AMD - Zen chitchat

[scaryjim]

... you are probably looking at a doubling in transistors in that leap. There will be pressures to reduce die size though if cost per transistor is improved but cost per mm^2 worsens, which is what makes me wonder if 12 core might be the sweet spot. Marketing might say sod the cost though, I think Nvidia and Intel have demonstrated that high performance is something that people are willing to buy. That would make a drop to a quad core APU seem quite steep. *shrug*, who knows...

Another thing which occurs to me is that, rather than using that transistor budget to cram more cores in, they might look at using some of it to improve clocks - much like they did with Vega, where transistor count increased 40% compared to Fury X, despite having the same number of shaders and a narrower memory interface...

[DanceswithUnix]

Another thing which occurs to me is that, rather than using that transistor budget to cram more cores in, they might look at using some of it to improve clocks - much like they did with Vega, where transistor count increased 40% compared to Fury X, despite having the same number of shaders and a narrower memory interface...

It sounds like AMD have an architectural limit on 4096 shaders, in which case they *had* to make those shaders more effective. I have also heard rumour that it is a hard to break ceiling which is why the next gen is taking time. Some of those transistors will be the double rate fp16 support, pointless for mainline gaming.

But otherwise I'm sure they are always looking at options, and I always wondered how much a bigger L3 cache would help gaming.

[CAT-THE-FIFTH]

So Zen 3 is out in H2 2020 apparently:

https://twitter.com/witeken/status/1...733448704?s=19

TSMC says to expect meaningful revenue on 7nm+ "not the first half of 2020". This suggests that AMD Zen 3 is scheduled to launch in H2'20 (and go up against Tiger Lake on 10nm++).

Zen 2 is meant to be on TSMC 7NM HPC.

[CAT-THE-FIFTH]

Oracle is now using Epyc:

https://www.anandtech.com/show/13504...c-in-the-cloud

[CAT-THE-FIFTH]

This is worth a read:

https://www.fool.com/earnings/call-t...rnings-co.aspx

Maybe I would answer the question this way. When we look at our 7-nanometer product and its positioning in 2019 across the server landscape, we feel very good about the positioning. I think it's not just 7-nanometer, 7-nanometer is important, but we've also made some significant changes to the architecture, as well as how -- sort of the system. So, I think overall we feel with the design and process capabilities at our 7-nanometer products will be quite competitive.

Well, we have a great relationship with TSMC, I think they're very supportive of our roadmap in 7-nanometer. So it's not due to any supply constraints. It's just due to the time that we believe it will take for vendors to really qualify new systems.

So, the way I look at it, Vivek is, it's really a continuum. And the continuum is, we have a number of customers that are very actively working with EPYC. I think the first 5% does include some cloud customers ramping and that's important. And then as we go beyond that, we would expect that they're more used to our architecture. Our architecture socket compatible between the first and second generation we're sampling it now. And so, I think the idea is we would like to see some acceleration in that as we bring in the 7-nanometer product but we'll certainly have to go through that process.

And we will go through this in a lot more detail on November 6. But at a high level, I think 7-nanometer gives us a better density. So for a given system we can put more cores on it. It gives us better power. So that gives us total cost of ownership. And it does give us better performance as well.

It appears AMD will have a Horizon event on November 6th talking about Zen2.

[CAT-THE-FIFTH]

The US DoE next supercomputer is using Epyc:

https://www.reddit.com/r/Amd/comment...ced_that_epyc/

Edit!!

https://www.hpcwire.com/2018/10/30/c...sc-9-contract/

In tandem with Cray’s Shasta disclosures, the U.S. Department of Energy (DOE) is announcing that NERSC, the National Energy Research Scientific Computing Center, has chosen a Cray Shasta supercomputer for its NERSC-9 system, slated for delivery in late 2020. Named “Perlmutter” (after Nobel Prize winning astrophysicist Saul Perlmutter), the system will feature AMD Epyc processors and Nvidia GPUs offering a combined peak performance of ~100 petaflops and a sustained application performance equivalent to about 3X that of the Cray Cori (NERSC-8) supercomputer. The new contract, which includes a Cray Clusterstor storage system, is worth a reported $146 million, one of the largest in Cray’s history. (We’ll be reporting additional details of the NERSC-9 system soon.)

