AMD - Zen chitchat

[kompukare]

Somebody on The Stilt's Ryzen Strictly Technical thread over at AT, found this Thai comparison of the 7 1800X forced to 4C/8T (single Cand i7-7700K both at 4.0GHz:
http://www.zolkorn.com/reviews/amd-r...core/view-all/
(Google translated)

No summary, but a quick tally up gave me this:

At the same clock, gaming doesn't really look that bad and bodes well for Ryzen 5 1400X (or Ryzen 5 1300 overclocked).
The gaming advantage Intel continue to hold is that 4.0GHz is the max speed for Ryzen (for now), while Kabylake can be clocked to 4.8GHz+.

[CAT-THE-FIFTH]

Have you noticed how AMD successfully got most reviewers to test newer games??

The biggest weaknesses are in games using older engines,which is where AMD had weaknesses in the past. Ryzen is nowhere near Skylake levels of IPC sadly,and is closer to SB or IB.

[kompukare]

But then again AMD PR are not known for being able to set their agenda in reviews, so it's equally possible that reviewers are mostly using modern games.

And the only real big recent releases with very poor scaling are both from Bethesda (aka Bugsoft) aren't they? Fallout 4 and Skyrim:SE are well known for being based on old tech. The initial Skyrim release was compiled for x87 and it seems only after being embarrassed by the likes of the SkyBoost mod did Bethesda compile a SSE version. Unsure about FA4 and SSE as I haven't tried any SSE mods until SKSE comes out.

Of course, while not AMD's fault it doesn't help those who want to play those games now. Personally I'd rather have more cores so if I were to upgrade from i5-3570K (not even overclocked), I would not want more than 4C/8T anyway. My biggest concern with the current Ryzen is actually their idle power which I feel could be lower especially since FM2 for instance used to have better idle than Intel.

The other thing I want from an upgrade is to go for a smaller case. The reason I am full-ATX currently is that someone gave me the Z77 motherboard so I bought the IB i5 even though Haswell had just been released. Normally I buy mATX, although am looking at mITX too. But really would like something way smaller than 20L otherwise there's no point.

[CAT-THE-FIFTH]

But then again AMD PR are not known for being able to set their agenda in reviews, so it's equally possible that reviewers are mostly using modern games.

And the only real big recent releases with very poor scaling are both from Bethesda (aka Bugsoft) aren't they? Fallout 4 and Skyrim:SE are well known for being based on old tech. The initial Skyrim release was compiled for x87 and it seems only after being embarrassed by the likes of the SkyBoost mod did Bethesda compile a SSE version. Unsure about FA4 and SSE as I haven't tried any SSE mods until SKSE comes out.

Of course, while not AMD's fault it doesn't help those who want to play those games now. Personally I'd rather have more cores so if I were to upgrade from i5-3570K (not even overclocked), I would not want more than 4C/8T anyway. My biggest concern with the current Ryzen is actually their idle power which I feel could be lower especially since FM2 for instance used to have better idle than Intel.

The other thing I want from an upgrade is to go for a smaller case. The reason I am full-ATX currently is that someone gave me the Z77 motherboard so I bought the IB i5 even though Haswell had just been released. Normally I buy mATX, although am looking at mITX too. But really would like something way smaller than 20L otherwise there's no point.

Well the distinct lack of FO4 testing in many reviews is what I am hinting at - it used to be more common and sadly I play both FO4 and Skyrim:Special Edition.

In its current state Ryzen is not brilliant.




Now look at some forum testing.

Default Win = 47,5fps.
HT ON + lock one CCX = 49fps.
HT OFF + lock one CCX = 50fps.

6700k 4.5Ghz 3Ghz ddr4 80Fps
xeon 4030Mhz 2120DDR3 59Fps
2600k 4500mhz DDR3 1600mhz 54fps.

Broadwell server CPU at 2.5GHZ~3.3GHZ 58FPS

Area tested:Boston Commons

The problem is that once you start building settlements and using the contraptions DLC to make factories,etc it really starts pushing the CPU.

Mods can cause even more problems.

[kompukare]

CB has some details about Naples.
https://www.computerbase.de/2017-03/...pu-benchmarks/ (Goggle Translated)
4000+ pins (Xeon E7-8xxxx v4 is still LGA2011 according to the Wiki) which sounds rather big.
No estimate of die size, but some specs include 128 PCIe lanes, and of course it's a SOC.

