64 core AMD Rome CPU

[samcool55]

3 minutes ago, leadeater said:

That's what I said, single core perf is just that. IPC is IPC. Single core performance is the result of IPC and your frequency (clocks) for the task being run. When most people talk about IPC, as in Intel's being better they are actually comparing and talking about the single core perf.

True, actually, this might become interesting very soon.

Right now the IPC of the ryzen 2 cpu's and 8th-gen intel cpu's (can't find numbers of 9th-gen but afaik the IPC is the same) are quite close.

Intel is a few % ahead and right now the main reason why intel is ahead when it comes down to single-core performance is clockspeed.

 

So, if AMD gets ahead IPC wise, which is possible because the gap isn't that big, and manages to increase the clockspeed a bit, which is again possible but i have no idea if that will actually happen, we might see depending how it all goes that AMD will be ahead of intel with a lower clock speed.

 

Again, 0 source for it and just thinking about "what if" basically, but it could start an interesting debate (or fanboy fight depending who you ask).

Not all people understand that cpu X can be faster than cpu Y while the clockspeed of X is lower than Y.

 

It's going to be an interesting few months, that's for sure.

[leadeater]

6 minutes ago, samcool55 said:

or fanboy fight depending who you ask

 

[kuddlesworth9419]

All these cores are very nice and all but they are useless for gaming considering games mostly only use like 4 cores, older games only use like 1 core maybe 2 if your lucky. I don't even know if any use 6 cores.

[leadeater]

2 minutes ago, VegetableStu said:

4 for CS, 6 for GTA, 4 for CIV, 2 for OBS, 6 for chrome...

If you have no friends you could easily make many multiplayer sessions of Civ off the one CPU and play yourself 

[GoldenLag]

5 minutes ago, kuddlesworth9419 said:

All these cores are very nice and all but they are useless for gaming considering games mostly only use like 4 cores, older games only use like 1 core maybe 2 if your lucky. I don't even know if any use 6 cores.

Battlefield uses 6 threads last time i checked. And Civ is known to use quite a lot of cores. Most modern titles use 4 or more threads not to mention Game engines adapting to use extra cores. 

 

I mean, 6 cores is standard these days. Any extra is sort of just bonus for the future

[TechyBen]

9 minutes ago, GoldenLag said:

Battlefield uses 6 threads last time i checked. And Civ is known to use quite a lot of cores. Most modern titles use 4 or more threads not to mention Game engines adapting to use extra cores. 

 

I mean, 6 cores is standard these days. Any extra is sort of just bonus for the future

Ashes of the Singularity is the only true multicore game benchmark software out there.

[leadeater]

So I was wondering since the I/O die is 14nm could AMD dual source that from TSMC and GloFo?

[porina]

36 minutes ago, kuddlesworth9419 said:

All these cores are very nice and all but they are useless for gaming considering games mostly only use like 4 cores, older games only use like 1 core maybe 2 if your lucky. I don't even know if any use 6 cores.

This is the server stuff, so not intended for gaming. What we can see is what Ryzen 3000 series might be like, and I think this is when the fight with Intel really kicks off.

[Doobeedoo]

Damn those chiplets, dies looking nice. I take it CCX is 4 core and not 8 core. 

Either way, architectural improvements all great and for consumer chips can't wait to see how much they can be clocked this time. Reaching 5GHz wold be amazing! 

[williamcll]

Reminds me of the 140 core ARM chip baidu made sometime ago.

[cj09beira]

1 hour ago, leadeater said:

So I was wondering since the I/O die is 14nm could AMD dual source that from TSMC and GloFo?

maybe not tsmc, but samsung why not

[Taf the Ghost]

2 hours ago, leadeater said:

That's what I said, single core perf is just that. IPC is IPC. Single core performance is the result of IPC and your frequency (clocks) for the task being run. When most people talk about IPC, as in Intel's being better they are actually comparing and talking about the single core perf which is a function of IPC and clocks.

Plus, these days, it matters what Instructions are being run. As we've found with the current AMD & Intel generations, the actual IPC difference is in the memory systems and AVX2. It makes Intel's IPC advantage pretty large in workstation-type tasks (about 25% by The Stilts research/heavy workstation testing), down to about 3% in 1080p Gaming, clock for clock. 

[Brooksie359]

10 minutes ago, Taf the Ghost said:

Plus, these days, it matters what Instructions are being run. As we've found with the current AMD & Intel generations, the actual IPC difference is in the memory systems and AVX2. It makes Intel's IPC advantage pretty large in workstation-type tasks (about 25% by The Stilts research/heavy workstation testing), down to about 3% in 1080p Gaming, clock for clock. 

