Zen cores actually have half sized SIMD units compare to Intel. They will execute AVX instructions, but for programs that use AVX well, Intel should win. Using AVX at all is still rare, using it well even more so, but Intel's AVX SIMD performance is pretty stellar.
Yeah, the execution unit differences are really complicated.
Ryzen has four ALUs plus four partial vector units. These vector units are all 128 bit. Two can add and two can multiply. An adder and a multiplier can combine to do FMA.
Skylake has four math ports total, all of which can do basic ALU operations. Two of them have 256 bit vector capabilities, and each of those can add or multiply or FMA. A third has limited vector capabilities.
10 core (and higher?) Skylake-X chips upgrade those limited capabilities into an entire new 512-bit vector unit, also capable of FMA.
Depending on your exact needs, Ryzen can be slower or faster than Skylake, but it's usually slower on programs stuffed full of vector math, especially FMA.
Presumably, as a brand new architecture, there's more chance of AMD improving single-core performance (either through frequency or IPC) in the next iteration of Ryzen than Intel doing the same with Core i3/5/7?
Clock speed increases are a likely benefit we'll see when Zen 2 is released on Global Foundries 7nm+ process, right now that's the biggest reason Zen seems to fail against Kaby Lake in single-threaded performance minus some minor IPC differences. A Ryzen 7 1800X tops out at 4GHz boost (it can usually run stable at 4GHz overclocked with a vCore bump, but you're playing the silicon lottery) meanwhile a Core i7-7700K has a base frequency of 4.2GHz - a 5% difference, ignoring the 7700K's 4.5GHz boost frequency.
On a similar note, it looks like fab processes are likely hurting AMD with their GPU's. From what I can gather they opted to produce Vega at GloFo's fab on their 14nm proccess, which might explain why they are shipping the air-cooled Vega 64 with a 1.24GHz base clock at a pretty ludicrous 295W TDP.
I'm impressed too, if only it wouldn't have those nasty mce and segfault bugs, and AMD could actually explain what's going on after all this time testing it.
Segfaults happen under heavy compilation workload - nothing to do with PHP specifically, for instance compiling Mesa produces it. And mce bug means random freezes and reboots, when CPU core wakes up from some sleep C state.
It is a roadblock, if you need to build something from source and it segfaults. Can be quite useful for gaming too (building Mesa, Wine, open engines and what not).
Besides, mce bug is more critical, since it actually makes the whole system unstable.
These bugs aren't consistent, and likely affect only part of Ryzen CPUs of the same model, not all of them. That's why the results are so over the place for different people. It also can be affected by power setup, which adds a lot of variables and makes it harder to reproduce.
By the way, for users or Gentoo or other distros which actually build packages when installing them, these segfaults are a very major roadblock for any kind of use case.
There are however reports of segfaults during compilations that seem more Ryzen-specific, though even there GCC versions were said to play a role. It is all a bit obscure currently.
You can make a test that will segfault everywhere. We aren't talking about those, but about regular workloads, like compiling Mesa for example. If Mesa compilation segfaults everywhere, something is really wrong.
I like my Ryzen 1700X which I use as a home server. But after overclocking my 5960x runs most software faster. The Ryzen CPUs seem to have some kind of hard speed limit right around 4 GHz. My 5960x can go to 4.3 although I keep it at 4.1 normally. That's 4.1 GHz on all cores at full load, not boost clocks.
But yeah, AMD is cheaper. And if you compare it to a 5960x at stock 3.3 GHz clock, it is faster.
I got a 1700 because it uses 65W instead of 95W but still has eight cores.
There are some minor glitches with the system, but as far as performance on CPU-bound workloads is concerned, I am very, very happy with it. When I look at what comparable chip from Intel would have cost me, I am even happier.
I also have a 1700 and with a $100 closed loop liquid cooler was able to clock it up to 4ghz without any trouble at all. Obviously it draws more power that way but it's nice to know the option is there if you ever want more out of it.
CPU/GPU benchmarks always look like complete ass, I've considered just designing a simplified design template people can use on websites or in video reviews.
Digital Foundry's frame rate and frame time graphs are particularly egregious.
I had the same reaction immediately as I opened that page. Without even known the context, I could have guessed this was a Linux related site. It's as if the entire Linux ecosystem consistently wants to present the ugliest interfaces to their end users. I don't get it.