No info on what they've done with the Infinity Fabric, but obviously the interconnect for each CCX has to be a lot faster than with Ryzen 7 or the whole thing will have huge cache misses etc.

The benchmark compared it to a Broadwell-E Xeon E5-2699 v4 which runs at 2.4GHz/3.6GHz Turbo and the Naples was a lot quicker.

Ars, Tom's, and ZDNet have that story too:
https://arstechnica.com/information-...ry-than-intel/
http://www.tomshardware.com/news/amd...ver,33819.html
http://www.zdnet.com/article/amd-pre...k-with-naples/

EDIT: about the die size. If Ryzen 8C is about 200mm², then Naples surely has to be close to 800mm². Which is huge. Wouldn't Infinity Fabric and the CCX clusters be suitable for use with an Interposer though? Huge yield difference between four 200mm² dies and one 800mm² monolith.

[Bagnaj97]

Well the distinct lack of FO4 testing in many reviews is what I am hinting at - it used to be more common and sadly I play both FO4 and Skyrim:Special Edition.

and if anyone's wondering why CAT seems more than slightly disappointed with Ryzen, this is your answer! Poor performance in FO4 would affect him greatly...

[scaryjim]

... No info on what they've done with the Infinity Fabric, but obviously the interconnect for each CCX has to be a lot faster than with Ryzen 7 or the whole thing will have huge cache misses etc. ....

Unless they're relying on the type of software run in server loads to schedule better than desktop loads - don't forget that in some workloads there doesn't seem to be any impact at all from cross-CCX communication, presumably because those workloads don't require cross-CCX communication.

Alternatively they might have made some tweaks between bandwidth and latency, so the bandwidth between CCXes might be lower, but the latency of fetch operations from other L3 caches might be improved. Hardware.fr tested this quite extensively, as did legitreviews and Ryzen 7's cache latency deteriorates past 4MB usage, and tanks completely once you go over 8MB usage (90ns - 100ns is main memory latency).

I don't think the bandwidth between CCXes is that much of a problem - it's hugely unlikely you'd ever need to pull the entire 8MB from one L3 cache in one go - but the fact that the latency between CCXes is at least as high as back to main memory means the second half of the L3 cache is pretty much useless in those scenarios. If they could even half the latency between L3s, then you'd basically be in a situation where each CCX had 8MB of L3 cache and 8MB of L4, which would massively improve the situation. No real way of knowing if that's possible, though....

EDIT: Just thought of something that puts context on this:

in the first 4MB - 6MB of L3 cache latency is around 12ns, meaning the CPU has to wait ~ 50 cycles for the data it wants to be ready.

In that last 8MB - 16MB of L3 cache, it has to wait > 350 cycles for the data. I suspect you can do an awful lot in 300 CPU cycles....

[CAT-THE-FIFTH]

and if anyone's wondering why CAT seems more than slightly disappointed with Ryzen, this is your answer! Poor performance in FO4 would affect him greatly...

TBF,about half that is having in it menus,or minimised when I walk off and do something else. I think real gameplay is around 800ish hrs,but that is spread over 18 months or so.

The other games I played are Planetside 2 which is similar in scaling too.

[kompukare]

Well, maybe the deal between AMD and Bethesda will yield something but currently it looks like FO4 doesn't really scale past 4C/8T:

Source:
http://www.pcgameshardware.de/Fallou...h-194-1220187/

Or even just 4 threads

Source:
http://gamegpu.com/rpg/%D1%80%D0%BE%...-pack-test-gpu

Skyrim SE is even worse:


Source: http://gamegpu.com/rpg/%D1%80%D0%BE%...ition-test-gpu
That 8C/16T Intel HEDT chip is pretty wasted there. And Ryzen 7 would be as well.
EDIT: No ignore the last line.
The 8C/16T Intel HEDT chip was totally wasted in the original, but is getting somewhat used now. But then the question is, where does the game put user scripts? Like the automated stuff in the user mods which Cat keeps mentioning. If they are all on one thread on one CPU that would become the limit.

