PCPer has an hour with David Kanter discussing AMDs next CPU arch. Fresh off the press :-)
https://www.pcper.com/reviews/Processors/Dissecting-AMD-Zen-Architecture-Interview-David-Kanter
If you don't know who Kanter is, he's been around a while. He wrote an article on AMD ZEN back in August:
http://www.linleygroup.com/mpr/article.php?id=11666
Youtube:
Anyone care to summarize the video? I don't have the time to watch it.
It's a high level chat on Cpu architecture in general and the landscape as it is now, and going back 20 years, very sweeping talk. Good to watch, would have been a better audio cast though; though ppl mentioned this in comments and Ryan suggested they would look into it.
They talk a lot about the inner workings (what is known so far) of ZEN and compare to Intel Skylake client cores. Kanter also talk a bit about diffrences in Client loads and Server etc. He is actually quite good at explaining complicated subjects. For example, Kanter talks a bit about the differences in how the caches work in Zen and how it is different from Skylake.
I really hope they do another interview after Zen has been launched and we have actual performance numbers.
So @Cavemanthe0ne you say you know better when pretty much everyone says the Blender benchmark is a cherry picked one by AMD?
Kanter thinks Ryzen will have about Haswell IPC generally, and about half when comparing FPU perf (AVX-256) and no AVX-512. Maybe Naples will have AVX-512 (Skylake EP is planned to have it) and maybe not just half as fast AVX 2, but we'll see. Now, blender doesn't use AVX at all IIRC per default, you need to set the flag at compilation. People have compiled Blender with AVX 2.0 and Skylake does really well then, Ryzen can probably not keep up at all as it needs two full cycles to do AVX 2, Skylake (beginning with Haswell) is about twice as fast as it does one operation per cycle.
So of course it is a cherry picked benchmark. AMD would have been stupid to compile blender with AVX 2.0 and then compare it to Broadwell-E. It would have been humiliating. Buuut it also shows that Ryzen has a good FPU for general tasks and AVX-1 (128 bit operations). So if you don't really need AVX 2.0, you'll be good. And you can do it if it is really needed, albeit at a slower pace. So you get die space savings and power savings with Ryzen and consumers doing desktop stuff doesn't really need AVX 2 to run Word etc. Kanter seems to agree with me on this latter part, AMD has not tried to take on Intel head on.
Going on, stuff from the interview and my own thoughts (@1920.1080p.1280.720p a bit of a summary too ):
It looks like the ordinary Integer operations will be good, and arguably they are already pretty good in Excavator. ZEN seems to improve on this, front end looks good etc. Integer pipelines are more similar to Skylake, it's not the same but more comparable than the FPU design. Kanter touch on how Intel probably has an advantage in loads like Linpack where you move around a lot of data, but then again Ryzen is a Desktop CPU so it will probably not matter as much. Kanter is much happier in the design of ZEN compared to Bulldozer that he thinks had several bad designs. Annd I think most people agree there.
Caches are discussed in the video, and while AMDs cache layout is a little different, it might not be much worse than Intels. The benches seems to suggest it isn't too far behind. This is really good, as AMD has been almost garbage at the cache level compared to Intel. Or you could say that Intel caches are God-like, but whatever. ZEN has a micro-op cache which is a first fro AMD - very good. Also AMDs caches seems to be smaller in size as in die area? Kanter was at the ISSCC where AMD showed that their cores are smaller compared to Intel (not strange as AVX capability is less) but also that their caches are smaller. That is weird, and Kanter can't really answer that, more than AMD uses some "interesting measures". Intel is the king of making super dense SRAM, so it is notable if AMDs caches are smaller on Samsung 14nm. Kanter also doesn't think that the cache design is bad in any way, it is different than intel, but not a bad design at all. I've seen som discussions about the L3 victim cache and that is Very Bad for some reason, but Kanter explains this and talks a bit about pro and con. Also, this needs testing ofc to work out the specifics.
They talk about power savings too a bit. Again it comes down to comparing Client application to Server. Broadwell-E is a server part first, and ofc power savings are important on server, but also throughput is important. Still AMDs TDP numbers are impressive according to Kanter and will be interesting to test in a couple of weeks.
