New Ryzen Processors and VRay

[JD3D_CGI]

So AMD have announced their new line of Ryzen processors

http://www.kitguru.net/components/cp...-pricing-hard/

I was wondering, have Chaos had any contact with AMD and tweaked anything?

I remember having a bad time a couple of years back with the 8350 using some dodgy trick where it shared maths coprocessors between cores or something, which gave me an exceptionally slow (about half) speed renders from Vray.

As a result of that I'm pretty reticent to use their processors again - I was just wondering what you guys know?

[vlado]

I was wondering, have Chaos had any contact with AMD and tweaked anything?

Nope, AMD have not contacted us regarding their CPUs.

I remember having a bad time a couple of years back with the 8350 using some dodgy trick where it shared maths coprocessors between cores or something, which gave me an exceptionally slow (about half) speed renders from Vray.

We can only hope that things will be better

As a result of that I'm pretty reticent to use their processors again - I was just wondering what you guys know?

We'll have to get our hands on these CPUs to be able to say anything for certain.

Best regards,
Vlado

[vlado]

Ok, a small correction: it looks like we might be getting a Ryzen system for testing in our LA office. I will let you know how it goes.

Best regards,
Vlado

[hardrock_ram]

Vlado:
One thing I have been missing from you guys is some benchmarking.
You have previously recommended running Vray on machines with fewer cores (<16) and rather use more nodes rather than a few monsters with many cpu`s.

The thing is it is obviously much less time consuming to administrate those huge monsters than many small nodes. Its usually pretty easy to get money for equipment, but not so much for hiring of additional personell, which is why this is important.
I have a farm with a mix of computers, some with 2 cpu sockets and some have 4. And contrary to what you have recommended, Vray runs very well on machines with 4 sockets and many cores (85-95% of theoretical performance). Those are Xeon E5 with two QPI-links, and I suspect it could get even closer to 100% theoretical perf with E7s that have three links.

I understand where you are coming from when you recommend "smaller" nodes, since you get more stable performance. I guess plugins that generate shitloads of data can slow down 4-socket machines for example?
Anyway, some representative benchmarks for a few different setups would be great, as we dont have to buy in blind

[vlado]

Yes, I know that a benchmark is important and it's almost done actually. One of the last remaining parts is to hook up the submission part to the website...

Best regards,
Vlado

[Macker]

Any news on the benchmark?

[vlado]

Not yet, but there are some Ryzen tests with V-Ray here:
http://www.hardware.fr/articles/956-...ray-v-ray.html

Best regards,
Vlado

[Alex_M]

Not yet, but there are some Ryzen tests with V-Ray here:
http://www.hardware.fr/articles/956-...ray-v-ray.html

Best regards,
Vlado

Looking good! Speeds look the same and even a bit better than 6900k. We can safely say that R7 1800X is a 6900 killer with that price.

Here I want to make a very important point. It's very likely that it would have got even better bench results if they had manually overclocked 1800X to 4Ghz. You'll ask "why when it has a turbo boost of 4 Ghz"? The answer is, if you don't manually put all cores at 4 Ghz, by default Ryzen runs at 3.7 Ghz on 7 of the cores and turbo's just one to 4 Ghz when all are under full load. Similarly, if half of the cores are under full load, 6 cores will run at base clock with the remaining 2 at turbo speed.

I speculate that this is to keep the Ryzen CPU within the healthy TDP range when it's at 100% load on all cores. If all cores turbo'ed to 4 Ghz, the TDP and heat would be much higher so you would need a good cooler to keep it cool when all cores are at 4 Ghz. This is something that motherboard manufacturers can tackle so it turbo's all cores automatically. Intel processors are the same - by default they too boost just one core under full load but motherboard manufacturers have added a feature to enable their boards to boost all cores regardless of how many of them are under full load. Maybe in the very near future we'll see such motherboards for Ryzen chips too so we won't have to manually put them at 4 Ghz.

So if the testers from that review had put the 1800X manually to 4 Ghz before benchmarking, I would expect 10% better results for 1800X (3.6 base clock +10% = 4Ghz). This would translate to give or take 278 seconds for Ryzen in that V-ray test they did.

So if anyone gets a 1800X, please do a Vray bench without any manual overclock first (let Ryzen boost by itself) and then do a bench with manual overclock on all cores to 4Ghz so we can see the difference. Should be around 10% faster.

[vlado]

Just got my hands on our Ryzen system, we'll see how it goes.

Best regards,
Vlado

[tinybot]

Very interesting! I'm just now in the process of investing in new hardware for a render node, and have my eyes on building a Ryzen 7 1800 system! This news from yesterday had me concerned:

http://www.digitaltrends.com/computi...ix-fma3-crash/

As with any new platform there are always trade-offs for jumping on to something so new, but the price for performance seems too good to resist! Looking forward to your thoughts!

(Also... http://www.digitaltrends.com/computi...ryzen-16-core/ )

[Morbid Angel]

it runs at 3.6 ghz! that's pretty crazy, does it have a turbo clock like xeon?

[Alex_M]

Nice! Waiting for some bench scores and first hand impressions.

If I were to build a render node today, 1700 would be my top choice. It's the best bang for the buck CPU on the market right now. Not to mention enthusiast overclocking AMD motherboards are twice as cheap than enthusiast X99 boards.

If you know how to overclock, you can easily get 1800X performance out of 1700 and save $170 for a motherboard or other parts. I've seen many people running 1700's fine at 4.0 Ghz so basically the difference in the 7 lineup is the default base and turbo clock speeds which they run at. We are yet to see if AMD is doing any binning for the different R7 models. So for anyone who know how to overclock I would say skip the rest of the R7's, get the 1700, overclock it and be happy with the performance for the small amount of money you've spent compared to an Intel system. At $330 USD, an overclocked 1700 matches and surpasses in performance 6900k, a $1000-1100 CPU. If my math is correct, an entire 1700 system complete with overclocker motherboard, RAM and PSU will cost you the same as a 6900k CPU.

[glorybound]

Start building them, I am sure you'll sell them quickly. I'll be your first customer.

Nice! Waiting for some bench scores and first hand impressions.

If I were to build a render node today, 1700 would be my top choice. It's the best bang for the buck CPU on the market right now. Not to mention enthusiast overclocking AMD motherboards are twice as cheap than enthusiast X99 boards.

If you know how to overclock, you can easily get 1800X performance out of 1700 and save $170 for a motherboard or other parts. I've seen many people running 1700's fine at 4.0 Ghz so basically the difference in the 7 lineup is the default base and turbo clock speeds which they run at. We are yet to see if AMD is doing any binning for the different R7 models. So for anyone who know how to overclock I would say skip the rest of the R7's, get the 1700, overclock it and be happy with the performance for the small amount of money you've spent compared to an Intel system. At $330 USD, an overclocked 1700 matches and surpasses in performance 6900k, a $1000-1100 CPU. If my math is correct, an entire 1700 system complete with overclocker motherboard, RAM and PSU will cost you the same as a 6900k CPU.

[Jeffrey]

Latest rumor: AMD has a 16 core Ryzen cpu in development:
http://www.guru3d.com/news-story/doe...the-works.html

[Alex_M]

Start building them, I am sure you'll sell them quickly. I'll be your first customer.

Why wait? Just buy the parts and put it together yourself. Or ask for a friend or the PC store to do it for you.