Dual EPYC

[MikeB]

Does anyone have experience with this enough to recommend a workstation manufacturer? I'm looking for a large manufacturer at tier 1 level - such as Dell - but I'm really struggling to find a company which sells dual EPYC machines.

The only feedback I have is that it's hard to get 2 EPYCs in a case and Dell don't go with AMD at all.

Bizarre.

It appears none of the Tier 1 companies sell EPYC serves other than in rack format and as of today, I could find anyone selling a 128 core dual EPYC configuration. A dual EPYC with 128 cores will cost close to $20,000 or more from any Tier 1.

In all honesty, I would just go with a 3990x with a high end gpu. No to need to blow $13000 on something that would add 50 Elo max. Of course if you're filthy rich , that would not matter.

[Dann Corbit]

It's also possible he needs it for more than chess.
And the I/O on a full 8-channel RAM Epyc is mind blowing.
With two 7742's (or those even higher speed exotic versions) you can run a database at speeds never imagined before.\
I just bought a RAID card for my machine with 4 gumstick SSDs that are PCIE 4.0 capable.
Here comes the data stream of my dreams.
(I have 4 ultra high speed PCIE 4.0 2 TB gum sticks to put in it, and I can smell the data sizzling already)

[Werewolf]

It appears none of the Tier 1 companies sell EPYC serves other than in rack format and as of today, I could find anyone selling a 128 core dual EPYC configuration. A dual EPYC with 128 cores will cost close to $20,000 or more from any Tier 1.

In all honesty, I would just go with a 3990x with a high end gpu. No to need to blow $13000 on something that would add 50 Elo max. Of course if you're filthy rich , that would not matter.

It's odd about the tier 1 companies, isn't it?

On the CPU choice, if this was about a single chess engine I would indeed make do with Threadripper. But I use Aquarium IDeA (yes some of us still use it!) for chess analysis where the extra CPU will literally double performance.

In the old days I clustered together half a dozen PCs for Aquarium and used that for fairly cheap analysis. I gave up 3 years ago vowing never again - Windows Nazi-enforced updates made managing even a small cluster a nightmare for 24/7 365 analysis - now I want it all in a single quiet, reliable box.

[MikeB]

...

I gave up 3 years ago vowing never again - Windows Nazi-enforced updates made managing even a small cluster a nightmare for 24/7 365 analysis - now I want it all in a single quiet, reliable box.

I ended up thermal throttling my threadripper at 74C - heat is not real bad ( i know some people do not get near 74C - congrats to you!) - still goes steady 3833 Mhz while using all 64 cores - thats setting runs at 265 watts I could crank it to 4.0 Ghz - but difference is only 2 or 3% nps ( now benches with current SF and large pages at around 99M nps)- I like the less heat and the less fan noise - my setup is in the loft looking over the family room and just barely hums.

[smatovic]

Just curious, what happened out of this?

--
Srdja

[Werewolf]

I’m considering.

I’m very interested, may wait for Milan, not sure

[smatovic]

I’m considering.

I’m very interested, may wait for Milan, not sure

Just thinking loud...TSMC will have 5nm at risk production in 2nd half 2020,
AMD may switch maybe 2021 or 2022 (Genoa*), TSMC aims 3nm for 2022, AMD may
adapt in 2023...AMD plans some new stuff for the ExaFLOP computer El Capitan in
2023...so we may see some new sockets with DDR5 already in 2021/2022, and I am
not sure what will descend from El Capitan into the market after 2023...

*
https://images.anandtech.com/doci/15594 ... oadmap.png


I ponder myself on a dual EPYC machine, need mostly RAM and bandwidth and some
cores for a graph-database, but I am in no hurry, I can wait till some fancy new
stuff arises >=2023, maybe HBM on CPU interposer or alike, that with 3nm
production would be a burner for me.

