news/intel-cpu-28-core-5-ghz,37201.html

 

I get a better physics score, something is up there. Everyone knew 5GHz was a sham.

 

That score should be at the bare minimum, 41K+ on Physics

Sham part was the cooling(They should have flat out said it was a chiller - and reasons why - Personal opinion only), not the fact that it can't do 5 ghz. What that means is all extreme over clockers with pockets that deep or people backed by Nvidia and intel allowing them to use those processors for benching will pretty much clean out the top 10 of anything.

 

5GHz on a chiller or LN2 is not a feasible system. More realistically is the other system they had with a 480mm cooler and massive PSU and VRM's doing 4.0 GHz again overclocked with what I am sure was a golden chip. Again we shall see.

Edit;

 

Yep. Totally get that part. Your totally right. Seen that hashed out over and over again.

The part that interest people like me more is what it's capable of doing. Since I have what they have. Water chiller/Phase change/LN2 Pot & tank. Once it's public knowledge we will find out what part of the golden scope it falls under since that is another chip being currently tested under NDA


Edit:
That board looks pretty bad ass!

 

I thought we were talking about 32 cores AMD Ryzen ™ Threadripper ™ 2990WX. Or maybe I missed your point.

 

So, I'm estimating here, but if you guess at the prior leak on scores being correct (getting 6300 or 6400 while at 4.1GHz), along with the scores that Der8auer was getting on the EPYC chip, I'd say subtract around 300 points per multiple, meaning stock would be around 3.6GHz all core or 3.7GHz all core.

But if the stuff on the 4GHz on air with Wraith Ripper is true, a water cooled workstation doing the 4GHz and being around 6000 CB score should be normal.

On the Intel 28-core chip that Tanware shared, we have no idea exactly what speed it was run at (some have guessed 4.4GHz, but no one really knows), and it being on water with at least a 480mm rad, we are seeing comparable performance. Since these are built for heavily multithreaded workloads, I will ignore the single thread performance advantage at the moment, although it is a factor if someone uses those alongside the multithreaded loads.

And I agree with Johnksss on that board, it looks fun AF!

 

I found out that certain 3DMark ones are only able to use certain core counts when watching the GN interview of Kingpin with the roboclocker at computex. I think it is capped where above 16C, your score goes down, so you want to disable down to that for FS and I think time spy, while you can use all 18 cores on time spy extreme for the 7980XE.

It was discussed somewhere in this video on core count for which benches.

 

Hummm, that is interesting. I did see where he talks about that. I just tested it and my scores totally dropped when I ran 16C vs 18C/36T
This was for Firestrike. With Vantage I did have to drop to 18C.

 

Curious. I haven't checked it out yet myself (turning SMT off to run it). So wonder what is going on, or if they updated it since, or what.....

 

Not sure, but give it a quick test when you have time.

 

Later today I should be able to give that a test. Just going to do a quick on and off on SMT for it. Unfortunately, I've only been able to get 3466 stable since the bios update to make it compatible with gen 2, rather than my normal 3600, so won't be my best showing... But should give an idea on it.

 

AMD Threadripper 2000 Series (aka Threadripper 2) vs. Intel Core X
https://www.tomshardware.com/news/amd-threadripper_2-vs-intel-core_x,37550.html

 

Using the above and some math;

 

Probably on same level in Cinebench for Intel’s coming 28 cores chips. But nice numbers.

 

I think for it to run clocks at stock that will match performance wise the CPU will require more power overall. This means a heftier VRM, PSU and cooling solution along with power consumption. That ROG system was a monster that Linus ran and it did only 6,000 CB R15. If this OC's on air to 4 GHz it should do 6,000 on CB R15 as well and at way below Intel's power consumption and 6 channel memory. I am sure 7nm will improve on this as well.

 

No worries, not looking for best scores just to see if it improves or tanks the scoring.

Example on mine: 31k disabled and 37.4K enabled.

This should be interesting in the up and coming months once results get published with LN2 on the 28 core Intel chip. Water chiller(Modded -10C to-30C) vs LN2 (-190C) is quite the jump in temp control.

 

AMD 2nd Gen Ryzen Threadripper MCM Configuration Confirmed, More Details Techpowerup.com Tuesday, August 7th 2018

Tech Day slides leaked to the web by Kitguru provide a confirmation of how AMD is wiring out the additional two dies on the 24-core and 32-core Ryzen Threadripper 2000-series MCMs on the socket TR4 platform. We had speculated that because the quad-channel DDR4 memory and PCIe interfaces are wired to two diagonally-opposite dies on AMD X399 chipset motherboards; in the interest of backwards compatibility, AMD could wire out memory and PCIe from just two out of four dies on the multi-chip module, and have the two additional dies seek memory and PCIe over the InfinityFabric interfaces.

