Zen 2 samples are with AMD, pumping away with 8C/16T at 4.5GHz .. TweakTown

[scannall]

You said: "AMD has promised to live within the constraints of their current sockets."

Why does this mean that there can't be an improved CPU and socket with more channels memory and still have the CPU backward compatible with the previous. Threadripper 2990WX has the resources to use 8 channels but since the boards can only use 4 then that is what is enabled.

Basically, is saying "we will maintain socket compatibility" mean no new sockets?

Their next socket should be in 2020, with DDR5 and Pcie 4/5 support.

 

[yuri69]

@Percode licensing: Oracle DB, MSSQL, Exchange, ... what next Novell? It's 2018, majority lives in a cloud on a OSS stack.

@IMC and stuff: AMD is tight on cash and time... They've been massively using 3rd party IPs (with all the drawbacks). Just an example: https://www.synopsys.com/dw/doc.php/ss/amd_ss.pdf

@Zen2: Don't let the hype train leave its home station...

 

[dnavas]

Basically, is saying "we will maintain socket compatibility" mean no new sockets?

It is exactly my suspicion that this might be their chosen path to maintain backwards compatibility and get greater performance. Performance would be CPU+socket upgrade. Caution: I know nothing about their possible plans....

 

[TheGiant]

It is exactly my suspicion that this might be their chosen path to maintain backwards compatibility and get greater performance. Performance would be CPU+socket upgrade. Caution: I know nothing about their possible plans....

I have bad expecience with so called socket compatibility. With upgrade like every 2-3 years yes the CPU might be able to fit into the socket, but unless your have top model board, bios won't be compatible with new CPU or the board is not able to sustain higher currents etc. It happens especially when the CPU node changes, like now from 14/12nm to 7nm.

So in practice, I don't think an average B350 board will run the 7nm zen and if yes, the OC potential will be lesser than expected. So the result, to get max of your new top technology CPU, you must change the board. I hope I am wrong.

 

[DrMrLordX]

They should be.

Again, why? AMD is selling dGPUs for any serious HPC application. I already articulated that. AMD has a "don't know/don't care" attitude towards 256/512-bit SIMD performance anywhere else.

They aren't Intel, and they aren't going to obsess over being the top performer everywhere. They already don't. They have their target markets, and everyone else gets to use their hardware to the best of its capabilities elsewhere, which as it turns out, is doing pretty well right now.

Maybe I don't agree with their apparent approach, but I recognize it for what it is.

You said: "AMD has promised to live within the constraints of their current sockets."

Why does this mean that there can't be an improved CPU and socket with more channels memory and still have the CPU backward compatible with the previous. Threadripper 2990WX has the resources to use 8 channels but since the boards can only use 4 then that is what is enabled.

Basically, is saying "we will maintain socket compatibility" mean no new sockets?

AMD is going to have to consider a platform upgrade for Zen2 Threadripper. Existing TR boards don't have enough memory channels for anything bigger/badder than the 2990WX. How they go about that transition will be interesting, especially since it could imperil sales of EPYC systems under some circumstances.

I have bad expecience with so called socket compatibility. With upgrade like every 2-3 years yes the CPU might be able to fit into the socket, but unless your have top model board, bios won't be compatible with new CPU or the board is not able to sustain higher currents etc. It happens especially when the CPU node changes, like now from 14/12nm to 7nm.

So in practice, I don't think an average B350 board will run the 7nm zen and if yes, the OC potential will be lesser than expected. So the result, to get max of your new top technology CPU, you must change the board. I hope I am wrong.

Somewhat agreed. Even "top" boards like the X370 Taichi get weird when you start updating the UEFI too often. Older UEFI revisions (sadly pre-Spectre) are the best for Summit Ridge while newer Pinnacle Ridge UEFIs have some oddball bugs that are not fun to deal with. Anyone looking for the best Pinnacle Ridge experience really need ASRocks X470 board, while someone looking for the best Summit Ridge experience should want the X370 Taichi using a UEFI from late 2017. But again, Spectre . . .

Anyway I expect support for Matisse on X370 to be rather poor across the industry.

 

[itsmydamnation]

Why did this post remind me of Bulldozer?

I don't know it's nothing like bulldozer, far more like power9

 

[dnavas]

Again, why?

Because their current set of advantages won't last.
Because increased cores are of limited value.
Because without their current set of advantages, their short-comings become more visible.
Did I not state that?

They aren't Intel, and they aren't going to obsess over being the top performer everywhere.

That's a mistake. You either run for the gold, or you don't compete. Second place is reserved for those that lose the race for gold. No one wins the race for silver.

Maybe I don't agree with their apparent approach, but I recognize it for what it is.

I think/hope you are mistaking short-term positioning for long-term goals. You prioritize what you can do quickly, where you can make the most sales, but you don't cede ground long-term. It might be ok if TR3 ships w/o 256bit vectors -- not for me, but for some. But if they never get wider vector support? It's a dead platform.

