buying a new computer

[Leo]

I would stay away from the 2060's because they only have 6GB of VRAM. The 2060 Supers all have 8gb of VRAM as far as I know. I would also stay away from reference design cards. Best to check reviews of the thermal performance of any card you select as this can be an issue in dual card configurations.

The cheapest 2080 Ti I've seen is $1050, not sure of the quality. You can get a 2060 Super of good quality for $400 and 2070 Supers are $500.

I'm going to build a 100M+nps machine for < $3000. So yes, you can get them for less than $5K US.

Regards,

Zenmastur

I don't think 6GB of VRAM is an issue when using Lc0 to play; it can be a factor when training nets (although batch size can be adjusted).

In the US (don't know if non-US shipping is an option) a 2060 can be had for $299 on sale, and I recently got a 2080ti for $949 (dual fan, quiet) from EVGA B-stock:
https://www.evga.com/products/productlist.aspx?type=8
Have to look from time to time (search for 'instant' savings); when on sale they sell out fast.

I primarily use the 2080ti to train nets (where the 11GB helps).
Check different models for clock speed differences.

For playing and particularly for analyzing for longer periods, more CPU RAM seems to be a more important factor.

Nice purchase.

[Dann Corbit]

If 4 cards don't work together, even with 8 channel memory and PCie4, you could always run 2 pairs of 2 cards with two instances of LC0 to double your throughput.

I do admit it is a gamble to go after any software implementation + hardware combination that has never been tried.

People reported here already thermal problems with one single RTX card and LC0
so they had to add additional fans or to optimize the air flow, so I see little
chance for an air cooled 4 way gpu PC, maybe with an custom build water cooling
solution, but then maybe 2 PCs with a 2 way gpu setup would make more sense.

--
Srdja

If I was going to run four cards, it would have to be an industrial machine that is made for that. They have boxes that can hold ten or more cards larger than any of the super cards. Of course, you need at least a kilowatt just for four cards, and if you want 7 man tb files, you need to power a lot of disks. I guess the safe size would be 2 kw or more, so that also means industrial grade. I found a nice configuration for a 7702p with 512gb of fast RAM at a decent price, but it only had a Windows configuration option and I will only accept a Linux variant.
I also want to see what the threadripper timeline nlooks like. I guess when they announce the 3rd generation, they will say something about what core counts will be available and when.
Considering that the first one is supposed to be only 24 cores, I will probably buy an industrial machine instead. But we will see. For now I can limp along on my current hardware. And with the weather turning colder, I can switch my big machines back on.

[Nordlandia]

Is the machine going to be dedicated to engine games or mix of analysis and engine game. If the first is the case then two identical machines for making ponder beneficial is wonderful.

[Zenmastur]

If 4 cards don't work together, even with 8 channel memory and PCie4, you could always run 2 pairs of 2 cards with two instances of LC0 to double your throughput.

I do admit it is a gamble to go after any software implementation + hardware combination that has never been tried.

People reported here already thermal problems with one single RTX card and LC0
so they had to add additional fans or to optimize the air flow, so I see little
chance for an air cooled 4 way gpu PC, maybe with an custom build water cooling
solution, but then maybe 2 PCs with a 2 way gpu setup would make more sense.

--
Srdja

If I was going to run four cards, it would have to be an industrial machine that is made for that. They have boxes that can hold ten or more cards larger than any of the super cards. Of course, you need at least a kilowatt just for four cards, and if you want 7 man tb files, you need to power a lot of disks. I guess the safe size would be 2 kw or more, so that also means industrial grade. I found a nice configuration for a 7702p with 512gb of fast RAM at a decent price, but it only had a Windows configuration option and I will only accept a Linux variant.

That seems odd that they would force you to put an OS on it. I would just request it with out an OS and install it myself. I'm curious as to which vendor is doing this.

I also want to see what the threadripper timeline nlooks like. I guess when they announce the 3rd generation, they will say something about what core counts will be available and when.
Considering that the first one is supposed to be only 24 cores, I will probably buy an industrial machine instead.

Not sure you are correct about there only being a 24-core part. I've seen nothing from AMD to indicate they will only release a single SKU to start with. My suspicion is that in November we will see at least two SKU's with 4-memory channels released (24-core and 32-core) and 2 SKU's announced (but not yet released) for the 8-channel memory parts. If they delay the 8-memory channel parts for very long then it seems reasonable that it will be released on the Zen 3 architecture possibly using 7nm+ process.