The AVX2 disadvantage should be fixed with the larger memory bandwidth they have with Zen 2. Going from 128bit to 256 will help them alot. 

[Taf the Ghost]

20 minutes ago, Brooksie359 said:

The AVX2 disadvantage should be fixed with the larger memory bandwidth they have with Zen 2. Going from 128bit to 256 will help them alot. 

For the Epyc server parts, they had more memory bandwidth than Intel, which is what made for some of the hilarious early AVX512 benchmarks. (Optimized code is still a long ways off.) The issues were how the front-end handled the situation, which appears to have one of Zen 1's biggest bottlenecks. We'll see what happens, but this the "low hanging fruit" stuff that gets mentioned.

[Brooksie359]

16 minutes ago, Taf the Ghost said:

For the Epyc server parts, they had more memory bandwidth than Intel, which is what made for some of the hilarious early AVX512 benchmarks. (Optimized code is still a long ways off.) The issues were how the front-end handled the situation, which appears to have one of Zen 1's biggest bottlenecks. We'll see what happens, but this the "low hanging fruit" stuff that gets mentioned.

They said that they had more room to add 256bit units rather than 128 but units so it should be much better. It's why they said double performance in that area. 

[schwellmo92]

3 hours ago, kuddlesworth9419 said:

All these cores are very nice and all but they are useless for gaming considering games mostly only use like 4 cores, older games only use like 1 core maybe 2 if your lucky. I don't even know if any use 6 cores.

You have obviously never played either of the two last Assasins Creeds

[Fasauceome]

6 hours ago, leadeater said:

Yep but that's the problem with what people term as IPC when they mean single core perf in actuality.

Well the 7nm shrink and the microarchitecture refinement is what leads to a huge gain in IPC, and the increased efficiency allows for higher clocks, which will possibly be higher than a 25% gain. I do really see a 25% increase to only IPC a possible outcome

[Fasauceome]

4 hours ago, samcool55 said:

Different frequencies make it impossible to say how good the IPC is of a cpu compared to another one.

Actually instruction sets do that. You can change the clock speeds of a CPU to match a competitor, but they're still too different to compare apples to apples.

[porina]

1 hour ago, Taf the Ghost said:

Plus, these days, it matters what Instructions are being run. As we've found with the current AMD & Intel generations, the actual IPC difference is in the memory systems and AVX2. It makes Intel's IPC advantage pretty large in workstation-type tasks (about 25% by The Stilts research/heavy workstation testing), down to about 3% in 1080p Gaming, clock for clock. 

There's different ways to cut it. I prefer to look at architecture performance, as far as is practical, in the absence of external limitations like ram bandwidth. Of course, in real world tasks that is hard to avoid, but the limits will vary depending on wider configuration so isn't even constant.

 

For those that haven't seen it, I did test earlier a bunch of mostly synthetic tests, and you can see how Skylake vs Zen vs Zen+ differ, with and without HT/SMT. I hope to add Zen2 to this as soon as practical, but that'll probably have to wait until consumer models are launched.

1 hour ago, Brooksie359 said:

The AVX2 disadvantage should be fixed with the larger memory bandwidth they have with Zen 2. Going from 128bit to 256 will help them alot. 

The biggest weakness of Zen/Zen+ architecture compared to Intel is the size of the AVX unit. The upgrade in Zen2 will bring AMD near enough to per-core per clock parity with Intel. Ram bandwidth limitations for sure are also a limitation on both sides, but is ever more noticeable as AVX in general gets faster and/or with more cores relative to the available ram bandwidth.

[Taf the Ghost]

35 minutes ago, porina said:

There's different ways to cut it. I prefer to look at architecture performance, as far as is practical, in the absence of external limitations like ram bandwidth. Of course, in real world tasks that is hard to avoid, but the limits will vary depending on wider configuration so isn't even constant.

 

For those that haven't seen it, I did test earlier a bunch of mostly synthetic tests, and you can see how Skylake vs Zen vs Zen+ differ, with and without HT/SMT. I hope to add Zen2 to this as soon as practical, but that'll probably have to wait until consumer models are launched.

The biggest weakness of Zen/Zen+ architecture compared to Intel is the size of the AVX unit. The upgrade in Zen2 will bring AMD near enough to per-core per clock parity with Intel. Ram bandwidth limitations for sure are also a limitation on both sides, but is ever more noticeable as AVX in general gets faster and/or with more cores relative to the available ram bandwidth.

