#1 - Page 16

pinhodecarlos · 2020-06-09, 17:16

Quote:

Originally Posted by moebius

F*** off current Navi, even old Vega10 chip , is 4 times faster than an 8-core Ryzen 3700X, and also at least 50 Euros cheaper than a AMD RX 5700XT....

It is not about how fast but how efficient it is. How much is your CPU consuming against the GPU on a daily basis?

moebius · 2020-06-10, 14:54

Quote:

Originally Posted by pinhodecarlos

How much is your CPU consuming against the GPU on a daily basis?

Sure, an RX 5700 Xt consumes e.g. 225W, which is 50 watts less than my RX Vega64 with a maximum consumption of 275 watts. My objection is also more related to GIMPS (Single Precision Flops and Memory Read / Write are better with the Vega), although the 5000 card actually performs better in many games .I'm more annoyed about, that AMD no longer sells Vega 10 and 20 chips, but sells only the gamer stuff.

It is not against CPU'S (especially not against my current one) because they have a different purpose.

pinhodecarlos · 2020-06-10, 19:00

Quote:

Originally Posted by moebius

Sure, an RX 5700 Xt consumes e.g. 225W, which is 50 watts less than my RX Vega64 with a maximum consumption of 275 watts. My objection is also more related to GIMPS (Single Precision Flops and Memory Read / Write are better with the Vega), although the 5000 card actually performs better in many games .I'm more annoyed about, that AMD no longer sells Vega 10 and 20 chips, but sells only the gamer stuff.

It is not against CPU'S (especially not against my current one) because they have a different purpose.

You were comparing CPU speed with GPU speed, which they are not normalised to the same conditions. I’ve only addressed on that perspective. Take care.

Xyzzy · 2020-06-13, 21:15

We now have two GPUs in this box. They use about 100W each.

Unfortunately, they are nowhere near as fast as Radeon VII cards.

However, we were surprised that just one of these cards (PRP/gpuowl) is 113% faster than the CPU (PRP/mprime) on all twelve cores!

GalebG2 · 2020-06-25, 10:20

I was wondering, with 64GB of RAM, won't I be too much behind the curve these days? Regarding the graphics card, I'm looking at several options, would the NVidia Titan RDX be a good choice? It's quite expensive, and I really want a good return on investment in terms of performance.

VBCurtis · 2020-06-26, 06:04

Behind the curve to do what?
You can solve a rather large NFS matrix with 64GB; a problem that doesn't fit would also need multiple months to solve on any sub-$5000 CPU setup.

GalebG2 · 2020-06-26, 13:01

Okay, thanks for the answer, that's more or less what I imagined, but with the talks about higher RAM counts, I was sort of lost.

VBCurtis · 2020-06-26, 15:20

I'm not sure what your experience level is with NFS matrices, but an msieve matrix roughly 26M in size fits on a 16GB machine, and something in the 45-to-50M range fits on 32GB. I haven't run a large enough matrix yet to need more than 64GB, but it is likely to scale similarly so projects to ~80M in size. Filtering is likely to need more RAM than the matrix somewhere around there (say, as you move to larger than 33-bit large primes or jobs tougher than GNFS-210).

If you're using CADO for postprocessing, memory needs are higher for both filtering and LA; but a 50-60M matrix will fit in 64GB still. I found 64GB not enough for CADO filtering for GNFS-207, but I was using large prime bounds too big (34/35 bit) and had over 3e9 raw relations.

M344587487 · 2020-06-26, 16:42

Quote:

Originally Posted by moebius

Sure, an RX 5700 Xt consumes e.g. 225W, which is 50 watts less than my RX Vega64 with a maximum consumption of 275 watts. My objection is also more related to GIMPS (Single Precision Flops and Memory Read / Write are better with the Vega), although the 5000 card actually performs better in many games .I'm more annoyed about, that AMD no longer sells Vega 10 and 20 chips, but sells only the gamer stuff.

It is not against CPU'S (especially not against my current one) because they have a different purpose.

