Intel Knights Mill, Stratix 10 + future of cpu making, and Scorpio

[MyName13]

10 minutes ago, The Benjamins said:

We knew it was Jaguar cores, if they wanted zen it would be delayed by at least a year.

Making a console with a such powerful CPU (assuming they still want more than 4 cores) would be pointless because most gamers wouldn't have that processing power.

[The Benjamins]

Just now, MyName13 said:

Making a console with a such powerful CPU (assuming they still want more than 4 cores) would be pointless because most gamers wouldn't have that processing power.

well 4 zen cores is faster then 8 jaguar cores, but if Microsoft wanted to do that they would have to delay the release due to AMD not ready to do custom zen designs. and the current 8 Jaguar cores is the one part that will hold back the Xbox One X the most.

[MyName13]

2 minutes ago, The Benjamins said:

well 4 zen cores is faster then 8 jaguar cores, but if Microsoft wanted to do that they would have to delay the release due to AMD not ready to do custom zen designs. and the current 8 Jaguar cores is the one part that will hold back the Xbox One X the most.

Yes, Zen still can't be put into consoles even with 4 cores, especially because PS4 has more users.

[The Benjamins]

1 minute ago, MyName13 said:

Yes, Zen still can't be put into consoles even with 4 cores, especially because PS4 has more users.

what does PS4 have to do with this?

What do you mean by "Zen still can't be put into consoles even with 4 cores"

[MyName13]

5 minutes ago, The Benjamins said:

what does PS4 have to do with this?

What do you mean by "Zen still can't be put into consoles even with 4 cores"

If Xbox gets at least 4 Zen cores it will be pointless because it would be far more powerful than PS4 and probably many PCs (because of console optimisation), developers wouldn't make 1 edition with better physics or A.I. for example, and a dumber version for PS4.All that power would be unused and by the time sony releases ps5 Xbox one x would be obsolete because ps5 would probably have more cores or a better architecture.

[The Benjamins]

5 minutes ago, MyName13 said:

If Xbox gets at least 4 Zen cores it will be pointless because it would be far more powerful than PS4 and probably many PCs (because of console optimisation), developers wouldn't make 1 edition with better physics or A.I. for example, and a dumber version for PS4.All that power would be unused and by the time sony releases ps5 Xbox one x would be obsolete because ps5 would probably have more cores or a better architecture.

well the XOX SOC has a hardware DX12 engine in silicon and if they had the zen cores they wouldn't need as custom of a CPU implementation. but i get what you are saying.

 

once the next Xbox and PS come out with AMD zen/vega(navi) then we should see games push the CPU more.

[Jito463]

3 hours ago, Sniperfox47 said:

Except Intel was using interposers first, long before AMD... on Xeon Phi...

Only in the server market, and that only up to at most a year before the launch of Ryzen, so it's not like they were doing this for several years.  Yes, there was the Pentium DC with its MCM, but that was a horrible implementation.  It's been evident for a while that Intel was pushing the limits of single die packages, they just refused to change while they could still milk people on the current technology without as much cost to them (actually changing the fundamental structure of the chip inherently requires R&D expenses).

 

*EDIT*

2 hours ago, leadeater said:

Clovertown and Core 2 Quad too 

 

That's 2006 btw.

Err, that's what I was thinking of.  Which, as I said, had a horrible implementation.  The inter-core communication was extremely slow.

[Sauron]

2 hours ago, The Benjamins said:

if they wanted zen it would be delayed by at least a year.

Yes, and it may also have been worth paying attention to.

[The Benjamins]

Just now, Sauron said:

Yes, and it may also have been worth paying attention to.

I think the XOX is a nice improvement it is able to scale 1080p XO games to 4k so they hit their goal. But I would love to see the potential of a high end AMD APU in action (zen based)

[Sniperfox47]

1 hour ago, Jito463 said:

Only in the server market, and that only up to at most a year before the launch of Ryzen, so it's not like they were doing this for several years.  Yes, there was the Pentium DC with its MCM, but that was a horrible implementation.  It's been evident for a while that Intel was pushing the limits of single die packages, they just refused to change while they could still milk people on the current technology without as much cost to them (actually changing the fundamental structure of the chip inherently requires R&D expenses).

Except they have been doing the R&D on faster and more stable interposers anyways, so I'm not sure what you mean. It's not as if EMIB is a new thing. They've teases it and leaked it multiple times in the past.

 

AMD's"Infinity Fabric" has issues with it. It's not a bad solution with the resources (time+money) they had, and it's decently scalable, but that doesn't make it a perfect solution. Basing it off HyperTransport allows for a nice open ecosystem and ease of development/implementation but it also limits it in certain aspects of performance and how deeply integrated it can be.