AMD is going to have to consider a platform upgrade for Zen2 Threadripper. Existing TR boards don't have enough memory channels for anything bigger/badder than the 2990WX.

Arguably they don't do the 2990WX justice. Bifurcate your TR market. X series is backwards compatible. WX series becomes new-socket only. So, 3950X is 16 - 32 cores with two lanes, 3990X is 32 - 64 cores with four. The way you preserve your Epyc market is to add even more memory channels there. They're likely needed if they're really getting 64 cores, and no one expects socket compatibility there anyway. Or you add more lanes. Support for 2x and 4x systems. Maybe the WX series doesn't support two slots per lane. Lots of ways to keep these lines differentiated.

 

[itsmydamnation]

Increased AVX unit width is of limited value, whats more important is total load/store in and out of the Core. There is an argument for 256bit SIMD units and datapaths. I would rather see 4 symmetric 128bit SIMD/FMA units with double the amount of 128 bit load and store ports. That will benefit far more workloads then 256bit units and datapaths and provide the same peak throughput, but it would also be significantly harder to implement.

 

[JDG1980]

@Percode licensing: Oracle DB, MSSQL, Exchange, ... what next Novell? It's 2018, majority lives in a cloud on a OSS stack.

MSSQL is still used extensively in the enterprise (most of my current day job involves working with it). That said, MS doesn't have any particular allegiance to Intel, and if per-core licensing becomes a problem, I can imagine them going to some other form of metered licensing instead.

@IMC and stuff: AMD is tight on cash and time... They've been massively using 3rd party IPs (with all the drawbacks). Just an example: https://www.synopsys.com/dw/doc.php/ss/amd_ss.pdf

That was true for the original Zen release, but for Zen2, AMD has had more time to prepare plus a considerably better revenue outlook. We can't be sure, but it's certainly not out of the realm of possibility that one of the Zen2 improvements was a better, in-house-designed memory controller.

@Zen2: Don't let the hype train leave its home station...

What constitutes hype? I think AMD's success with Zen so far entitles us to some degree of optimism. Are you expecting Zen2 to be just a die-shrink of Zen+? Even that would likely mean a few hundred extra MHz at the top end, due to the transition from a Global Foundries mobile process to a TSMC high-performance process. I think a reasonable expectation for desktop Zen2 is turbo clock speeds of ~4.5GHz, maybe a bit better if we're lucky, and something like 10% IPC improvement in average applications over Zen+.

 

[maddie]

Because their current set of advantages won't last.
Because increased cores are of limited value.
Because without their current set of advantages, their short-comings become more visible.
Did I not state that?

That's a mistake. You either run for the gold, or you don't compete. Second place is reserved for those that lose the race for gold. No one wins the race for silver.

I think/hope you are mistaking short-term positioning for long-term goals. You prioritize what you can do quickly, where you can make the most sales, but you don't cede ground long-term. It might be ok if TR3 ships w/o 256bit vectors -- not for me, but for some. But if they never get wider vector support? It's a dead platform.

Arguably they don't do the 2990WX justice. Bifurcate your TR market. X series is backwards compatible. WX series becomes new-socket only. So, 3950X is 16 - 32 cores with two lanes, 3990X is 32 - 64 cores with four. The way you preserve your Epyc market is to add even more memory channels there. They're likely needed if they're really getting 64 cores, and no one expects socket compatibility there anyway. Or you add more lanes. Support for 2x and 4x systems. Maybe the WX series doesn't support two slots per lane. Lots of ways to keep these lines differentiated.

I have to strongly disagree with this.

There is no race. There are a multitude of races. Each one different and trying to effectively compete across the board is impossible for a smaller competitor. You will lose in all, as your resources will everywhere be spread too thin.

Your strategy will be the death of AMD in its present position

 

[beginner99]

Most software solutions have moved away from per-core to per-cpu (physical).

Except oracle. They charge you for the whole machine even if you run it in a VM with only 2-cores. Then they realized well you have a full vmware cluster and you could run it everywhere so even if you run a small databases on 2-cores you now need to pay for the full cluster. Then you split your cluster, then a tool gets released that combines clusters Then oracle tells you you have o pay for both clusters.

I vote for a new rule on forums: profanity allowed if it targets oracle. Company i work for is actively moving to PostgreSQL. Sadly we have 3rd party applications that only work on oracle as most of the competitor ones. So it's hard to completely get rid of them.

 

[yuri69]

MSSQL is still used extensively in the enterprise (most of my current day job involves working with it). That said, MS doesn't have any particular allegiance to Intel, and if per-core licensing becomes a problem, I can imagine them going to some other form of metered licensing instead.

