Ryzen 3000 CPUs and Core_a7 (and Core_21)

[JimboPalmer]

Core_a4 does most of the math using SSE2 registers, present on pretty much all CPUs since 2004 including Pentium 4 (2000) and later and Athlon 64 and later. https://en.wikipedia.org/wiki/SSE2

I am not expecting a4 to get any faster on the Ryzen 3000 Desktop CPUs.

Core_a7 has two modes, SSE2 and AVX2. The software chooses mode based on the CPU, AVX2 was introduced by Intel with Haswell (2013) and implemented by AMD with Excavator in 2015. All Ryzen CPUs can do AVX2. https://en.wikipedia.org/wiki/Advanced_ ... tensions_2

Current AMD CPUs break down the 256 bit registers by using a 128 bit math unit. so you do 128 bits, remember the carry, and then do 128 bits with the carry you got the first time. 3 micro instructions to do 1 instruction. The new Ryzen 3000 Desktop CPUs due this year, will also do AVX2, instructions, but will use a 256 bit math units (accumulator) so they are much faster at SIMD math. Now the rest of the program is unchanged, but the the over all speed up of the Core_a7 WU should be at least 10% at the same clock speeds.

So the new chips will be faster at esoteric math, but F@H is pretty much all esoteric math.

For GPU projects, Ryzen offers the possibility of PCIE 4.0 if both your motherboard and GPU card support that, which should double the speed of the PCIE bottleneck F@H sometimes feels. https://en.wikipedia.org/wiki/PCI_Expre ... xpress_4.0 Again double the PCIE speed may mean a 10% speed up for the whole Core_21 WU. Existing motherboards MAY be able to make the closest slot to the CPU PCIE 4 with a new BIOS, new motherboards may support PCIE 4 with all slots.

It will be fun to see if reality matches advertising, we are certainly the target audience for these changes.

[gordonbb]

I noticed during coverage of the Ryzen preview at CES when they were comparing the new Zen to Intel one of the benchmarks was reporting in ns/d. Sounds like a workload we’re all familiar with!

Now if only they could get their video drivers working better ...

[foldy]

@JimboPalmer: Core_a7 uses AVX (non 2)

@gordonbb: Which problem do you have with AMD video drivers? Or is this the AMD drivers classic joke?

[gordonbb]

@JimboPalmer: Core_a7 uses AVX (non 2)

@gordonbb: Which problem do you have with AMD video drivers? Or is this the AMD drivers classic joke?

I _thought_ I’d read here, somewhere, that there were performance issues with the AMD OpenCL implementation which caused their GPUs to underperform.

@bruce perhaps might be able to shed some light on this or correct me if I’ve misunderstood something. I’d gladly support team red and buy more of their cards if that’s the case but currently they appear to be a poor value proposition for folding but that might just be due to the inflated mining prices I was seeing when I bought my 1070 Ti six or 8 months ago.

I use the FirePro and RadeonPro cards at work and at home on my daily driver and find AMD cards to be more stable and friendlier on the pocketbook for my use cases (Lightroom and Photoshop at home with the odd bit of Premiere).

I’m actually waiting for the new RadeonPros to trickle down to replace a W4100 with a 5300 if they ever come out. My next daily driver will likely be moving from Windows 10 on a Haswell Xeon to Windows 10 virtualized on Linux using a Threadripper

[bruce]

I'm not really the one to ask to compare GPU X with GPU Y. My biases are no better than anybody else's.

There's really no indication that A<D's OpenCL is a problem. Originally it was bundled with the AMD video drivers and now it's unbundled ... meaning you may need to install it separately, but that's not a showstopper.

"More stable" isn't in my vocabulary. A GPU is either stable or not stable, and it depends on whether you accept the manufacturer's numbers or you try to beat them with your own settings. Personally, I've bought both factory-overclocked GPUs and non-OC versions and had no troubles. The guys in the factory do a lot better testing than any of us and they don't want to get cards returned so they leave enough of a margin to be sure it's stable under all conditions.

[JimboPalmer]

I _thought_ I’d read here, somewhere, that there were performance issues with the AMD OpenCL implementation which caused their GPUs to underperform.

This is not about right and wrong, just different approaches.

Nvidia uses an infinite loop of CPU polling to keep the GPU fed, This chews up an entire CPU for each GPU but is very quick to begin a transfer between the GPU and CPU.

AMD uses interrupts which use WAY less CPU time, but has to wake up and then begin transferring data between the GPU and CPU. There is a good chance that one CPU could keep many AMD GPUs fed.

Which is 'best'? Beats me.

[ProDigit]

I _thought_ I’d read here, somewhere, that there were performance issues with the AMD OpenCL implementation which caused their GPUs to underperform.

This is not about right and wrong, just different approaches.

Nvidia uses an infinite loop of CPU polling to keep the GPU fed, This chews up an entire CPU for each GPU but is very quick to begin a transfer between the GPU and CPU.

AMD uses interrupts which use WAY less CPU time, but has to wake up and then begin transferring data between the GPU and CPU. There is a good change that one CPU could keep many AMD GPUs fed.

Which is 'best'? Beats me.

You can share CPU cores with nVidia cards as well, provided the CPU is powerful enough.
See here:
viewtopic.php?f=38&t=31320#p304780

As far as AMD vs nVidia,
I've always been an AMD fan, always have ran AMD cards, but for folding I went with nVidia.
Not so much a performance issue (yes, their top of the line RX cards, are about as fast as a GTX 1060), and their Vega cards are still a bit slower than the top of the line RTX cards, but the main issue they perform less good than nVidia, is because of their lithography.
AMD GPUs are running on 14nm, Nvidia's are running on 12nm, or 15% of power and heat savings.
That's 40Watts of power savings on 300W, which for 24/7 folding or mining, is important.

AMD already has announced their 7nm cards for release somewhere in 2019 (rumors state february, but it could easily be may before stores carry them), and will release mid-range cards first (in line with a GTX 1050, 1060), for a much higher price than equivalent GTX cards.
They should cut electric cost down by almost 50% compared to similar older cards from AMD, and by 40% compared to Nvidia's GTX line of cards.

I wouldn't say they're performing less. They're just consuming more electricity (and thus generating more heat) at the moment.

[artoar_11]

https://techreport.com/news/34388/zen-2 ... his-summer

AMD showed a brief demo of a single 64-core Epyc processor running a NAMD molecular dynamics simulation in direct comparison against a machine with two 28-core Xeon Platinum 8180 chips. Given the highly-parallel nature of the NAMD software, it's not too surprising that the 64-core Epyc walked away with a decisive victory, but the circa-20% performance advantage for the Epyc is a good bit greater than the simple difference in core counts (64 vs. 56) would imply.

[bruce]

Xeon 8180 -- `$10000 each Compared to WPYX 7551P: 32 cores (64 threads) $2100. That would pay for an upgraded MB and there will be a reduction in power/heat. If you're still running a Xeon, it would pay you to upgrade.
The WPYX is a cluster of 8 7-nm processors connected to a central 14nm I/O chip.

[ProDigit]

I got my E5-2600 v2 for $50 (LGA 2011) and my new v4 (LGA 2011-3) for $75.
Total cost of the server $500, and the second one I got for $300, saving money on a case, operating system, and graphics cards. Found 2x cheap chinese ($50) 1060s out there.
I see no reason to pay that much, besides, a 1060 is better for folding than even a $10k Xeon.

Xeons are interesting, because they have many pcie lanes. You can add in some cases up to 4x full 16x speed pcie graphics cards, depending on the motherboard.