Trying to make a single design fit gamers and compute simultaneously is partly why AMD has struggled with GPU performance particularly on the gaming side. We've been lucky that GCN favoured compute so much and that HBM was made available in a consumer product to kick things up a notch with our hobbyist compute. Navi finally closes the gap with nvidia gaming cards which I don't grudge them for doing, but it does so with a process advantage (aka nvidia is going to kick AMD's butt when it finally does transition to 7nm) (from the look of it in compute too but at nvidia's trademark we-have-no-competition insane pricing). RDNA2 is rumoured to be a pretty massive jump in performance so I hope AMD at least stays in the running for gaming, would be a shame to go back to bugger all competition at the high end.

The next compute chip from AMD is CDNA. We won't see a consumer version but might get a pro version if we're lucky. A larger chunk of the die is probably dedicated to tensor-flow/machine-learning/other-buzzwords than with Vega20 so it may not be as FP64 performant as it could be. Do these newer categories have any bearing on prime hunting?

https://www.anandtech.com/show/15593...r-data-centers

https://www.servethehome.com/amd-cdn...ture-5nm-epyc/

https://videocardz.com/newz/amds-mar...00-accelerator

moebius · 2020-06-27, 09:27

Quote:

Originally Posted by M344587487

Trying to make a...

Does anyone have a gpuowl benchmark for this graphics card. If you find a new Mersenne prime number, you may even be able to afford this astronomical price.
https://www.mindfactory.de/product_i...5_1335677.html

kriesel · 2020-06-27, 17:32

Quote:

Originally Posted by moebius

Does anyone have a gpuowl benchmark for this graphics card.

Doubtful at that price.

https://www.amd.com/en/products/prof.../instinct-mi25 768GFLOPS FP64
https://www.techpowerup.com/gpu-specs/radeon-vii.c3358 3.36TFLOPS FP64
So MI25 ~1/4 the speed of a RadeonVII?!
Demonstrating again what a good deal the RadeonVII is, and the difference between HPC&professional pricing and consumer pricing. NVIDIA does the same with Tesla and Quadro lines, versus GTX & RTX. High Professional/server use pricing, quality drivers, ECC, higher speed interfaces, vs. gaming-oriented drivers and who cares if a bit here and there is bad, and make do with PCIe rates. https://docs.nvidia.com/gameworks/co...te_cluster.htm

2020-06-13, 21:15	#169
Xyzzy Aug 2002 7×1,237 Posts	We now have two GPUs in this box. They use about 100W each. Unfortunately, they are nowhere near as fast as Radeon VII cards. However, we were surprised that just one of these cards (PRP/gpuowl) is 113% faster than the CPU (PRP/mprime) on all twelve cores! Attached Thumbnails

2020-06-25, 10:20	#170
GalebG2 Jun 2020 7 Posts	I was wondering, with 64GB of RAM, won't I be too much behind the curve these days? Regarding the graphics card, I'm looking at several options, would the NVidia Titan RDX be a good choice? It's quite expensive, and I really want a good return on investment in terms of performance.

2020-06-26, 06:04	#171
VBCurtis "Curtis" Feb 2005 Riverside, CA 2·2,927 Posts	Behind the curve to do what? You can solve a rather large NFS matrix with 64GB; a problem that doesn't fit would also need multiple months to solve on any sub-$5000 CPU setup.

2020-06-26, 13:01	#172
GalebG2 Jun 2020 7 Posts	Okay, thanks for the answer, that's more or less what I imagined, but with the talks about higher RAM counts, I was sort of lost.

2020-06-26, 15:20	#173
VBCurtis "Curtis" Feb 2005 Riverside, CA 2×2,927 Posts	I'm not sure what your experience level is with NFS matrices, but an msieve matrix roughly 26M in size fits on a 16GB machine, and something in the 45-to-50M range fits on 32GB. I haven't run a large enough matrix yet to need more than 64GB, but it is likely to scale similarly so projects to ~80M in size. Filtering is likely to need more RAM than the matrix somewhere around there (say, as you move to larger than 33-bit large primes or jobs tougher than GNFS-210). If you're using CADO for postprocessing, memory needs are higher for both filtering and LA; but a 50-60M matrix will fit in 64GB still. I found 64GB not enough for CADO filtering for GNFS-207, but I was using large prime bounds too big (34/35 bit) and had over 3e9 raw relations.