Yea, that's the legacy world. You can regularly meet AS/400 based large enterprises these days. But still, that doesn't mean it's a "2018 mainstream".

That was true for the original Zen release, but for Zen2, AMD has had more time to prepare plus a considerably better revenue outlook. We can't be sure, but it's certainly not out of the realm of possibility that one of the Zen2 improvements was a better, in-house-designed memory controller.

The Zen2 design phase was not affected by revenue generated by anything Zen1 based. Mind you, Zen1 project started someday in 2012. Given that, the Zen2 design started ~2 years after Zen, that indicates 2014.

What constitutes hype? I think AMD's success with Zen so far entitles us to some degree of optimism. Are you expecting Zen2 to be just a die-shrink of Zen+? Even that would likely mean a few hundred extra MHz at the top end, due to the transition from a Global Foundries mobile process to a TSMC high-performance process. I think a reasonable expectation for desktop Zen2 is turbo clock speeds of ~4.5GHz, maybe a bit better if we're lucky, and something like 10% IPC improvement in average applications over Zen+.

Personally, I expect a major core count increase => push for densish servers, classic VMs, etc. 7nm should bring efficiency benefits and lower risks due to clock requirements.

The hype train is fueled by articles/posters like "Intel in trouble, Zen2 on 7nm", "Zen2 will rule 2019", etc. Still, there is a big unknown about the TSMC's 7nm built hiperf CPUs. Seeing Intel getting aggressive (4->6->8c, 9900K running at 4.5-5GHz turbo) - just don't hype the thing.

 

[DrMrLordX]

I have to strongly disagree with this.

There is no race. There are a multitude of races. Each one different and trying to effectively compete across the board is impossible for a smaller competitor. You will lose in all, as your resources will everywhere be spread too thin.

Your strategy will be the death of AMD in its present position

Agreed. Again, I would like to see better SIMD performance out of Zen2 than what we get from Zen+. I just don't expect it, nor do I see that as being the best path forward for AMD right now.

Seeing Intel getting aggressive (4->6->8c, 9900K running at 4.5-5GHz turbo) - just don't hype the thing.

Hype aside, if you see the 8700k and 9900k as "Intel being aggressive", then I think you are looking at things with blue-colored glasses. That smell? It's desperation.

Yeah Zen2 might provide something lame like 5% IPC improvement and no clockspeed increase versus Zen+. It could happen. Most of AMD's gains could be in a few datacenter-specific workloads that won't affect desktop users. It could shine most at clockspeeds that will be a major yawnfest for us as well (arguably, this is already the case for Zen+). I just don't expect that.

In fact, expecting 4.5-4.6 GHz operation (versus 4.3 GHz for Zen+) and 12-13% higher IPC across the board for Zen2 is rather conservative. 5 GHz? Nahh. Probably not gonna happen without exotic cooling.

 

[thecoolnessrune]

Yea, that's the legacy world. You can regularly meet AS/400 based large enterprises these days. But still, that doesn't mean it's a "2018 mainstream".

You mentioned these points in the previous post as well. It's worth pointing out that AS/400 hasn't been used since the turn of the Millenium, and its been a unified IBM Power Systems product stack for a decade. While there are definitely AS/400 systems out there, there are also still a large amount of systems running in enterprise on IBM i, AIX, and Linux using IBM Power Systems. You could definitely consider it mainstream, but just because a system architecture has been out for a long time, doesn't means it's legacy.

That all stated, it's less of a strange statement than stating as if Relational Databases are somehow legacy. They still exist in the cloud, they still need licensing, and they're not going anywhere for a long time. Hell, most Linux installs include MariaDB to leverage (which is just the open source version of MySQL it used to use). Oracle alone, the arse that it is, owns 4x as much of the market as the nearest non-relational DB.

Transactional DBs aren't going anywhere anytime soon, they'll just continue growing in the cloud, where for now you're still licensing them by core count and memory pool.

 

[dnavas]

Agreed. Again, I would like to see better SIMD performance out of Zen2 than what we get from Zen+. I just don't expect it, nor do I see that as being the best path forward for AMD right now.

I was reacting to "AMD is not losing any sleep." This is a competition -- you should definitely be losing sleep over choices made and tradeoffs taken. I, also, expect tradeoffs. Those tradeoffs will piss someone off. TR was an engineer-driven product and may well not figure prominently in company resuscitation. Memory latency and peak clocks for desktop and lower idle for laptops would have to be more important. Such is life. They are definitely going all-out for the server market as evidenced by moving to 7nm and a new arch on their high end first. It's a risky move -- mistakes will count quite heavily there.

 

[yuri69]

That all stated, it's less of a strange statement than stating as if Relational Databases are somehow legacy. They still exist in the cloud, they still need licensing, and they're not going anywhere for a long time. Hell, most Linux installs include MariaDB to leverage (which is just the open source version of MySQL it used to use). Oracle alone, the arse that it is, owns 4x as much of the market as the nearest non-relational DB.

