AMD - Zen chitchat

[CAT-THE-FIFTH]

I had an interesting thought regarding all this raytracing stuff. Navi was rumoured to be a MCM type setup with multiple identical chips,trying to appear as one GPU to software. Apparently this is hard to do and AMD said Navi was a normal GPU. However,what if Navi or a future AMD GPU line are normal monolithic GPUs but using IF to integrate a co-processor for things like ray tracing?? Hence two smaller chips on a MCM.

It probably won't happen,but I thought it would be an alternate way forwards,instead of large chips??

[scaryjim]

… using IF to integrate a co-processor for things like ray tracing?? Hence two smaller chips on a MCM.

It probably won't happen,but I thought it would be an alternate way forwards,instead of large chips??

I'm pretty sure that at some point MCMs will become the way forward for GPUs, but AFAIK at the minute the extra latency of taking trips off-die is just too high to make it effective. OTOH having on-die IF linking modular logic blocks might make it easier for AMD to provide different levels of ray-tracing capability on different GPUs, if they can come up with a repeatable block that can be implemented multiple times...

[Biscuit]

I had an interesting thought regarding all this raytracing stuff. Navi was rumoured to be a MCM type setup with multiple identical chips,trying to appear as one GPU to software. Apparently this is hard to do and AMD said Navi was a normal GPU. However,what if Navi or a future AMD GPU line are normal monolithic GPUs but using IF to integrate a co-processor for things like ray tracing?? Hence two smaller chips on a MCM.

It probably won't happen,but I thought it would be an alternate way forwards,instead of large chips??

I think, just as Jim said, this is inevitable.

I think a lot of the progress with MCM will happen on CPU before it makes its way to GPU. Threadripper is really just the start in a long path of advancements.
Suppose we could a parallel compute card that really isn't for gaming at all, but uses GPU-based-tech with MCM for better yeilds.

For data centre/AI/science that could be pretty useful.

[CAT-THE-FIFTH]

Have people seen this:

https://sconedocs.github.io/microcode/
https://labs.vmware.com/flings/vmwar...river/comments

That is the microcode update for Intel CPUs under Linux for the Foreshadow vulnerability which points to this:

https://downloadmirror.intel.com/280...e-20180807.tgz

The latest microcode updates are listed here so its a valid link:

https://downloadcenter.intel.com/dow...-Data-File?v=t

Now if you download,look at the license,which says this:

https://paste.ubuntu.com/p/z2F3Cj6R8Q/

3. LICENSE RESTRICTIONS. All right, title and interest in and to the Software
and associated documentation are and will remain the exclusive property of
Intel and its licensors or suppliers. Unless expressly permitted under the
Agreement, You will not, and will not allow any third party to (i) use, copy,
distribute, sell or offer to sell the Software or associated documentation;
(ii) modify, adapt, enhance, disassemble, decompile, reverse engineer, change
or create derivative works from the Software except and only to the extent as
specifically required by mandatory applicable laws or any applicable third
party license terms accompanying the Software; (iii) use or make the Software
available for the use or benefit of third parties; or (iv) use the Software on
Your products other than those that include the Intel hardware product(s),
platform(s), or software identified in the Software; or (v) publish or provide
any Software benchmark or comparison test results. You acknowledge that an
essential basis of the bargain in this Agreement is that Intel grants You no
licenses or other rights including, but not limited to, patent, copyright,
trade secret, trademark, trade name, service mark or other intellectual
property licenses or rights with respect to the Software and associated
documentation, by implication, estoppel or otherwise, except for the licenses
expressly granted above. You acknowledge there are significant uses of the
Software in its original, unmodified and uncombined form. You may not remove
any copyright notices from the Software.

Wait,wut?? People can't benchmark the effect of the microcode update??

[kompukare]

Wait,wut?? People can't benchmark the effect of the microcode update??

That seems an incredible abuse of power and display of corporate arrogance.
Actually, I think this deserves its own thread in as many forums as possible.

EDIT:
Doesn't this mean that logically Intel CPUs can no longer be benchmarked or reviewed at all from now on?

Because otherwise the details of any tests would read something like this:

All benches were run with the latest patches, drivers, and updates except for the Intel CPUs where the performance impact of the latest security patches is top secret and we may not disclose their performance impact.

[kalniel]

The comparison clause is usually included in beta software, I'm surprised if it's present in any release version. That said, I also haven't seen any such terms and conditions when applying updates to computers.

[CAT-THE-FIFTH]

That seems an incredible abuse of power and display of corporate arrogance.
Actually, I think this deserves its own thread in as many forums as possible.

EDIT:
Doesn't this mean that logically Intel CPUs can no longer be benchmarked or reviewed at all from now on?

Because otherwise the details of any tests would read something like this:

Read the other parts of the clause which says the US government is not allowed to benchmark,them of the fact to read the clause you need to download the software which means you are bound by the clause.

Well other parts of the clause have annoyed Debian:

https://www.theregister.co.uk/2018/0...patch_licence/

Also,I won't purport to be a Linux expert so do you want me to start a new thread here or do you want to??

[CAT-THE-FIFTH]

Also a rumour AMD might launch APUs on 7NM by the end of the year:

https://videocardz.com/77660/exprevi...d-of-this-year

[Biscuit]

Also a rumour AMD might launch APUs on 7NM by the end of the year:

https://videocardz.com/77660/exprevi...d-of-this-year

Would make sense as the 2xxx series APU's were released with Zen+ before the full Zen+ release IIRC?

