CPCHardware：2nd gen AMD EPYC will have 64 cores, 256 Mo (!) L3, 8x DDR4-3200 and 128 PCIE-4 lines

[CatMerc]

If Zen 2 EPYC is based on some sort of interposer tech (EMIB, classic interposer, whatever), then die to die latency could effectively be on nearly the same level as intra die.

 

[scannall]

Plus I don't think you can OC on Epyc to the same extent?

You can't overclock Epyc at all. And you shouldn't be able to.

 

[raghu78]

If Zen 2 EPYC is based on some sort of interposer tech (EMIB, classic interposer, whatever), then die to die latency could effectively be on nearly the same level as intra die.

I don't see major changes to packaging tech with Zen 2. It will still be a MCM package. Zen 3 is likely to move to a chiplet based design with advanced packaging tech. SLIM by Amkor could be available by the 2020 timeframe and should be a good fit.

https://www.amkor.com/go/technology/slim

 

[CatMerc]

I don't see major changes to packaging tech with Zen 2. It will still be a MCM package. Zen 3 is likely to move to a chiplet based design with advanced packaging tech. SLIM by Amkor could be available by the 2020 timeframe and should be a good fit.

https://www.amkor.com/go/technology/slim

I'd consider it a viable possibility. Server part margins means they can afford to offer interposer packaged EPYC, even if they would cost more than EMIB.

 

[raghu78]

I'd consider it a viable possibility. Server part margins means they can afford to offer interposer packaged EPYC, even if they would cost more than EMIB.

Si Interposer based packaging is not a suitable choice as it suffers from yield and cost issues. I think AMD will not change packaging until Zen 3 arrives.

 

[Hans de Vries]

I don't see major changes to packaging tech with Zen 2. It will still be a MCM package. Zen 3 is likely to move to a chiplet based design with advanced packaging tech. SLIM by Amkor could be available by the 2020 timeframe and should be a good fit.

https://www.amkor.com/go/technology/slim

This type of technology is already used now for Xilinx and NVidia HBM2 designs (The version of SPIL/ASE)
http://www.semicontaiwan.org/en/sites/semicontaiwan.org/files/data16/docs/(3) 03_SPIL_Mike_3DIC forum_9082016.pdf
It eliminates the TSV's through the silicon interposer. (see page 8 versus page 7)


It is also possible to do decent size APU/GPU+HBM2 designs without any interposer or EMIB with newer fine line (2/2.5μm lines/spaces) organic substrates currently in High Volume Manufacturing. (ASE FOCoS)

TSMC can produce larger substrates (max 830 mm2) starting Q1-2018 (InFO-oS)

Much larger fine line organic substrates are in principle possible with Flat Panel lithography equipment. http://www.nikon.com/products/fpd/lineup/pdf/FX-68S_e.pdf

 

[DrMrLordX]

You can't overclock Epyc at all. And you shouldn't be able to.

Can you at least OC the RAM, and in so doing raise IF speeds? Or does Epyc determine IF speeds differently?

 

[scannall]

Can you at least OC the RAM, and in so doing raise IF speeds? Or does Epyc determine IF speeds differently?

No. Never, ever ever on server parts. From either Intel or AMD. Servers are a different kind of beast. They typically run with very large amounts of ECC ram. Doing jobs that MUST have reliability and stability.

 

[DrMrLordX]

No. Never, ever ever on server parts. From either Intel or AMD. Servers are a different kind of beast. They typically run with very large amounts of ECC ram. Doing jobs that MUST have reliability and stability.

So what's the maximum RAM speed supported by Epyc? Certainly you aren't stuck with DDR4-2133.

 

[raghu78]

Excellent discussion with Lisa Su at Credit Suisse conference.

http://phx.corporate-ir.net/phoenix.zhtml?c=74093&p=irol-eventDetails&EventId=5265650

Lisa states a 5% market share expectation by end of 2018 in servers which imo is conservative. 2018 is the year of the revenue ramp according to Lisa Su. 2 more unannounced semi custom designs other than Atari which will launch in H2 2018. Zen 2 is more competitive than Zen according to Lisa. All in all very exciting times ahead for PC and server industry.

 

[scannall]

So what's the maximum RAM speed supported by Epyc? Certainly you aren't stuck with DDR4-2133.

2666 on Epyc. 2400 or 2666 on Xeon's, depending on the series.

 

[amd6502]

Plus I don't think you can OC on Epyc to the same extent?

Also servers are pricy and usually don't even come with a VGA card (much less gaming card):

https://shop.softiron.com/products/

https://shop.softiron.com/product/overdrive-1000/

Also game makers are not going to code for more than 8 threads for quite some years (possibly decades) and DX12 has very limited returns on threads over 8.

 

[Yotsugi]

Lisa states a 5% market share expectation by end of 2018 in servers which imo is conservative.

AMD is in no position to say anything but the most conservative estimates possible.
Because everyone, absolutely everyone loves bashing AMD for every mistake they make, ever.

 

[krumme]

AMD is in no position to say anything but the most conservative estimates possible.
Because everyone, absolutely everyone loves bashing AMD for every mistake they make, ever.

Also servers are pricy and usually don't even come with a VGA card (much less gaming card):

https://shop.softiron.com/products/

https://shop.softiron.com/product/overdrive-1000/

Also game makers are not going to code for more than 8 threads for quite some years (possibly decades) and DX12 has very limited returns on threads over 8.

Bf1 issues 10 threads. And as Dan Baker feom oxide have shown new game model enables game beeing drivin without a main thread. High end stuff but its comming although it seems to drag on for years.

 

[Yotsugi]

Bf1 issues 10 threads. And as Dan Baker feom oxide have shown new game model enables game beeing drivin without a main thread. High end stuff but its comming although it seems to drag on for years.

BF1 chokes 2/4 and 4/4 CPUs in MP.
Most of the market is still 2/4 or 4/4 CPUs. That's the problem.
It'll take at least one more gen of moarcoarz for market to move forward.

 

[dfk7677]

BF1 MP loads 5 threads to ~100%, the rest barely load 1 more core.

 

[IRobot23]

BF1 MP loads 5 threads to ~100%, the rest barely load 1 more core.

Well with my RX 480 and R7 it loads 8 threads. To be more specific :
8T = 20-90% , 6T = 50%+, 2T = 50%-
8T = 0-5%

Sometimes load of 16T is near or just over 50%.

 

[dfk7677]

Well with my RX 480 and R7 it loads 8 threads. To be more specific :
8T = 20-90% , 6T = 50%+, 2T = 50%-
8T = 0-5%

Sometimes load of 16T is near or just over 50%.

Try using process hacker to see how it really loads threads.

 

[IRobot23]

Try using process hacker to see how it really loads threads.

Why? Its best threading that way. Performs really good.

 

[dfk7677]

Why? Its best threading that way. Performs really good.

Because threads change cores and core load cannot be accurately measured.

You can try using affinity for each thread and will notice the same.

 

[IRobot23]

Because threads change cores and core load cannot be accurately measured.

You can try using affinity for each thread and will notice the same.

Anyway, it performs really good.

As far as I am aware it loads 1T per core. While Rianbow6Siege will load 40-50% per threads all 16 between 40-50% and there is almost no improvement over 4C/8T, no matter if you use 8C/8T vs 4C/8T almost no improvement.