Expectations of Ryzen Threadripper 6000 Series

[tuckerkao]

I've checked the CPU models of Threadripper 5000 Series, the base clock is still under 4.0 GHz -
https://www.cyberpowerpc.com/system/...eadripper-Core

Heard some rumor that the Threadripper 6000 Series will be clocked faster to impress the PC gamers. I believe these CPU will be able to finish the PRP tests of exponents such as M199,003,267 within 9 days.

I'm waiting for the AMD RDNA 3 GPUs too.

[tServo]

Originally Posted by tuckerkao

I've checked the CPU models of Threadripper 5000 Series, the base clock is still under 4.0 GHz -
https://www.cyberpowerpc.com/system/...eadripper-Core


I'm waiting for the AMD RDNA 3 GPUs too.

(yawn......) Not to rain on your parade but ......
Nvidia's rtx 3090 & 3080 were "released" mid September and AMD's rx 6900xt was "released" at the end of October and here, in March. they are still rare or downright unavailable. I could find none in stock at 3 big US retailers.
( please don't respond by saying so-and-so retailer had them available last week for 17 seconds before they sold out )
I would guess that RDNA should be available mid 2024 or maybe 2025.

[M344587487]

Originally Posted by tuckerkao

I've checked the CPU models of Threadripper 5000 Series, the base clock is still under 4.0 GHz -
https://www.cyberpowerpc.com/system/...eadripper-Core

Heard some rumor that the Threadripper 6000 Series will be clocked faster to impress the PC gamers. I believe these CPU will be able to finish the PRP tests of exponents such as M199,003,267 within 9 days.

I'm waiting for the AMD RDNA 3 GPUs too.

We're limited by memory bandwidth, cache helps. Zen4 cache per chiplet will be at least 16MiB, I wouldn't be surprised to see that increase but they may choose to increase L1/L2 cache instead if anything. There's a rumour that the Zen4 generations will have up to 12 chiplets of 8 cores, naturally more cache in this case but if >8 chiplets will exist for Epyc it won't necessarily exist for Threadripper. Zen4 is also when DDR5 is introduced so all bets are off.

Originally Posted by tServo

(yawn......) Not to rain on your parade but ......
Nvidia's rtx 3090 & 3080 were "released" mid September and AMD's rx 6900xt was "released" at the end of October and here, in March. they are still rare or downright unavailable. I could find none in stock at 3 big US retailers.
( please don't respond by saying so-and-so retailer had them available last week for 17 seconds before they sold out )
I would guess that RDNA should be available mid 2024 or maybe 2025.

The shortage will last for at least another year. The best we can hope for AMD hardware is that console demand dies down so they can focus on more profitable dies but even if that comes to pass it'll only make a dent at best. Nvidia is spinning up older nodes to meet mining demand, it's unknown how much mining demand there actually is. Even if mining demand dies down and/or supply ramps up you can bet there'll be significant inertia in the price as every retailer has to re-learn that the price can go more than one direction.

[tServo]

Originally Posted by M344587487

We're limited by memory bandwidth, cache helps. Zen4 cache per chiplet will be at least 16MiB, I wouldn't be surprised to see that increase but they may choose to increase L1/L2 cache instead if anything. There's a rumour that the Zen4 generations will have up to 12 chiplets of 8 cores, naturally more cache in this case but if >8 chiplets will exist for Epyc it won't necessarily exist for Threadripper. Zen4 is also when DDR5 is introduced so all bets are off.


The shortage will last for at least another year. The best we can hope for AMD hardware is that console demand dies down so they can focus on more profitable dies but even if that comes to pass it'll only make a dent at best. Nvidia is spinning up older nodes to meet mining demand, it's unknown how much mining demand there actually is. Even if mining demand dies down and/or supply ramps up you can bet there'll be significant inertia in the price as every retailer has to re-learn that the price can go more than one direction.

You made a valid point about the consoles. WRT mining, my impression for the last year or 2 is that serious miners use ASICs rather than GPUs. Perhaps the GPU makers are blaming the miners to hide other problems like bad yield problems, perhaps?

[pepi37]

Originally Posted by tServo

You made a valid point about the consoles. WRT mining, my impression for the last year or 2 is that serious miners use ASICs rather than GPUs. Perhaps the GPU makers are blaming the miners to hide other problems like bad yield problems, perhaps?

I have 12 GTX/RTX cards. I do mining. ASIC is better option, but it is limited to algo, and cannot be used on anything else. GPU can do many thing, and can also be sold years after with decent price. Current ASIC in two years from will be useless ( or near that point)
Current GPU , that is different story

[M344587487]

Originally Posted by tServo

You made a valid point about the consoles. WRT mining, my impression for the last year or 2 is that serious miners use ASICs rather than GPUs. Perhaps the GPU makers are blaming the miners to hide other problems like bad yield problems, perhaps?