[CAT-THE-FIFTH]

Would make sense as the 2xxx series APU's were released with Zen+ before the full Zen+ release IIRC?

Also I have some concerns about 7NM too(think about AMD and the fact it is dual sourcing) - I do think it is quite possible AMD might be able to beat Intel to a proper rollout of parts on 7NM/10NM but clockspeeds might be the issue at least initially.

On mobile clockspeeds are a lesser issue,and also Intel has technically "beaten" AMD in the core count there as they have 6C and soon to be 8C APUs. Moving to 7NM allows AMD to add more cores,whilst keeping chip size and power consumption in check. Also the APUs can act as a 7NM CPU test.

[Biscuit]

Also I have some concerns about 7NM too(think about AMD and the fact it is dual sourcing) - I do think it is quite possible AMD might be able to beat Intel to a proper rollout of parts on 7NM/10NM but clockspeeds might be the issue at least initially.

On mobile clockspeeds are a lesser issue,and also Intel has technically "beaten" AMD in the core count there as they have 6C and soon to be 8C APUs. Moving to 7NM allows AMD to add more cores,whilst keeping chip size and power consumption in check. Also the APUs can act as a 7NM CPU test.

I wonder if the 2018 late release is when we will find out how many cores are in the new CCX then...

[MLyons]

I wonder if the 2018 late release is when we will find out how many cores are in the new CCX then...

Unlikely.

[Biscuit]

Unlikely.

But otherwise, that would mean they're running out 2 x CCX version on 7nm, which would go against their approach so far of having basically one universal scalable design?

[MLyons]

But otherwise, that would mean they're running out 2 x CCX version on 7nm, which would go against their approach so far of having basically one universal scalable design?

The 2018 release IIRC is 12nm. Zen 2 is 2019 some time. That;s when we'll find out. Unless these new rumoured APUs are in fact 7nm. If they are then we'll find out.

[scaryjim]

But otherwise, that would mean they're running out 2 x CCX version on 7nm, which would go against their approach so far of having basically one universal scalable design?

They could run out a 2x CCX APU and a 4x CCX main die, which could then be scaled up to a 64 core 4 die MCM.

I'd personally find it really odd if the 4 core CCX only lasted one (and a half?) generation before they redefined the CCX again - particularly when the whole point of the CCX is that it's modular. Perhaps they'll decide that 4 cores is actually enough for an APU. After all, the current APUs aren't priced or specced to compete with i7s, or even i5s really. We might see a 1 CCX APU with more shaders, and maybe a 3 CCX CPU for up to 12 cores. 4C/8T APUs are still going to compare favourably to the i3s anyway (which are currently, and likely to remain, 4C/4T).

And then again maybe they'll just concentrate on using the transistor budget to improve clockspeeds, like they did with Vega 64. If they can pull a similar use of transistor count to bump the clock speeds of Zen 2 they won't need to increase core counts, because they won't be compromising on single-threaded performance anymore....

[CAT-THE-FIFTH]

AMD PR has commented,Epyc and Vega are only going to be released this year on 7NM. Apparently AMD is releasing a new super efficient H series CPU.

They appear to have less cache,but the memory support is interesting.

"AMD Reference Platform, “Raven Ridge 2018” AMD Ryzen™ 7 2800H, 2x8GB DDR4-3200, Samsung VLV2560 SSD
Windows 10 x64 16299.64, Graphics Driver: 23.20.768.0, 1920×1080"

https://www8.hp.com/h20195/v2/GetPDF...A7-3217EEP.pdf

They could run out a 2x CCX APU and a 4x CCX main die, which could then be scaled up to a 64 core 4 die MCM.

I'd personally find it really odd if the 4 core CCX only lasted one (and a half?) generation before they redefined the CCX again - particularly when the whole point of the CCX is that it's modular. Perhaps they'll decide that 4 cores is actually enough for an APU. After all, the current APUs aren't priced or specced to compete with i7s, or even i5s really. We might see a 1 CCX APU with more shaders, and maybe a 3 CCX CPU for up to 12 cores. 4C/8T APUs are still going to compare favourably to the i3s anyway (which are currently, and likely to remain, 4C/4T).

And then again maybe they'll just concentrate on using the transistor budget to improve clockspeeds, like they did with Vega 64. If they can pull a similar use of transistor count to bump the clock speeds of Zen 2 they won't need to increase core counts, because they won't be compromising on single-threaded performance anymore....

Having said that - look at the AT article on the 2990WX and look how much of the power budget Infinity Fabric is taking up. Now I suspected why they had it in a 1:2 ratio with memory frequency was down to power,but its surprising how much Infinity Fabric takes up.

I also looked at the TH article on Ryzen+ and cache latencies are as good as Intel now - so it is basically latency due to the inter-CCX communication and system memory.

So it could be quite possible a lot of improvements are to the Inifinity Fabric instead,instead of pure clockspeed,especially if they end up adding more cores.

Now this might have an impact on games,since if they can get latencies down,that would probably help a lot with games.

Plus this is GlobalFoundries we are talking about,there claims of 5GHZ clockspeeds need to be taken with a grain of salt,until proof is shown!!