Intel to possibly announce even more impressive chip than 7980XE at Computex

[Deleted member 66701]

Makes me wonder how they will enforce distinction between TR and Epyc when they are happy to stack cores on the consumer part.

EPYC is 8 channel mem and TR is only 4 channel, meaning on 24 and 32 core TR two cores will have no direct access to memory and will have to go through the infinity fabric. This will increase memory latencies and will affect applications sensitive to that (i.e. database systems - one of EPYCs main markets). Shouldn't affect typical HEDT type workloads such as video and 3D rendering. That's the differentiation between EPYC and TR.

Looking forward to dropping a 32 core into my system in August

 

[Panos]

Well the problem for Intel is they rely on monolithic dies, you can only make them a certain size and 28 cores on 14nm seems to be the limit, that CPU is 700mm2^, its vast and with that because of yields very expensive to make, you might get 2 or 3 such CPU's out of one $10,000 wafer.

AMD with Infinity Fabric get much higher yields with much higher core count CPU's because its a bunch of little ones 'glued together'.
With that said its not that AMD are cannibalising EPYC with Threadripper 2, EPYC 2 will have 48 and 64 cores.

And next year with Zen 2 the future is unknown Going for half the die size or double the cores per CCX

 

[Rroff]

And next year with Zen 2 the future is unknown Going for half the die size or double the cores per CCX?

It would make most sense of them to go for best yields/profit in the short term with the option of scaling up core count, etc. in the longer run. Whether AMD will do that though is another matter.

(It also gives them more manoeuvring space to answer anything by Intel).

 

[4K8KW10]

It would make most sense of them to go for best yields/profit in the short term with the option of scaling up core count, etc. in the longer run. Whether AMD will do that though is another matter.

(It also gives them more manoeuvring space to answer anything by Intel).

But this anything from Intel will have to wait for some later times, meanwhile AMD can and should go very aggressive and take as much market as possibly they can squeeze, equal many sales/profit, too. At some future point, Intel may inevitably catch, but AMD should have plenty of time to work on Zen 5.

 

[Rroff]

But this anything from Intel will have to wait for some later times

That is an assumption however. Despite the issues (and some hilarious ineptitude) counting Intel out at this point would be premature.

 

[Beren]

I think from the news we already have that we are going to get a 48 then 64 core Epyc chip that it is pretty certain AMD are going to shift to an 8 core ccx on 7nm. That is going to be the wedge they drive into the gap they have created in that market.

They cant be certain of yields straight out on the process so first they are going to 48 core which will maximise yields then go to a full 64 core once yields are high enough to make it viable.

It also means that we could see a 3800x as a 12 core desktop part in 7nm which could be released as early as next April and then a 4800x as a 16 core in 2020.

We can then expect a 6 core 12 thread APU with double the Vega cores to replace the 2400g 2h2019. And an 8 core 16 thread APU with Navi cores in 2020.

Basically I think that AMD should be aggressive but controlled because they have an opportunity to build on the last 18 months with a really strong ramp over the next 24 months.

<edit> It is this ability to target all 4 major market sectors with one major architectural shift that makes AMD such a beast and what they have done just so clever.

 

[Beren]

That is an assumption however. Despite the issues (and some hilarious ineptitude) counting Intel out at this point would be premature.

Completely agree, with their reserves and R&D budget you have to believe that they will respond with something at least as good. That is why AMDs roadmap is so awesome Intel are very likely to come up with even better at some point during this cycle.

 

[Dicehunter]

Seeing as the Xeon that this is based on costs £10,000 it wouldn't surprise me if it came in at around half that.

 

[BongoHunter]

EPYC is 8 channel mem and TR is only 4 channel, meaning on 24 and 32 core TR two cores will have no direct access to memory and will have to go through the infinity fabric. This will increase memory latencies and will affect applications sensitive to that (i.e. database systems - one of EPYCs main markets). Shouldn't affect typical HEDT type workloads such as video and 3D rendering. That's the differentiation between EPYC and TR.

Looking forward to dropping a 32 core into my system in August

TR is not that competitive in the data center as the motherboards are simply not suitable with the lower memory capacity limits and a complete lack of enterprise features like IPMI etc, current board's have no optimisations for getting the typically required rack density either.

Epyc works well from a TCO point of view as Intel can't match the memory capacity in a single socket or 2 socket server - and the generous amount of PCIE lanes you get ensure you can attach as many GPU / NVME / 10-100 GB networking as required in what is essentially a cheaper platform

 

[Panos]

I think from the news we already have that we are going to get a 48 then 64 core Epyc chip that it is pretty certain AMD are going to shift to an 8 core ccx on 7nm. That is going to be the wedge they drive into the gap they have created in that market.

They cant be certain of yields straight out on the process so first they are going to 48 core which will maximise yields then go to a full 64 core once yields are high enough to make it viable.

It also means that we could see a 3800x as a 12 core desktop part in 7nm which could be released as early as next April and then a 4800x as a 16 core in 2020.

We can then expect a 6 core 12 thread APU with double the Vega cores to replace the 2400g 2h2019. And an 8 core 16 thread APU with Navi cores in 2020.

Basically I think that AMD should be aggressive but controlled because they have an opportunity to build on the last 18 months with a really strong ramp over the next 24 months.

<edit> It is this ability to target all 4 major market sectors with one major architectural shift that makes AMD such a beast and what they have done just so clever.

Yeah, we are just 10 months away from 12 core main stream part. That's bonkers. And Intel cannot do much against this, with their monolithic CPUs designs planned until 2021.
They will have to change sockets on "mainstream" market and start cannibalising HEDT CPUs just months after the 8 core CL comes out. Otherwise going to leave AMD having a field day eating fast market share.
As for HEDT AMD will have no competition. And tbh already doesn't have. 1950X already goes for £700 while 7900X costs £830 and 7960XE £1500. Let alone X399 is better platform.

 

[smilingcrow]

I think from the news we already have that we are going to get a 48 then 64 core Epyc chip....

They cant be certain of yields straight out on the process so first they are going to 48 core which will maximise yields then go to a full 64 core once yields are high enough to make it viable.

Are you suggesting they will use dies that are faulty but can be harvested to work as 12 cores rather than 16 or that they will have a separate smaller design that natively uses 6 core CCXs?

 

[Beren]

Are you suggesting they will use dies that are faulty but can be harvested to work as 12 cores rather than 16 or that they will have a separate smaller design that natively uses 6 core CCXs?

I am guessing. My guess is that they go straight for an 8 core ccx, they can then build stock quickly by harvesting 4, 6 and 8 core chips depending on viability to populate all their skus.
I assume that is what every multicore manufacturer does.