Transactional DBs aren't going anywhere anytime soon, they'll just continue growing in the cloud, where for now you're still licensing them by core count and memory pool.

This discussion has revolved aroud the percore licensing issue, not the RDBs...

The legacy theme is bound to licensing, not the fact they are RDBs. Nowadays you are really not bound to commercial licenses, unless you have to cope with legal issues. The landscape and mindest has changed from "don't even think of doing business w/o Oracle" to "let's just push it on a pgsql (or mysql/maria whatever)". Is there a percore issue?

 

[thecoolnessrune]

This discussion has revolved aroud the percore licensing issue, not the RDBs...

The legacy theme is bound to licensing, not the fact they are RDBs. Nowadays you are really not bound to commercial licenses, unless you have to cope with legal issues. The landscape and mindest has changed from "don't even think of doing business w/o Oracle" to "let's just push it on a pgsql (or mysql/maria whatever)". Is there a percore issue?

I'd be curious if there's any data to back up this statement if its available. Most commercial projects very much have a list of supported Databases. Now that is not to say the landscape is not changing. A large example I can think of is VMware vCenter. For a long time, vCenter was only supported on big-iron MSSQL and Oracle SQL Databases. VMware got tired of losing their cut of the pie to extra costs associated with the Database, so they have their own version of PostgreSQL called vPostgres that they now use to do the Database themselves. They also used to run on Windows and SUSE Linux, but they got tired of paying licensing to SUSE so they built their own Linux distro called PhotonOS and use that for their Appliance now.

That all above stated, the model works because of the application. VMware has worked hard over the years to minimize the impact of a corrupt or destroyed Database, to make it where restoring the Appliance is the quickest and safest way of getting vCenter back vs. troubleshooting the Database. This used to not be the case. You absolutely needed that transactional Database in-tact, and were required to purchase licensing to big-iron databases that assured that integrity.

The mindset changed because we're building better apps, not because of some sort of stigma. In the old days, the relational database was one of the most important pieces of a large company. It got all the resources (dedicated boxes long after almost everything else was virtualized), consumed a large portion of licensing costs, and had full time staff. Because all the applications said "just rely on the Database to keep me intact". Now we build better apps that recover better, that rebuild better, that restart better. That's when we started going "you know, we don't need these heavy iron DBs."

And that's a big difference. The reason you see mariadb and pgsql in a lot of places is because they can come standalone in a very small footprint. The only big competitor to those open source standalone DBs is Microsoft SQL Express, which has a very large footprint, and is resigned to running on Windows only.

When you have those DBs start running heavy iron workload, you find that there isn't that big of a cost-parity. Getting a supported MariaDB Cluster for instance is going to cost you 30K for 3 nodes. That can buy you 24 cores of Microsoft SQL Standard licensing at Advertised Price. While MariaDB can easily license better when developing huge DB servers, under 20 cores (which is the vast majority of SQL clusters), the Microsoft Licensing is actually better. And Microsoft also has their Server + CAL Licensing model, but you better have a Microsoft Licensing Expert available for navigating that quagmire.

The summary of all that is that yes there are absolutely times where you are bound to commercial support licenses. Those aren't going away anytime soon. And when you move your commercially supported Database to the cloud, you are still paying for a bundle of cores and RAM in a large number of cases.

Could that model be legacy someday and go purely to transaction billing? Sure. Is it even remotely close to be considered legacy today? Absolutely not.

 

[jpiniero]

BTW AMD will apparently discuss Zen 2 on November 6th, per the earnings conference call.

 

[Abwx]

Future 4C/8T mobile APU on sight :

ZM370SC4T4MFG_38/22_Y

https://www.userbenchmark.com/UserRun/12123813

https://www.computerbase.de/2018-11/amd-ryzen-7-3700u-picasso/

 

[Glo.]

Not related. Mobile, Picasso APUs are Zen+, and on 12 nm process.

 

[cortexa99]

Interesting, some old slide imply this R7-3700u could be zen2:
https://cdn.videocardz.com/1/2018/03/AMD-Ryzen-2018-2020-Roadmap.jpg
https://videocardz.com/75217/amd-ry...astle-peak-matisse-picasso-vermeer-and-renoir

In the slide both Pinnacle Ridge(12nm) & Raven Ridge(14nm) were above 'Optimization - Process maturity efficiency enhancement' tag, but Matisse & Picasso were 'Inflection - new process tech new cpu core'.

That said this R7-3700u is
1. belongs to RR(14nm) rebranded
2. Picasso(7nm/12nm)
3. some codename that hasn't been unveiled(XXX Ridge, 12nm)

 

[Topweasel]

Picasso is a Raven Ridge Arch upgrade much like Pinnacle Ridge was a Summit Ridge upgrade.