AMD - Zen chitchat

[kompukare]

PCGH is running a story about that Patch Tuesday was missing in February
http://www.pcgameshardware.de/Ryzen-...Patch-1222588/
(translated)
The article talks about a 3DCenter story about Windows 10 seeing an AMD chip and using the scheduling for Bulldozer etc. Wonder if a lot of this is then to do with the numbering of the cores. And that it doesn't understand that moving threads willy-nilly between CCXs is not a good idea.
This also links back this reddit thread
https://www.reddit.com/r/Amd/comment...ndows_updates/
Of course that goes straight into conspiracy theories. Guess the question is how often does Patch Tuesday get cancelled?

[scaryjim]

... the AT news item about Naples seems to have a bit of extra info compared to the others. ...

If it has 128 PCIe lanes per CPU it can't *just* be four Ryzen 7 dies on an MCM - 4 x 24 != 128. Of course, it's possible the Zeppelin die actually has 32 (or more) lanes baked in but the AM4 platform doesn't support the extra ones...

[CAT-THE-FIFTH]

Ryzen 2 early next year it seems:

https://translate.google.com/transla...-text=&act=url

I think that is where it will be at TBH.

If it has 128 PCIe lanes per CPU it can't *just* be four Ryzen 7 dies on an MCM - 4 x 24 != 128. Of course, it's possible the Zeppelin die actually has 32 (or more) lanes baked in but the AM4 platform doesn't support the extra ones...

Could be product segmentation - AMD will eventually released Opterons for DIY builds,and nobody would buy one if the consumer version had exactly the same feature set.

[kompukare]

Anyone care to do some speculation what Zen2 would bring?

Which low-hanging fruit didn't make it into the current design?

The CCX interconnect seems a bit slow for high-end desktop use, but I guess if it had been wider or faster then when scaled up to Naples' 32C/64T it might have been too much. Guess scheduler fixes might solve it, but logically the current octo and the hex cores are basically dual-CPU designs.

Better ICM is another one. Faster max DDR4 support, better latencies.

Process improvements are another obvious one, but unless they move a different process there's probably not much they can do for the vocal desktop uses who want 5GHz. According to those test The Stilt did, the current process is really good for servers and mobile especially under 3.0GHz and the only ones who could benefit from a different process with a higher frequency are overclocking desktop users. Might be worth AMD's while since a 5GHz capable chip would get the halo effect, but for the far more important server, mobile, and normal desktop (APU) users, the current process is better.

[CAT-THE-FIFTH]

Anyone care to do some speculation what Zen2 would bring?

Which low-hanging fruit didn't make it into the current design?

The CCX interconnect seems a bit slow for high-end desktop use, but I guess if it had been wider or faster then when scaled up to Naples' 32C/64T it might have been too much. Guess scheduler fixes might solve it, but logically the current octo and the hex cores are basically dual-CPU designs.

Better ICM is another one. Faster max DDR4 support, better latencies.

Process improvements are another obvious one, but unless they move a different process there's probably not much they can do for the vocal desktop uses who want 5GHz. According to those test The Stilt did, the current process is really good for servers and mobile especially under 3.0GHz and the only ones who could benefit from a different process with a higher frequency are overclocking desktop users. Might be worth AMD's while since a 5GHz capable chip would get the halo effect, but for the far more important server, mobile, and normal desktop (APU) users, the current process is better.

The biggest improvement would be by then there will be better software optimisations! The obvious change would be having improved AVX2 ability to keep Ryzen in line with Intel in that regard and if you look at the TR review,they were surprised how Intel has been quietly adding more and more AVX2 support to less obvious things like music software.

IMC we don't know - it might be quite good,but we not know since the motherboards are in a poor state - on OcUK they are mentioning the Asus ROG Crosshair VI Hero is having random bricking problems.

Well the issue is if Intel can push higher and higher stock clockspeeds,they can push less cores to do the same work as AMD,ie,smaller chips,whilst having a lead in anything which is not threaded. AMD has an IPC deficit in games currently,so Intel also have higher stock clockspeeds is not helping. Its why Apple still is on 2 cores for its phones and tablets - they went the high IPC route at highish core clockspeeds.

Remember this - the Phenom II X4 had an IPC deficit to the 45NM Intel quad cores - but AMD could ship them at higher stock clockspeeds so the difference vanished.