Serious or casual, people mine whatever is profitable. ASICs have been made to mine certain algorithms applicable to certain coins. Ethereum has an algorithm designed to be somewhat ASIC-resistant and is ideally suited to GPUs thanks to scaling with memory bandwidth (there do exist Ethereum ASICs but they're not orders of magnitude better than GPU's and the ROI doesn't/didn't make sense unless bulk buying with other considerations like space limitations and hardware maintenance at scale). Ethereum is a popular coin with a lot of mining potential so there is high GPU demand to mine it.



Miners are a common scapegoat, but not generally by GPU makers. If anything they're very quiet about selling directly to miners because it's bad press. It's possible that there's yield issues with Samsung's 8nm node but there's also just high demand, whatever the reason they aren't producing enough. There would be a shortage even without mining, Nvidia's MSRP is back to sensible territory now but there's been a few years of silly pricing (thanks partly to AMD not being competitive) so a demand for an affordable upgrade path has been building for years. TSMC's 7nm node is so constrained that AMD was never going to be a big part of the equation.

[tuckerkao]

I hope that the DDR5 Memory modules will be available for the compatible motherboards and CPUs by mid year 2022. If the total memory capacities reaches 512 gigabytes per machine, the P-1 factoring should be under 2 hours for most exponents under 200M.

[tuckerkao]

Originally Posted by tServo

I would guess that RDNA should be available mid 2024 or maybe 2025.

I searched "AMD RDNA 3 Release Date" on Google, it said -> The likely target for RDNA 3 launch, therefore, seems to be H1 2022, but this is just speculation for now.

Zen 4 CPUs also seem to be scheduled for H1 2022.
https://www.techradar.com/news/amd-p...rd-as-big-navi

[drkirkby]

Originally Posted by tuckerkao

I hope that the DDR5 Memory modules will be available for the compatible motherboards and CPUs by mid year 2022. If the total memory capacities reaches 512 gigabytes per machine, the P-1 factoring should be under 2 hours for most exponents under 200M.

Is that really true? Money permitting I could stuff 3 TB in my PC, but it has 384 GB RAM now. Can I actually; use that amount of RAM to speed things up? No, I'm not going to upgrade to beyond 384 GB. To go to 768 GB would be easy, but any more than that and I would have to use bigger DIMMs.

[LaurV]

Originally Posted by drkirkby

Is that really true?

No. Or, say, partially. Having more RAM helps to run the Stage 2 of the P-1 algorithm more efficient, but the approach we have here is that we prefer to keep the same time but increase the chance of finding a factor. The best tradeoff effort vs. factors found is when the two stages take about the same amount of time. If you can run stage 1 very fast, you go for a higher B1. If you can run stage 2 very fast, you go for a higher B2. At the end, having more memory will allow you to go for a higher B2, so you will have a higher chance to find a (larger) factor. But what this clown says is mostly vegetable market talk. He comes here with his loco predictions and expects us to do the work for him. Whatever math he talks, you can safely ignore.

[drkirkby]

Originally Posted by LaurV

No. Or, say, partially. Having more RAM helps to run the Stage 2 of the P-1 algorithm more efficient, but the approach we have here is that we prefer to keep the same time but increase the chance of finding a factor. The best tradeoff effort vs. factors found is when the two stages take about the same amount of time. If you can run stage 1 very fast, you go for a higher B1. If you can run stage 2 very fast, you go for a higher B2. At the end, having more memory will allow you to go for a higher B2, so you will have a higher chance to find a (larger) factor.

At the moment I have

Daytime P-1/ECM stage 2 memory in GB (100.000000):
Nighttime P-1/ECM stage 2 memory in GB (100.000000):
Max emergency memory in GB/worker (70.000000):

However, it so happens that I'm currently running two exponents (one on each CPU), with each taking around 48 hours. Exponents are starting around 24 hours apart, so when one exponent starts, the other will be 50% complete. This would imply that no two exponents would be doing P-1 at the same time. Would it be worth my why increasing the memory for P-1? Unless I'm running EM simulations, which is not that often, I can probably give mprime more than 300 GB RAM. Would there be an advantage in doing so? From what you say, this would have a higher chance of finding a factor, so I assume waste less time doing PRP tests on composite numbers. I guess this could go wrong, if one P-1 test finds a factor, it could start another exponent whilst the other CPU is also doing a P-1 test.

Anyway, I can very safely increase the above to 170 GB, from 100 GB. I never gave it much thought before.

Dave