Fabs, fabs are talked about. Ofc Intel has a huge advantage running their own Fabs, Kanter has some insight on that. AMD on the other hand has worked hard to be able to use pretty much any Fab it seems, using standard tools etc. Sure, they can't do some of the kinda insane Intel silicon process things, but they can do other stuff. Intel control their Fabs so they work hard on process tech, while AMD needs to think about what Fab companies like GloFo can deliver and work with that. Intel always release small chips on an new process, the large server and E parts come about a year later when problems have been worked out. AMD did something similar last year when they released Polaris on the new-ish Samsung 14nm. They probably have worked out the problems and ZEN/Ryzen will have avoided some pitfalls thanks to that. Yields are probably significantly better now compared to a year ago too.
In closing Kanter thinks Intel will have the single core performance crown. Most people agree on this, and I don't think AMD will be able to match Kaby Lake clock for clock, much less a Kaby at 5+ GHz. On the FPU side the gap will be bigger if you want to use the bigger vector instructions. Integer looks better. Also clocks, how good ZEN/Ryzen clocks will ofc be very important. I'm super curios about this. I don't think that "auto overclock" thing will be significant, more like 10 MHz here, 50 there. But if I can overclock Ryzen close to 5 GHz under water, it will be a smackdown I think. Sure a 7700K will be faster overclocked to 5 GHz in single thread, but if Ryzen is kinda close I won't care. It is about price and performance, not ultimate maximum perf for me. Plus I kinda love to play with overclocking, lol.
So march will be interesting. Let the benchmarking begin :-D
Soo
You made a post that agrees with me, tries to prove you are better than me, and then also disagrees with me.
What is the point of it?
Seriously.
What? I'm interested in your input. Ofc you can't say anything if you're under NDA.
You took something you were going to say in another thread that you knew you couldn't and moved it over here because you knew you could post it here
Also
They picked blender
BECAUSE ITS BLENDER
and it happens to show the things Ryzen does well, like the guy who Was doing the rendering demos
Idk what they did with the streaming demo though that one seems odd considering I can stream 1080p60 on my FX 8370 fairly well, especially on a game like LoL or Dota 2
In gaming I'm going to guess
Guess mind you
That Ryzen probably is going to end up at the performance of a 6900k on the top end
Why you even are comparing it to single thread of 7700k is beyond me lol, and considering the fact that kaby lake overclocks well its not a fair ipc ccomparison
I'm afraid I'm not following you. If you're upset over something I wrote I'm sorry?
I responded in this thread 'cause it would be on topic here and I find it an interesting topic. It also relevant to the PCPer interview with Kanter as stuff like IPC is discussed.
People who use Blender a lot tend to compile it with relevant optimizations for their systems. I'm a programmer so it is second nature to me.
The comparison with a 7700K is kinda important. Single thread performance is still very important in games. So a 7700K overclocked will probably be, if not the best CPU in most games, very close to the top. Remember that according to the silicon lottery, over 50% of 7700K parts overclock to 5+ GHz. Plenty points to Sky/Kaby Lake having better IPC in general and in float ops in particular (up to twice as fast).
Ofc this is all speculation from the information available. I would love it if Ryzen is fantastic and blows everything out of the water.
I'm basing my thoughts on the matter entirely on what I saw when I was at the New horizon event. I am 100% ignoring anything else because I don't care about BS lol, so if you are taking into account any of the leaked infos... which are mainly WCCFtech spewing crap... meh.
I just am stating what i know from first hand experience and i don't understand when people argue with that :P
If i get any new info i'll go off of that but for now that's what i have and any thing more is pure speculation on my end and on everyone elses, which I don't care about :P
Well there is nothing else to do but speculate while being excited for the Ryzen launch, what am I supposed to dooo?!?
:-)
There were some Passmark scores leaked (or on purpose, idk). Integer looks good as assumed. Memory perf and in the extension cache looks a bit on the low side though. But as it is a "leak" we don't know the caveats, might have ran in single channel for all we know. Still I think memory overclocking will probably be important, it kinda was on AM3+ too. I had a much better experience on my Sabertooth AM3+ board when I overclocked the FSB to 300 and the RAM to 2333. Memory and cache is probably considerably better on ZEN compared to Bulldozer/Piledriver/Excavator, but might not catch up to Intel. But then again, it probably doesn't have to if it is close otherwise.