26-Way Intel/AMD CPU System Comparison With Ubuntu 16.10 + Linux 4.10 Kernel

[duby229]

I'm a bit confused by what you mean by this. How does the fabrication process affect the cache latency? The latency of the cache was the issue with Bulldozer through Excavator.

It's not obvious? https://en.wikipedia.org/wiki/Standard_cell Every fabrication process has a library of cells designers can use to build their products.

[nomadewolf]

In all honesty, those old AMD CPUs are not as bad as they are projected to be. We also keep forgetting that Intel has had a process node / technology (finfet) advantage for quite a long time now. Most users may not even notice a difference between the two under normal day to day tasks. Games are a different story, but with AMD's push of Vulkan and their gaming consoles, they're also slowly warming up to the idea of multi-core rendering. But I can't wait for Ryzen though. Especially because it will mean more competition in the high-end computing space. More competition will lead to more awesome technologies and better prices, and that's gonna be good for us consumers. All in all, 2017 seems to be very exciting! Thanks for the round-up!

AMD's error was thinking that it could direct the market simply by showing the way.
Even though AMD was better at multithreading at the time and that should've been the way the market progressed, Intel just used it's market share to steer the market in a different direction (IPC). Since it's much easier to program for single thread and the majority of the software developers continued to do same they always did, and we got to this situation...

[kram]

Don't cheer too early: Zen does not come with AVX2. Nothing big when it comes to gaming, but quite the difference for server and compute workloads.

Regarding the performance difference between Haswell and Skylake with Darktable: Skylake has a performance penalty for transitions from AVX to SSE modes. It's my speculation that Darktable mixes those, because in an image editing program you'd like to use AVX2, and mfpmath=both or =387,sse is a common default setting with most distros.

[M@GOid]

AMD's error was thinking that it could direct the market simply by showing the way.
Even though AMD was better at multithreading at the time and that should've been the way the market progressed, Intel just used it's market share to steer the market in a different direction (IPC). Since it's much easier to program for single thread and the majority of the software developers continued to do same they always did, and we got to this situation...

Exactly. At that time they started working at Bulldozer it looked like every simple calculator app was gonna be multithreaded, but very few programmers jumped in that wagon, so when Bulldozer came nobody had converted their software. Only a few niche programs did it, but the big names in the market, the ones that make hardware sales, did not.

Even in the game market, a market capable of making people upgrade their machines almost annually, are a sad thing to look in the multithread camp. I can only remember CryEngine (Crysis) and 4A Engine (Metro) as examples of engines that can tax equally all the cores you trow at them. The rest barely touch more than 2 cores in many cases.

[LinuxID10T]

It's not obvious? https://en.wikipedia.org/wiki/Standard_cell Every fabrication process has a library of cells designers can use to build their products.

I have some serious doubt AMD used Samsung's standard libraries. Work on Zen started way before the Samsung/GloFo deal happened. Furthermore with Jim Keller leading the design I even further doubt they used them.

[duby229]

Exactly. At that time they started working at Bulldozer it looked like every simple calculator app was gonna be multithreaded, but very few programmers jumped in that wagon, so when Bulldozer came nobody had converted their software. Only a few niche programs did it, but the big names in the market, the ones that make hardware sales, did not.

Even in the game market, a market capable of making people upgrade their machines almost annually, are a sad thing to look in the multithread camp. I can only remember CryEngine (Crysis) and 4A Engine (Metro) as examples of engines that can tax equally all the cores you trow at them. The rest barely touch more than 2 cores in many cases.

Yep, that's exactly right. What most engines do is they spawn a helper thread to split off loads from the main thread. It can help, it really does, but it doesn't scale to the number of loads CMT type architectures are designed for.

[duby229]

I have some serious doubt AMD used Samsung's standard libraries. Work on Zen started way before the Samsung/GloFo deal happened. Furthermore with Jim Keller leading the design I even further doubt they used them.

I'm sorry, perhaps I should have explained that not all fabs are designed to be fully compatible. It's very possible the place and route tools simply won't allow the older cache architecture to work. I very strongly suspect that indeed AMD had no other choice but to reimplement the cache design. Of course I should have included the disclaimer that it is no more than suspicion as of right now.

edit: The reason I suspect this so strongly is because AMD's cache cells as implemented currently are physically freakin' huge, literally ginormous. I very seriously doubt any other process would allow for such aberrations. At the very minimum I hope AMD had no choice but to redesign their cache architecture.

[mlau]

Don't cheer too early: Zen does not come with AVX2.

Of course does it have AVX2, even the lastest Excavator (Carrizo) does have it. You probably meant AVX512? That's not going to be relevant to the consumer space as long as Intel keeps it a Xeon-only feature.

[VikingGe]

The rest barely touch more than 2 cores in many cases.

That's not true anymore. Such games still exist, sure, but among the games I bought since 2015 I cannot think of a single one that doesn't scale decently across cores - in fact, this is pretty much the only reason my old Phenom II X6 is still capable of running modern games at all. I often see it sitting at >90% load when playing stuff like Battlefield 1, I'd expect much worse performance if it was only an X4.

Anyway, looking forward to Ryzen benchmarks - in the hope that those things are going to be released any time soon. Kaby Lake on the other hand will probably be rather boring for the most part, since Intel hasn't really changed much since Skylake.

[LinuxID10T]

I'm sorry, perhaps I should have explained that not all fabs are designed to be fully compatible. It's very possible the place and route tools simply won't allow the older cache architecture to work. I very strongly suspect that indeed AMD had no other choice but to reimplement the cache design. Of course I should have included the disclaimer that it is no more than suspicion as of right now.

edit: The reason I suspect this so strongly is because AMD's cache cells as implemented currently are physically freakin' huge, literally ginormous. I very seriously doubt any other process would allow for such aberrations. At the very minimum I hope AMD had no choice but to redesign their cache architecture.

The libraries AMD used also changed significantly for Carizzo and it shrunk them significantly. My understanding is that Ryzen was a clean slate design putting performance and power efficiency first. I would assume that would mean the libraries are custom for Ryzen. This is how Jim Keller got so much out of Apples A series processors.