I’m putting a Ryzen system together for evaluating the AMD work. It’s important to support the Intel competition. My system is a 1700 on a B350+ Asus , 128GB M2 SSD, 4 xGB DIMMs 2400. 8GB? this is a cheap eval. If I like it, I’ll stick 32GB in it.
Tried to install Debian 9 AMD64 net install. The installer chooses the wrong network driver . It’s an RTL 8168 etc. So I downloaded the Non-Free firmware Debian image, 3.4GB, and that installed just fine.
Ran up Vivado 2017.3 on it. Copied over a Windows project directory, and Vivado was not happy. All the MIF file paths are screwed up. For some reason, VIvado is on the windows install storing the full path including the drive name when pointing to the MIF files ? WTF Xilinx ??? so, there would be a script to write to fix that up. Still, the Vivado install worked like a charm – good one Xilinx ! It ran just fine, all the graphics and boxes all works as it should. Hooray Java.
OK, so that runs OK. Linux and Vivado and Debian will work for people.
Time for windows 10 pro. Installs OK. Geez there is alot of fluff now in a Windows install. Can’t we just have a Pro install without the cheese please MS ? Finally, after the endless reboots and updates I install Start-Is-Back, Opera, Ryzen Master , Vivado , Lasoo.
I run a few Vivado runs.
Compare box is Haswell… i5-4690k, turboing to 4 GHz, 32GB dual channel DDR3 1600 MHz and it benches Synthesis- 4:21, Implementation 5:29.
Still getting my head around The Ryzen Master Utility, and BALANCED/High performance mode etc. .
OK, Using ProcessLasoo to lock Vivado to Upper CCX cores 8-15…
general.Maxthreads=8 (was =4 for i5-4690k).
So 1st run 3.6GHz: 5:22 6:20. 3.6GHz (1.25V)
2nd run 3.8 GHz: 5:07, 6:32 (1.25V). slightly better synthesis.
3rd run 4 GHz : 4:47 , 6:11 (1.35V). The heat came up along way on this one (48 deg C) (ambient 23). synthesis still 9% faster , implementation 11% faster on Haswell
…another run at 4 GHz : 4:46 ; 6:10…….so that’s real.
Conclusion is that Zen slightly slower than Haswell, clock for clock ….approx even with 2400 versus 1600 memory.
The 2400 memory wont go any faster without overclocking voodoo which pro users are not going to bother with I think . Might get some 3200 memory and see how the benchmarks compare. I slowed he memory down to 2166, went about 8 seconds on synth and impl, or ~3% (for a 10% reduction- although the RAS/CAS timings got faster 14 14 14 instead of 15 15 etc)
On the other hand, the computer didn’t become useless when implementation really started cracking the whip – a full 4 cores, 8 threads left over to run other stuff, no problem.
Ran a mixed PRIME95 all thread stress test, with the supplied-with-the-CPU stock cooler, 1.35V, 3.9GHz, got a 46 deg C ambient rise.. Not bad. With a liquid cooler, would be no issue. Went to 4 GHz stress test- crash, did same with 1.4V no problem, but got too hot (>70). So there is room to experiment…