Take any tests with scaling governor "powersave" using the intel_pstate with a grain of salt. This powersave mode has nothing in common with the acpi_cpufreq's version of powersave and is quite aggressive at clocking down the CPU, even in periods of 100% CPU usage (when it decides that the throughput benefit of remaining highly clocked isn't worth it). This happens especially on memory-bound codes.
Despite having the same governor name, "powersave" with acpi-cpufreq has no relationship. It's using low-frequency software control and generally ramps up to 100% frequency (highest pstate) once CPU-bound code starts running and stays there.
Sounds like the perfect option for people in the market for a new gaming machine. You absolutely don't need top of the line performance in CPU anymore. My six year old Xeon X5675 is still killing it, which only recently replaced an I7 920 from 2008 on the same motherboard, and RAM. Besides swapping out video cards every couple of years, this thing maxes out anything at 2560x1600.
I'm blow away by the fact that nobody here is calling out the shit setup for testing. E.G. the Intel rigs have 16 gigs of RAM, but the 8370 rig has 8 gigs. Wat? They don't standardize RAM, SSD, or anything across platforms. Shit benchmarks.
Not an option until they fix Linux crashes and hangs. Threadripper would be pretty neat for a 4 GPU deep learning server, but I've had nothing but trouble with plain Ryzen
38 comments
[ 3.0 ms ] story [ 88.5 ms ] threadRyzen has four ALUs plus four partial vector units. These vector units are all 128 bit. Two can add and two can multiply. An adder and a multiplier can combine to do FMA.
Skylake has four math ports total, all of which can do basic ALU operations. Two of them have 256 bit vector capabilities, and each of those can add or multiply or FMA. A third has limited vector capabilities.
10 core (and higher?) Skylake-X chips upgrade those limited capabilities into an entire new 512-bit vector unit, also capable of FMA.
Depending on your exact needs, Ryzen can be slower or faster than Skylake, but it's usually slower on programs stuffed full of vector math, especially FMA.
I think I got that mostly correct.
On a similar note, it looks like fab processes are likely hurting AMD with their GPU's. From what I can gather they opted to produce Vega at GloFo's fab on their 14nm proccess, which might explain why they are shipping the air-cooled Vega 64 with a 1.24GHz base clock at a pretty ludicrous 295W TDP.
It outperforms the i7 5960 and it's similarly less than half the price.
AMD can compete well in the enthusiast-on-a-budget segment of the market.
Oh well.
Either way, IMO it's not a roadblock for office and gaming usage.
Besides, mce bug is more critical, since it actually makes the whole system unstable.
On the other hand, the compile thingy has never been an issue yet, despite me compiling a lot of things. (Ryzen 7)
And for gaming, I've never had a game crash, I do doubt that many people really game under Linux yet, most gaming happens on Windows.
Office use is not roadblocked at all.
By the way, for users or Gentoo or other distros which actually build packages when installing them, these segfaults are a very major roadblock for any kind of use case.
There are however reports of segfaults during compilations that seem more Ryzen-specific, though even there GCC versions were said to play a role. It is all a bit obscure currently.
But yeah, AMD is cheaper. And if you compare it to a 5960x at stock 3.3 GHz clock, it is faster.
[1]https://www.eteknix.com/memory-speed-large-impact-ryzen-perf...
There are some minor glitches with the system, but as far as performance on CPU-bound workloads is concerned, I am very, very happy with it. When I look at what comparable chip from Intel would have cost me, I am even happier.
That was a pleasant surprise. The fans Intel ships with its CPUs are ... not very good.
Digital Foundry's frame rate and frame time graphs are particularly egregious.
See for example https://stackoverflow.com/a/45201673/149138 and https://stackoverflow.com/questions/43343231/enhanced-rep-mo... - both of which show large performance variance between powersave and performance. In the first example, the CPU clocks down indefinitely to 1.1 GHz from a nominal 2.6 GHz despite being 100% CPU-bound, and performance is cut in half.
Despite having the same governor name, "powersave" with acpi-cpufreq has no relationship. It's using low-frequency software control and generally ramps up to 100% frequency (highest pstate) once CPU-bound code starts running and stays there.