It will be number 4 on the TOP500 list,similar to the Chinese Tianhe 2A.

[CAT-THE-FIFTH]

It looks like AMD might be winning the next US supercomputer contract too:

https://mobile.twitter.com/witeken/s...29817367379968

I have been tipped that AMD might have gotten an exascale design win (coral 2).

https://www.datacenterdynamics.com/n...shasta-system/

"The next big procurement after the NERSC-9 system is currently ongoing," Scott said. "It's CORAL 2, and they are planning on placing exascale systems in the 2021-22 timeframe at Oak Ridge National Laboratory and Lawrence Livermore National Laboratory. That procurement is ongoing so people have submitted bids and they have not yet announced winners for that system yet."

[krustylicious]

The 8+1 chiplet idea is interesting. At first I thought it was insane but after thinking about it, and seeing further tweets, its a pretty nifty idea. No increase in hopps, scaleable. As they could have reduced the die size for a ccx, which means we could see more cores per ccx, if not at least higher speeds.Plus they can go 1,2,4,6,8,10,12 ccx (in theory), By moving the if side to a signular chip and wiring the ccx's that way they have produced a master stroke

Intel are in a dark dark place

[Bagnaj97]

The 8+1 chiplet idea is interesting. At first I thought it was insane but after thinking about it, and seeing further tweets, its a pretty nifty idea. No increase in hopps, scaleable. As they could have reduced the die size for a ccx, which means we could see more cores per ccx, if not at least higher speeds.Plus they can go 1,2,4,6,8,10,12 ccx (in theory), By moving the if side to a signular chip and wiring the ccx's that way they have produced a master stroke

Intel are in a dark dark place

Could it also mean easier adoption of new tech too (DDR5, PCI-e 4, Sata 4, USB4 etc)? Core chiplets can be untouched and just the controller needs updating, although presumably it'll break socket compatibility.

[krustylicious]

Could it also mean easier adoption of new tech too (DDR5, PCI-e 4, Sata 4, USB4 etc)? Core chiplets can be untouched and just the controller needs updating, although presumably it'll break socket compatibility.

Sata 4 won't ever happen.

usb 4 is on the southbridge.

pci-e 4 will more or less need no new pins

But yes you are quite right, some of the newer tech can be replaced by a single chip; Ram and pci-e, and storage.

if you have moved the IF circuity from the zen core to a separate chip, whilst for a single zen core makes not much difference (space wise), it does if you add 2 or 4. I think we will see either a 64/96/128 core chip tomorrow. The 8+1 idea might make 4s epyc a reality (might need a rejigg on the pinouts) - aka threadripper.

Anyways all of this conjecture untill they spill the beans tomorrow at which point i think intel will cry

[CAT-THE-FIFTH]

Apparently the New Horizons event might be starting around 2pm.

[DanceswithUnix]

Apparently the New Horizons event might be starting around 2pm.

Cool. I bet that's when Openreach finally bother turning up and knacker my interwebz connection whilst attempting to fix it then

[CAT-THE-FIFTH]

Cool. I bet that's when Openreach finally bother turning up and knacker my interwebz connection whilst attempting to fix it then

I am with Sky and despite Openreach coming twice in the last month to fix things,replace loads of cabling,etc the internet is not working that well. Terrible uploads,disconnects and download speed slowly going down the drain.

[krustylicious]

its 5pm gmt 12 et

[DanceswithUnix]

I am with Sky and despite Openreach coming twice in the last month to fix things,replace loads of cabling,etc the internet is not working that well. Terrible uploads,disconnects and download speed slowly going down the drain.

Well I take it all back, I never saw "Dan" who the text message said was assigned to the job but the phone is working again and the internet connection is actually slightly faster than before.

[CAT-THE-FIFTH]

Chiplets confirmed:

https://www.anandtech.com/show/13547...9am-pt-5pm-utc

Jim from AdoredTV was correct.AVX256 confirmed.





7NM Vega is in the MI60 card!