Nice testing. Sorry I missed it the first go around.

 

As for the comparisons, bringing in the full AVX2 should really close up the gap. Apparently Icelake is going wider than Zen (Zen is wider than Skylake) at the front end, so it's interesting to see AMD putting a complete rework of the front-end for Zen2. Though scuttlebutt is that the backend will be completely different for Zen3. AMD wasn't joking about "Tock Tock Tock" strategy.

 

Though AMD either breaking even or getting a lead is going to be down to Intel's Node disaster. Icelake should have been out this year. Desktop might not be until 2020. Not that the 9th Gen isn't a paper launch until December, like 8th Gen was as well.

[porina]

16 minutes ago, Taf the Ghost said:

Though AMD either breaking even or getting a lead is going to be down to Intel's Node disaster. Icelake should have been out this year. Desktop might not be until 2020. Not that the 9th Gen isn't a paper launch until December, like 8th Gen was as well.

I am looking at it from my own need perspective. Zen family is already more than good enough for non-AVX workloads. AVX is a weak point for my interest areas, and that goes away with Zen 2. Assuming no surprises that will be my switch over point, unless AVX-512 gets a lot of traction before then, which seems unlikely.

 

By paper launch on Intel, do you mean the two phase release, high end initially, lower parts later? While availability of 9900k isn't great, I think it is better than the situation with 8700k. One UK retailer I'm following has had continuous stock available for a week or so now, although many others are still showing out of stock/preorder.

[Taf the Ghost]

5 minutes ago, porina said:

I am looking at it from my own need perspective. Zen family is already more than good enough for non-AVX workloads. AVX is a weak point for my interest areas, and that goes away with Zen 2. Assuming no surprises that will be my switch over point, unless AVX-512 gets a lot of traction before then, which seems unlikely.

 

By paper launch on Intel, do you mean the two phase release, high end initially, lower parts later? While availability of 9900k isn't great, I think it is better than the situation with 8700k. One UK retailer I'm following has had continuous stock available for a week or so now, although many others are still showing out of stock/preorder.

AVX512 is for DDR5. All of the instructions aren't even in hardware units yet. I can see why Intel would add them, but it also makes sense for AMD to just wait until the dies are small enough for those huge units to make sense.

 

On the 8th & 9th Gen launches, they have an initial supply that's quite small and sells out. Keeps the prices high going into the actual availability launch in December. I think 8th Gen had like 1000 8700k in each major European market. It was something very small for a generally high selling product.

[Stefan Payne]

7 hours ago, kuddlesworth9419 said:

All these cores are very nice and all but they are useless for gaming considering games mostly only use like 4 cores, older games only use like 1 core maybe 2 if your lucky. I don't even know if any use 6 cores.

...because you hadn't had more than 4 Cores in over 10 years in the Desktop Market....

 

4 Cores with SMT came out 10 years ago, at the end of 2008, but even before that, in 2006 or 7 the Dual Core 2 Duo (called Core 2 Quad) was released. So we're with 4 Cores for 12 years at least...

 

And software takes time to grow into new shoes. Did take long to profit from a second Core (althoigh there was a couple of games in the past that did that anyway, IIRC Quake or something like that in the beginning 2000s when Dual Celeron setups were popular for a short time)

 

So the Hardware usually comes first....

 

And there are already some games that benefit off of more than 4 or 6 cores already. They're just not that common YET...

[Fasauceome]

25 minutes ago, Stefan Payne said:

...because you hadn't had more than 4 Cores in over 10 years in the Desktop Market....

 

4 Cores with SMT came out 10 years ago, at the end of 2008, but even before that, in 2006 or 7 the Dual Core 2 Duo (called Core 2 Quad) was released. So we're with 4 Cores for 12 years at least...

 

And software takes time to grow into new shoes. Did take long to profit from a second Core (althoigh there was a couple of games in the past that did that anyway, IIRC Quake or something like that in the beginning 2000s when Dual Celeron setups were popular for a short time)

 

So the Hardware usually comes first....

 

And there are already some games that benefit off of more than 4 or 6 cores already. They're just not that common YET...

Of course there were those really nice 6 core xeons and i7s that were HEDT on the consumer socket, which may have been a subtle driving force for higher core count use in games.

[leadeater]

7 hours ago, cj09beira said:

maybe not tsmc, but samsung why not

TSMC fabs the 7nm chiplets and has 14nm capability so I don't see why they wouldn't source from them for the I/O dies, they still have contracts with GloFo so they also make sense since they are obligated to fab some stuff through them.