Why did the Phenom look crap?? AMD could not even match let alone exceed Intel clockspeeds on their Intel rivals.

Overclocking from 4.5GHZ to 5.0GHZ means nothing since most KL chips don't hit 5.0GHZ - more like 4.8GHZ/4.9GHZ,and its only Intel PR which is pushing that myth,and even then its not even a 10% difference. Its why I don't really care if Intel can overclock more - its not really adding much like in the SB days where it meant something.

[scaryjim]

... Could be product segmentation - AMD will eventually released Opterons for DIY builds,and nobody would buy one if the consumer version had exactly the same feature set.

Maybe, but on AM4+ I don't think they'll ever be able to enable more than 16 (20 in A/X300 form) PCIe lanes from the CPU, so that'd mean stepping up to a more expensive server socket to get the extra PCIe lanes - and it's hard to find motherboards from their existing server platforms (C32/G34) in the channel - e.g. scan have a number of C32/G34 Opterons for sale, but not a single motherboard.

Speculating a bit, between Ryzen 7 and full-blown Naples there's an obvious gap for a 2 die MCM in a package with ~ 2000 contacts. Since it'd be an SoC, there might be scope for cheapish prosumer grade motherboards to serve the workstation market: you'd be looking at 16C/32T, quad-channel DDR4, up to 64 PCIe lanes ... a genuine contender for s2011. In fact, given the AM4 chipsets are basically just a PCIe x4 expansion card providing a variety of storage and PCIe switching, you could just reuse the existing chipsets on the new platform.... one chipset, two platforms ...

[CAT-THE-FIFTH]

Maybe, but on AM4+ I don't think they'll ever be able to enable more than 16 (20 in A/X300 form) PCIe lanes from the CPU, so that'd mean stepping up to a more expensive server socket to get the extra PCIe lanes - and it's hard to find motherboards from their existing server platforms (C32/G34) in the channel - e.g. scan have a number of C32/G34 Opterons for sale, but not a single motherboard.

Speculating a bit, between Ryzen 7 and full-blown Naples there's an obvious gap for a 2 die MCM in a package with ~ 2000 contacts. Since it'd be an SoC, there might be scope for cheapish prosumer grade motherboards to serve the workstation market: you'd be looking at 16C/32T, quad-channel DDR4, up to 64 PCIe lanes ... a genuine contender for s2011. In fact, given the AM4 chipsets are basically just a PCIe x4 expansion card providing a variety of storage and PCIe switching, you could just reuse the existing chipsets on the new platform.... one chipset, two platforms ...

Well I just had another thought - would AM4 support dual sockets?? So instead of trying to pushing Naples down to DIY builders,AMD could start to release dual socket AM4 motherboards where you could use a pair of Ryzen 7 CPUs,and have double the PCI-E lanes.

[scaryjim]

... would AM4 support dual sockets?? ... you could use a pair of Ryzen 7 CPUs,and have double the PCI-E lanes.

The issue there would be inter-CPU communication. Naples dedicates half of its PCIe resources - 64 lanes per CPU - to inter-CPU comunication. Since the AM4 socket only appears to plumb out a maximum of 24 lanes (some of which are meant to be reserved for storage), I think you'd end up with a painfully slow link between the CPUs and no more free PCIe lanes than a single AM4 CPU.

The Naples PCIe setup implies that each die has at least 32 PCIe lanes, but we know those aren't all available on socket AM4, which has a best case officially supported scenario of 22 (that's the x16 "GPU" cluster, the x4 that connects to the chipset, and x2 from the storage controller if you only hang 2 sata ports off the SoC). But even if we take the AM4 setup as being the base, a 2 die MCM would support up to 44, and a minimum of 36 (that's the two x16 "GPU" clusters and one x4 chipset cluster (since you don't need to connect one CPU to two chipsets ).

So in fact a 2 die MCM on its own socket makes inherent sense for the workstation and 1P mid-range server market. It not only stacks up very nicely against Intel's equivalent platform - it would offer up to 16C/32T and at least 36 PCIe lanes - it would also sit nicely between AM4 Ryzen 7 and Opterons (up to 8C/16T and max 22 PCIe lanes) for the prosumer & small/home server market and Naples (up to 64C/128T and 128 PCIe lanes) for the high-end server market.