EDIT: ops, grabbed the wrong screenshots from game. Fixed now

[CAT-THE-FIFTH]

Well, maybe the deal between AMD and Bethesda will yield something but currently it looks like FO4 doesn't really scale past 4C/8T:

Source:
http://www.pcgameshardware.de/Fallou...h-194-1220187/

Or even just 4 threads

Source:
http://gamegpu.com/rpg/%D1%80%D0%BE%...-pack-test-gpu

Skyrim SE is even worse:


Source: http://gamegpu.com/rpg/%D1%80%D0%BE%...ition-test-gpu
That 8C/16T Intel HEDT chip is pretty wasted there. And Ryzen 7 would be as well.

The worst thing is that the FO4 testing is not even the worst part of the game.

Once you start using the settlement mechanics,ie,start making big settlements,with many NPCs,factories,logic gates,etc it starts to really push the CPU.

Add mods to that and you can see the problem.

Its a crap engine for threading for sure.

Edit!!

However with mods its almost like Minecraft lite in some ways.

[kompukare]

Its a crap engine for threading for sure.

Look at my edit: I had looked and grabbed the wrong CPU shots from gamegpu.
But, yes Skyrim, Oblivion and probably Morrowind too (never ran too many mods back then) have had the same problem going back years. Of course, it is generally helped by the fact that vanilla TES games tend to be rather empty and bland so there's not much happening. Plus the AI scripting (RadiantAI or whatever they call it) was less advanced in the older games.

Love the open-ended nature of TES games, but going back to Daggerfall it has often felt like interactions was just a dull database of factions and rather predictable.

Alternatively they might have made some tweaks between bandwidth and latency, so the bandwidth between CCXes might be lower, but the latency of fetch operations from other L3 caches might be improved. Hardware.fr tested this quite extensively, as did legitreviews and Ryzen 7's cache latency deteriorates past 4MB usage, and tanks completely once you go over 8MB usage (90ns - 100ns is main memory latency).

Ah, thanks I hadn't seen that part of the Hardware.fr review nor the LegitReviews part. Hardware.fr being one of the more in-depth sites is not new to me, but the detailed testing LegitReviews did is.

Still, most of these things look like fairly easy fixes for Zen2 or in some cases even for the Zen APUs or Naples.

[CAT-THE-FIFTH]

Look at my edit: I had looked and grabbed the wrong CPU shots from gamegpu.
But, yes Skyrim, Oblivion and probably Morrowind too (never ran too many mods back then) have had the same problem going back years. Of course, it is generally helped by the fact that vanilla TES games tend to be rather empty and bland so there's not much happening. Plus the AI scripting (RadiantAI or whatever they call it) was less advanced in the older games.

Love the open-ended nature of TES games, but going back to Daggerfall it has often felt like interactions was just a dull database of factions and rather predictable.



Ah, thanks I hadn't seen that part of the Hardware.fr review nor the LegitReviews part. Hardware.fr being one of the more in-depth sites is not new to me, but the detailed testing LegitReviews did is.

Still, most of these things look like fairly easy fixes for Zen2 or in some cases even for the Zen APUs or Naples.

FO4 is much intense as a world when it comes to even Skyrim,and it has a degree of persistence which very few games have. For instance you can drop some loot or move a cup somewhere and the game will remember where you placed it for weeks.

The build mechanics are almost like Minecraft Lite especially when you add mods.

For instance at Abernathy farm I have a 23 story Institute facility which I built in one play-through where you could make synths.

Like with Skyrim,FO4 has a massive modding community with some mods being downloaded 100s of 1000s or millions of times.

Some are optimisations for worse systems and others make it look totally different.

Its why when people just poo-poo me when they say on forums,SB or IB level performance is enough,it really isn't for certain games.

Planetside 2 is another one.

I mean I certainly had no impressions AMD would match Intel,but its still no excuses for them to only be at SB levels.

I mean it is getting to stupid levels on OcUK forums,where people have just lost the entire plot and have forgotten that certain games showed huge improvements going from SB to SKL/KL too.

I am starting to realise when I joined Hexus why so many were so pro-Intel and anti-AMD,because it must have been horrendous during the Athlon 64 days prior to Core2,so I think it must have been pay back of some sorts. I certainly can't remember it being this bad when the Phenom II was released either.

