AMD RyZen CPU Architecture for 2017

AMD will really have to work on the plug and play aspect. Ryzen is rough around the edges, but with enough effort you can make it shine (mainly tuning your memory correctly). Just in a purely CPU limited scenario the difference you can see is absolutely absurd:

0a9bmd.jpg

1nha6w.jpg


Of course this is exaggerated by the fact that I'm using an overclocked 1080 Ti to run what looks to be sub 720p res (creating a purely CPU limiting scenario), but it just shows how limited Ryzen is with a non-optimal memory setup (2666 MT/s with Auto timings). If AMD manage to make Ryzen perform more consistently across many different setups that'll be the real winner imo.
 
AMD used HBM on the wrong product line.
Instead, it should have been strapped together with Ryzen, while the DIMM slots would to be repurposed to host some uber fast NVMe storage. :p
 
Imagine EPYC with HBM2 chips on it as well. It would take up one half of the motherboard.
 
AMD will really have to work on the plug and play aspect. Ryzen is rough around the edges, but with enough effort you can make it shine (mainly tuning your memory correctly). Just in a purely CPU limited scenario the difference you can see is absolutely absurd:

0a9bmd.jpg

1nha6w.jpg


Of course this is exaggerated by the fact that I'm using an overclocked 1080 Ti to run what looks to be sub 720p res (creating a purely CPU limiting scenario), but it just shows how limited Ryzen is with a non-optimal memory setup (2666 MT/s with Auto timings). If AMD manage to make Ryzen perform more consistently across many different setups that'll be the real winner imo.
Whats the easier way to optimize this timings? I got a pair of cheap modules (https://www.amazon.com/gp/product/B01DPZVP3W/ref=oh_aui_detailpage_o02_s00?ie=UTF8&psc=1) and want to extract as much performance out of them as possible but with my work I don't have much spare time.
 
Imagine EPYC with HBM2 chips on it as well. It would take up one half of the motherboard.
If you look at a decapped epyc or threadripper you see there's monkeyloads of room under that giant heatspreader. The zephyr cores are actually quite small comparatively. It's not impossible to imagine that one stack of HBM per zephyr might be possible, and the interposer required would be rather small too...

While we're at it making up imaginary ryzen features, could we add AVX512 and transaction-based memory handling too? :p (And not artificially segment these features, as Intel always does.)
 
Prime.png
This is a learning experience for me, never really before have I delved this far into overclocking a CPU, or memory. I'm working on seeing how far I can bring down the voltage and keep it stable, then I'll go back to playing with the memory. The image above is running Prime95, used Ryzen Master to monitor temps.
 
Last edited:
Whats the easier way to optimize this timings? I got a pair of cheap modules (https://www.amazon.com/gp/product/B01DPZVP3W/ref=oh_aui_detailpage_o02_s00?ie=UTF8&psc=1) and want to extract as much performance out of them as possible but with my work I don't have much spare time.

Better use the timings The Stilt provided for Hynix dimms: http://www.overclock.net/t/1624603/rog-crosshair-vi-overclocking-thread/24130#post_26242714

Just as a base line. If you can get B-die memory, it's a lot easier to get it working properly.
 
If you look at a decapped epyc or threadripper you see there's monkeyloads of room under that giant heatspreader. The zephyr cores are actually quite small comparatively. It's not impossible to imagine that one stack of HBM per zephyr might be possible, and the interposer required would be rather small too...

While we're at it making up imaginary ryzen features, could we add AVX512 and transaction-based memory handling too? :p (And not artificially segment these features, as Intel always does.)

By Zephyr you mean Zeppelin, right?
 
Forget cheap modules and get something with Samsung B-dies
Thats not possible right now, my mom is sick and I simply don't have money for anything more than a 60 bucks ram. If it was for me I would buy the damn flagship of 4600MHz.

In my particular case I am aiming for a zen2(or ryzen refresh) mid to high-end update and in the mean time just have something "usable".
 
I feel as if my DDR3 @ 2400 is a tad pedestrian these days... It wasn't the fastest DDR3 exactly when I bought it, but it was near the peak of price/performance curve at the time amongst enthusiast DIMMs, clearly faster than average memory at the time, yet they didn't cost an arm and a leg either. (Also, Corsair Dominator heatspreaders look good! :p)

Today, enthusiast DDR4 is WAY faster than this - in bandwidth anyway. Latency, well, we know DRAM scales very poorly in that regard...
 
I have a set of six OCZ 1600MHz 7-7-7-24 DIMMs in my older PC... :p

Anyway, counting latency in cycles isn't so interesting unless one also factors in clock speed. It's better to go straight to nanoseconds... (Where DRAM has remained relatively stable, latency-wise for quite some time I believe.)
 
I have a set of six OCZ 1600MHz 7-7-7-24 DIMMs in my older PC... :p

Anyway, counting latency in cycles isn't so interesting unless one also factors in clock speed. It's better to go straight to nanoseconds... (Where DRAM has remained relatively stable, latency-wise for quite some time I believe.)
Actually on average I think latency/cost is getting a little worse each ddr gen. If you look on the high end of each gen we also see worse latency.
 
Actually on average I think latency/cost is getting a little worse each ddr gen.
Well, burst length keeps going up, because DRAM can't be clocked too high so they gotta keep making the internal architecture wider and wider to increase bandwidth. So you'd expect a bit more latency, I'd think. But it's not too bad - or at least not yet. Overall performance is still increasing.
 
Back
Top