See, now I'm going to be hugely disappointed if AMD don't launch a product that only exists in my imagination

[CAT-THE-FIFTH]

I dunno - it would mean AMD having three different sockets though??

[scaryjim]

I dunno - it would mean AMD having three different sockets though??

They currently have 4 - FM2, AM4, C32 and G34.

FM2 will be going very shortly (hopefully, if they get either BR APUs or new Zen based (Raven Ridge?) APUs into the channel quickly.

I don't think 3 sockets would be too many tbh - and they'd be clearly partitioned. They wouldn't have to release opteron on AM4 - leave that as the consumer/prosumer platform, then have a workstation/entry level server Opteron platform and a high end server Opteron platform above it.

[CAT-THE-FIFTH]

They currently have 4 - FM2, AM4, C32 and G34.

FM2 will be going very shortly (hopefully, if they get either BR APUs or new Zen based (Raven Ridge?) APUs into the channel quickly.

I don't think 3 sockets would be too many tbh - and they'd be clearly partitioned. They wouldn't have to release opteron on AM4 - leave that as the consumer/prosumer platform, then have a workstation/entry level server Opteron platform and a high end server Opteron platform above it.

If they do that,Intel will still have an edge in core counts upto 18 cores. If they could unify the equivalent of C32 and G34 into one platform,they will be able to offer more cores than Intel.

[scaryjim]

If they do that,Intel will still have an edge in core counts upto 18 cores. ...

Well, Intel's s2011 Xeons will give you up to 24C/48T in one socket at the minute, but those processors still top out at 40 PCIe lanes, so it's easy enough to see how they'd share a platform top-to-bottom. I'm not convinced that it'd be so easy for processors with ~ 40 PCIe lanes to share a platform with a processor with 128 lanes. Plus I doubt there's much of a value proposition in a 4000+ contact socket in the workstation environment.

The Naples platform feels like a special purpose halo server platform: it's just got an excessive number of everything. I do wonder if the 4 die 32 core variant is meant to be very low volume, and if the bulk of the server offering will be in a smaller platform.

Or perhaps they've found a way to engineer the platform at a sensible cost and they can trickle it right down to a set of smaller chips with lower number of dies. I suppose we won't actually know until they get released.

One thing I strongly suspect though, is that if they go for a single server platform we won't see any lower-end or workstation-class motherboards hit the channel....

[Agrippa]

I really, REALLY hope AMD is planning to release a platform to challenge Intel's s2011. While Ryzen appears to provide all the performance I want and need for a new build, I find the X370 platform sorely lacking. If I wanted a pure gaming PC for the boy's room it would be fine, but since I need at least one PCIe x8 for an HBA and would also like at least one x4 M.2 slot there's just no way I can build an AMD-based machine as things stand.

[scaryjim]

... since I need at least one PCIe x8 for an HBA and would also like at least one x4 M.2 slot there's just no way I can build an AMD-based machine as things stand.

Wouldn't an X370 be just fine for this? Use the first PCIe slot for graphics, stick the HBA in the second PCIe slot, you get a pair of x8 3.0 slots - you're going to lose negligible performance running a GPU at x8...

[Agrippa]

Sort of, I guess, but only today, not so much tomorrow I think. I build my computers to last as long as possible and the current one took shape around this time in 2012. The only changes I've made since then have been to swap graphics cards and hard drives.

I'm hoping the next one will last at least as long and in two, three or more years I'm absolutely expecting that x8 3.0 slot to hold me back significantly. I could be wrong about that of course, but I'm not willing to take the chance and would rather spend more money now to make sure I don't have to build again in a couple of years. I've built entirely too many computers as it is and I really don't enjoy the job any longer.

32 available lanes is basically what I need at a minimum and 40, which is what I have now, would be perfect.

[kompukare]

Wouldn't an X370 be just fine for this? Use the first PCIe slot for graphics, stick the HBA in the second PCIe slot, you get a pair of x8 3.0 slots - you're going to lose negligible performance running a GPU at x8...

Indeed, for the doubters I love posting the TPU PCIe scaling results:

That's the 4K results for the Nvidia GTX 1080, but W1zzard also previously this test for AMD Fury X:


So x8 PCIe 3.0 is somewhere between +0% and -1% difference. In other words, two x8 PCIe 3.0 slots is plenty for a graphic card and a SAS controller.