But we will see. For now I can limp along on my current hardware. And with the weather turning colder, I can switch my big machines back on.

I'm limping along as well. I thought about buying old servers for this purpose but I don't like the noise, high electricity bills, and massive heat output.

One other note, AMD has stated that they plan on a part with 15 chiplets on it. The rumors floating around is that the 4 extra chiplets will be for HBM2 (4GB ??) on the die. But I can see many other uses for these besides on die memory. e.g. more cores, on die tensor cores, on die graphics, on die inter-processor communication to support machines with large CPU counts, etc ,etc. This opens up a huge number of possibilities. Even the possibility to have custom chips on die.

Regards,

Zenmastur

[dannyb]

I'm limping along as well. I thought about buying old servers for this purpose but I don't like the noise, high electricity bills, and massive heat output.

I thought about buying some 2x Xeon v3 but the prices are still high: ~400$ for an E5-2695 v3 (14 cores) is too much. The QS and ES versions cost much less but I prefer not to take the risk.

[Nordlandia]

Graham Banks said that ponder works fairly well on dual xeon motherboard.

[Zenmastur]

Gigabyte has release an overclocking guide for the Ryzen 9 3950X along with Cinebench R15 benchmarks for the overclocked and non-overclocked parts. It appears the the 3950X will out perform the Core I9 10980XE by a small margin stock with overclocked parts being roughly equal (slightly favoring AMD depending on the particular silicon). I'm sure there will be other benchmarks where the chips are closer to parity and a few where Intel parts are superior.

The R9 3950x MSRP is 30% cheaper than the i9 10980XE and has a TDP advantage of about 57%. Neither AMD's nor Intel's TDP's are an accurate representation of "real" power consumption but the disparity between the two figure is enough to assure the the power draw difference between the two parts will be significant. This doesn't bode well for INTEL. Both chips will be roughly equal in available I/O bandwidth. Intel's part will be some what gimped by the "long in the tooth" X299 chip-set. The decision to keep X299 I'm sure was motivated by the need to get the chips to market quickly AND it gives X299 owners an upgrade path which will help the chips market acceptance.

In the short term X299 was probably a good decision. In the long term they should either design a new chip-set or move to a modified LGA 3647. LGA 3647 is a server chip set but could be modified to better serve the HEDT/enthusiast platforms. There are some MB's based on this chip set that are aimed at the HEDT market (as opposed to pure server platforms) but they cost well over $1,200 which puts them outside most of the HEDT/enthusiasts budget range. Modification would be needed to allow them to compete more directly with Ryzen Threadripper. Currently Intel has nothing that can compete directly with third gen. Threadripper. The Xeon W-3175X isn't competitive. E.G. GIGABYTE C621 AORUS XTREME is $1900 and the W-3175X is $3000. By the time you add the rest of the system components you are in the $7500-$8000 range, which is outside most peoples budget. 

So, It looks like Intel still has some ground to make up, and this will take time. No quick fixes seem feasable.

Edit: the Intel socket is LGA3647, the chip set used is C621.

Regards,

Zenmastur

[Zenmastur]

A link to the WCCFTECH Article:
https://wccftech.com/tsmc-achieves-hvm-7nm-plus-euv/

We know that RYZEN 3 started tapping out in the last month and now TSMC is ready to start high volume manufacturing. This can't be good for INTEL. The 7nm+ is a completely different process than the 7nm process which most likely means a new fab is going into production soon. SO, maybe, shortages will stop being a thing. Density is up 20% per wafer, % yields are about equal to the 7nm node, power consumption is down, and speed is up. Sounds like the best of all worlds. All was done in record time! Intel better get their 10nm node straightened out in a hurry!

One of the major differences between Zen 2 and Zen 3 is that ALL 8-cores of a chiplet will now share a unified 32MB L3 cache. This will cut L3 cache access latency for all program because less data will have to cross the infinity fabric. On a 2-chiplet design 75% of all L3 cache accesses hit the Infinity fabric. On ZEN3 this will drop to about 50%. On a single chiplet CPU no L3 cache access will use the Infinity fabric (only main memory etc. will use IF).

Regards,

Zenmastur

[dannyb]

I'm considering buy a RTX card for Leela and other NN engines. Do you think it's worth buying a 2070 or a 2060 would already be enough? Maybe I could buy later a second 2060. Do these engines work well with two cards?

[dannyb]

or probably I should just wait as I've read that nvidia should announce the new RTX cards on january so maybe the 2070 will be cheaper