The obvious trade-offs with this design choice is that latencies to the dies with indirect memory/PCIe access are higher, and that reflects heavily in AMD's own performance figures for comparing the 32-core 2990WX with the 16-core 2950X from the same generation. The 2990WX is "only" up to 64 percent faster than 2950X at Cinebench R15 nT, despite having double the number of cores. To its credit, the 2950X has higher clock-speeds (3.50 GHz nominal with 4.40 GHz boost) than the 2990WX (3.00 GHz nominal with 4.00 GHz boost). The presentation also puts out interesting bits of information such as AMD's own performance numbers showing 10-15 percent performance gains between the 2950X and the 1950X; and performance gains of the 2990WX over Intel Core i9-7980XE.


To combat problems intrinsic to MCM CPUs, which are essentially 2P-on-a-stick or 4P-on-a-stick, AMD introduced three operational modes for its Threadripper processors: Gaming Mode, Creators Mode, and Compatibility Mode. The Gaming Mode localizes less-parallelized workloads such as PC games to just one of the two/four dies, for the least memory latency. The Creators Mode makes the MCM optimized for content-creation tasks that scale with CPU and memory, by evenly spreading workloads and memory across the dies. For the 4-die MCMs, the dies with the memory/PCIe wiring are loaded first, and then additional dies are woken up. The Compatibility Mode is basically an SMT toggle. AMD demonstrated in its slides that NUMA-aware operating systems such as Windows 10 use the memory more efficiently.

AMD also detailed the latest version of Ryzen Master software, which comes with a special UI for 32-core/24-core Threadrippers, since you now have to monitor or thinker with those many more cores. Precision Boost Overdrive (introduced with 2nd Gen Ryzen) is also available with 2nd Gen Threadrippers, giving you more control over the Precision Boost algorithm, by making it respect either power-targets, current-targets, or temperature/TDP targets.


This explains the lower Cinebench scores vs. expected with double up of cores for Ryzen's top dog chips.

 

This is the main reason why I said @der8aur's simulation using Epyc was a waste of time. The TR2 4 CCX implementation differences introduce conditions that couldn't be simulated by using an Epyc CPU of the same core count.

We need to wait for actual hands on testing - careful testing and evaluation to characterize the memory access limitations.

Evening the playing field between the new 16c/32t and 32c/64t to get real world comparisons in actual applications used by owners to see the benefits between the CPU's.

It will also be interesting just to see the differences in performance between the 1950x and 2950x.

It may be we need to wait until the October time frame, and new motherboards as well as all the new TR2 CPU's to be released.

Faster is faster, but it's helpful to know how to tune using AMD's new tuning tools to get the fastest possible tuning and application tuning to use the new cores.

Intel's problems scaling monolithic design as it decreases yield are solved by AMD's CCX smaller CCX high yield design, but we are going to need to learn how to work with side effects and through software improve performance.

Intel is going to come up against similar issues soon enough too in their new distributed non-monolithic architecture, hopefully sooner than later.

Once again AMD is leading the way and solving the problems for Intel to benefit.

 

It's very amusing that AMD showed the comparison with Ryzen TR 2950X 16 cores vs. 10 core i9-7900x and not apples vs. apples the 16 cores Intel i9-7960x

 

AMD wouldn't compare them because the pricing of the 7960x is disproportionate:

Recommended Customer Price $1684.00 - $1699.00
https://ark.intel.com/products/1266...Processor-22M-Cache-up-to-4_20-GHz?q=i9-7960x

vs.

the AMD ThreadRipper 2950x @ $899.
https://www.forbes.com/sites/antony...-know-about-amds-new-processors/#666cc4b0f736

A better comparison will be the 2990wx @ $1799

In the AMD comparison charts, the Intel 7900x is also overpriced at $999:

Recommended Customer Price $989.00 - $999.00
https://ark.intel.com/products/1236...-series-Processor-13_75M-Cache-up-to-4_30-GHz

...but the Intel 7900x is closer in cost to the AMD 2950x for the price vs performance comparisons given.

 

But this was about performance, not the price differences Your point is invalide.

 

No, you are wrong, the legend clearly gives the price balance as the base for comparison vs performance:


Core count wasn't mentioned in the comparison, only price. With price constant - Intel / AMD CPU's at the same price - AMD performance is much better at the same cost.