 

Intel's solution with EMIB is fairly different in that it, at least appears to be, designed specifically as an on-package interconnect. The 1Tb/s listed on the Stratix diagram would put it at about 4X the performance of the Infinity Fabric interconnect used in Ryzen as an initial implementation (about 1/4 that used by Vega). That's pretty impressive. And if it is, in fact, a purpose built part it likely has a little less overhead than HyperTransport too. Don't know if it'll hit that. Don't know how scalable it is, and Intel has a habit of exaggerating, but if it hits even close that will be impressive.

 

But in reality all of that is moot anyways since we likely won't see it for a looooooooooooooong time on the consumer side. It's a die to die interconnect so if anything we *might* see it on laptops and mobile devices in the near future.

[Jito463]

9 minutes ago, Sniperfox47 said:

*snip*

I'll have to look into it more.  Admittedly, I don't research much of the server side, since my job doesn't really get involved in that aspect much.

[leadeater]

5 hours ago, Sniperfox47 said:

Intel's solution with EMIB is fairly different in that it, at least appears to be, designed specifically as an on-package interconnect. The 1Tb/s listed on the Stratix diagram would put it at about 4X the performance of the Infinity Fabric interconnect used in Ryzen as an initial implementation (about 1/4 that used by Vega). That's pretty impressive. And if it is, in fact, a purpose built part it likely has a little less overhead than HyperTransport too. Don't know if it'll hit that. Don't know how scalable it is, and Intel has a habit of exaggerating, but if it hits even close that will be impressive.

Infinity Fabric is also a purpose built solution for interconnecting dies and chips. It may share DNA with Hypertransport but it's no less purpose built for the task than EMIB is. The implementation on Zen only gets criticized because it's tied to the memory controller but that was likely a design choice to ensure aligned signal timing so it can more efficiently use the bandwidth that is present. Zen actually isn't that significantly limited by the IF badnwidth so long as you don't go below the critical point in ram speed.

 

Far as I could see most of the problems that people talk about like latency in Zen is actually the L3 cache and cache structure and not IF.

[MyName13]

21 minutes ago, leadeater said:

Infinity Fabric is also a purpose built solution for interconnecting dies and chips. It may share DNA with Hypertransport but it's no less purpose built for the task than EMIB is. The implementation on Zen only gets criticized because it's tied to the memory controller but that was likely a design choice to ensure aligned signal timing so it can more efficiently use the bandwidth that is present. Zen actually isn't that significantly limited by the IF badnwidth so long as you don't go below the critical point in ram speed.

 

Far as I could see most of the problems that people talk about like latency in Zen is actually the L3 cache and cache structure and not IF.

And what's the critical point in ram speed?Is there any explanation why IF is tied to memory speeds?It doesn't make much sense to tie your CPU to something that might limit it's performance.

[leadeater]

4 hours ago, MyName13 said:

And what's the critical point in ram speed?Is there any explanation why IF is tied to memory speeds?It doesn't make much sense to tie your CPU to something that might limit it's performance.

Think it was 2666, not sure on that not looked recently. There is a very obvious performance drop when you go below what ever it is.

 

It makes sense as your CPU only stores data in it's caches at it's own cycles and it has the memory controller. IF is tied to the CPU memory controller not the ram. If you cycle the IF faster than the CPU memory controller you're going to be out of sync a lot meaning you lose effective bandwidth.

[MyName13]

6 hours ago, leadeater said:

Think it was 2666, not sure on that not looked recently. There is a very obvious performance drop when you go below what ever it is.

 

It makes sense as your CPU only stores data in it's caches at it's own cycles and it has the memory controller. IF is tied to the CPU memory controller not the ram. If you cycle the IF faster than the CPU memory controller you're going to be out of sync a lot meaning you lose effective bandwidth.

So why don't intel CPUs benefit so much from fast ram?

[leadeater]

11 minutes ago, MyName13 said:

So why don't intel CPUs benefit so much from fast ram?

They do, it's just less apparent and mostly effects minimum frame rates if we are talking about gaming. AVX workloads are heavily impacted by RAM for Intel.

 

Either way you're also comparing completely different architectures, within a Zen CCX it is much like Intel but across CCX's is where it's a different story.

 

More bandwidth is better of course but you don't actually need any more than what the CPU can use, right now doubling the Infinity Fabric bandwidth will have minimal to no real effect on performance. Lowering the CPU memory controllers latency to access RAM and lowering L3 cache latency will do more.