--
Srdja

[Milos]

Just thinking loud...TSMC will have 5nm at risk production in 2nd half 2020,
AMD may switch maybe 2021 or 2022 (Genoa*), TSMC aims 3nm for 2022, AMD may
adapt in 2023...AMD plans some new stuff for the ExaFLOP computer El Capitan in
2023...so we may see some new sockets with DDR5 already in 2021/2022, and I am
not sure what will descend from El Capitan into the market after 2023...

*
https://images.anandtech.com/doci/15594 ... oadmap.png


I ponder myself on a dual EPYC machine, need mostly RAM and bandwidth and some
cores for a graph-database, but I am in no hurry, I can wait till some fancy new
stuff arises >=2023, maybe HBM on CPU interposer or alike, that with 3nm
production would be a burner for me.

You are really a lost cause optimist. TSMC originally planned 5nm risk production in April 2019. And it's really questionable if they are even gonna have it this year. And their high yield numbers of 15mm2 dies are a joke.
Basically the only one already in risk production (but with also quite bad yield) is Samsung.
And 3nm as things look today is a science fiction.

[smatovic]

Just thinking loud...TSMC will have 5nm at risk production in 2nd half 2020,
AMD may switch maybe 2021 or 2022 (Genoa*), TSMC aims 3nm for 2022, AMD may
adapt in 2023...AMD plans some new stuff for the ExaFLOP computer El Capitan in
2023...so we may see some new sockets with DDR5 already in 2021/2022, and I am
not sure what will descend from El Capitan into the market after 2023...

*
https://images.anandtech.com/doci/15594 ... oadmap.png


I ponder myself on a dual EPYC machine, need mostly RAM and bandwidth and some
cores for a graph-database, but I am in no hurry, I can wait till some fancy new
stuff arises >=2023, maybe HBM on CPU interposer or alike, that with 3nm
production would be a burner for me.

You are really a lost cause optimist. TSMC originally planned 5nm risk production in April 2019. And it's really questionable if they are even gonna have it this year. And their high yield numbers of 15mm2 dies are a joke.
Basically the only one already in risk production (but with also quite bad yield) is Samsung.
And 3nm as things look today is a science fiction.

Hehe, our insider speaks, it seems I am more a pessimist...

N5 Status
TSMC’s 5-nanometer node entered risk production early last year. C.C. Wei confirmed that N5 is now in volume production with good yield. We have recently covered TSMC 5-nanometer node disclose from a number of conferences. The node will deliver around 1.8x improvement in density along with 15% higher speed at iso-power or, alternatively, 30% lower power at the same speed. N5, like N7, is expected to remain a long-lasting node. “We expect a very fast and smooth ramp of N5 in the second half of this year driven by both mobile and HPC applications,” C.C. Wei added. TSMC estimates that N5 will contribute about 10% of wafer revenue in 2020. TSMC is observing a higher number of tape-outs compared to N7 at the same period of time during their ramp.

https://fuse.wikichip.org/news/3453/tsm ... illimeter/

But for sure your crystal ball has more to offer...

--
Srdja

[Milos]

Hehe, our insider speaks, it seems I am more a pessimist...

N5 Status
TSMC’s 5-nanometer node entered risk production early last year. C.C. Wei confirmed that N5 is now in volume production with good yield. We have recently covered TSMC 5-nanometer node disclose from a number of conferences. The node will deliver around 1.8x improvement in density along with 15% higher speed at iso-power or, alternatively, 30% lower power at the same speed. N5, like N7, is expected to remain a long-lasting node. “We expect a very fast and smooth ramp of N5 in the second half of this year driven by both mobile and HPC applications,” C.C. Wei added. TSMC estimates that N5 will contribute about 10% of wafer revenue in 2020. TSMC is observing a higher number of tape-outs compared to N7 at the same period of time during their ramp.

https://fuse.wikichip.org/news/3453/tsm ... illimeter/

But for sure your crystal ball has more to offer...

Well done, you quote TSMC PR guy advertising their product. And you can laugh as much as you like, but for your info, as someone who is actually doing ASIC design in 5nm as we speak, I know pretty well the situation.
And you, I bet you didn't even read a single paper related to 5nm technology, let alone ever seen a design manual or rule book.