The problem is I can see with so many pushing,single core IPC is not important but MOAR cores is,they are going to get a dose of that back once HEDT Skylake and Coffee Lake are released.

Intel will not only have an IPC and clockspeed advantage but MOAR cores too,so its going to be messy.

[kompukare]

Ah, the AT news item about Naples seems to have a bit of extra info compared to the others.

While not specifically mentioned in the announcement today, we do know that Naples is not a single monolithic die on the order of 500mm2 or up. Naples uses four of AMD’s Zeppelin dies (the Ryzen dies) in a single package. With each Zeppelin die coming in at 195.2mm2, if it were a monolithic die, that means a total of 780mm2 of silicon, and around 19.2 billion transistors – which is far bigger than anything Global Foundries has ever produced, let alone tried at 14nm. During our interview with Dr. Su, we postulated that multi-die packages would be the way forward on future process nodes given the difficulty of creating these large imposing dies, and the response from Dr. Su indicated that this was a prominent direction to go in.

Obviously packaging would be more expensive and the margins on such enterprise hardware should be a lot higher but in theory that would mean that this monster CPU would only cost four times the silicon cost of the current Ryzen 7 processors.

The next question is of course what's the max possible interposer size. And if it's a lot bigger than 800mm², then maybe they will offer some kind of Zen + Vega + HBM2 for the HPC market.

[CAT-THE-FIFTH]

Ah, the AT news item about Naples seems to have a bit of extra info compared to the others.


Obviously packaging would be more expensive and the margins on such enterprise hardware should be a lot higher but in theory that would mean that this monster CPU would only cost four times the silicon cost of the current Ryzen 7 processors.

The next question is of course what's the max possible interposer size. And if it's a lot bigger than 800mm², then maybe they will offer some kind of Zen + Vega + HBM2 for the HPC market.

AMD had 12 core and 16 core Opterons based on two Phenom II X6 and two BD/PD 8C chips,which were around 660MM2 to 700MM2 in size.

[flearider]

sorry cat you seem to be saying you like ryzen then rubbishrubbishrubbishrubbishting all over it .. ????
really cant see what side your trying to prove .. for a new chip it's great seeing as the 1700(£300) is quiet good and beats anything in it's price range

[CAT-THE-FIFTH]

sorry cat you seem to be saying you like ryzen then rubbishrubbishrubbishrubbishting all over it .. ????
really cant see what side your trying to prove .. for a new chip it's great seeing as the 1700(£300) is quiet good and beats anything in it's price range

Dude,seriously - where have I said its rubbish in everything?? I said it underperforms in certain games I play. You have an FX8350 - it will decimate your CPU. Games like FO4,Skyrim:SE,PS2,etc are played by loads of people including me - you must realise by now I am no Intel fanboi,but I have an IB Core i7 aleady. These are popular games.

AMD has been traditionally done badly in these games,and looking at the benchmarks,it looks like even an R7 1700 at 4GHZ won't beat it - you need to understand why should I buy an R7 now,if it cannot give me better performance in games I am playing currently??

What just to support AMD - my CPU was selling for only £175 in 2012,and don't you think me looking at a £320+ AMD CPU not doing any better is filling me with confidence. I could have got a Core i7 6700K from CEX,mini-ITX mb and RAM last year for not much over £320 but I decided to wait for Ryzen.

Now Ryzen appears and has worse performance - whats the point??

I can wait and wait,but there is no guarantee performance will get better in those titles.

AMD has even said they need better gaming performance so have sent out 300 kits to devs.

People need to look at this rationally - you could get a 6C/12T Core i7 5820K for as low as £248 18 months ago. These also overclock more than BW-E on average. In lightly threaded games it probably will perform better than Ryzen does in its current state.

I expected AMD would have less IPC than Intel for gaming,but not to this level.

Plus if the current level of performance is "good enough" then I might as well stick with what I have and wait for Ryzen 2 or Coffee Lake.

Edit!!

Look at my other usage,ie,image processing:

https://www.pugetsystems.com/labs/ar...rformance-910/

A SKL/KL CPU is more than competitive too - only if you export like 100 images at once an R7 1700X moves ahead(slightly).

I have not even mentioned that one before,but if you also want to know the other reason,then there it is.