[Techpowerup] AMD "Zen" CPU Prototypes Tested, "Meet all Expectations"

[III-V]

Sandy was way more important than Nehalem since Nehalem was still on the old Core architecture.

Did you forget to have your coffee this morning? Core 2 (Conroe) came before Nehalem, which improved upon it.

 

[DrMrLordX]

An APU part in this timeframe is very unlikely.

Not sure when the APUs will show up, but they're all supposed to be 2017. Q2 maybe? June? AMD seems to like June/July launches.

 

[Arachnotronic]

Did you forget to have your coffee this morning? Core 2 (Conroe) came before Nehalem, which improved upon it.

dark zero just says things to be controversial.

 

[myocardia]

Not sure when the APUs will show up, but they're all supposed to be 2017. Q2 maybe? June? AMD seems to like June/July launches.

Haha, so much so that I'm willing to bet that nobody around here would have argued if you had instead said "AMD goes out of their way to release nearly all of their products in June or July". Then again, it is in fact a perfect time to release CPUs and not at all a bad time to release GPUs, so I don't blame them for doing so.

 

[dark zero]

Did you forget to have your coffee this morning? Core 2 (Conroe) came before Nehalem, which improved upon it.

However it was Sandy who brought a MASSIVE revamp. Nehalem was good due HT and the native Quad Core, however the improvement weren't as abyssmal than Nehalem to Sandy.

 

[mrmt]

However it was Sandy who brought a MASSIVE revamp. Nehalem was good due HT and the native Quad Core, however the improvement weren't as abyssmal than Nehalem to Sandy.

You are forgetting that you are comparing a two generations jump because between Nehalen and Sandy Bridge we had Westmere. If we allow Nehalen the same courtesy then you would have to compare it against Conroe Xeon.

The native quad-core was less of an issue at the time in terms of performance than. The issue with Intel MCM FSB quad cores was that you could synchronize the first and the second thread very quickly but the third and the fourth would have to be synchronized via FSB, so in effect your quad-core server would be no faster than a 2P with 2 dual core processors, e.g., a problem already known for the software industry. On 4P and 8P servers the Caneland platform also made available a point-to-point link with the full bandwidth of the FSB of the platform, so no news here too.

And how did FSB fare in real life? IIRC up to four cores thread sync occurred as fast in AMD native quad-cores, only when going beyond that the FSB started to show its age. This is why partners like Cray stayed with AMD even after AMD was outclassed with Conroe.

In terms of relevancy I think Sandy Bridge-EP was a very interesting product, it was undisputed the king of the hill in every single scenario you could think of, but it didn't bring revolutionary changes to the platforms like Nehalen, and it didn't that much change in terms of TCO as Ivy Bridge-EP did. IVB-EP and EX caused a big splash on the markets because the economics were so much better than previous processors that 2P processors cannibalized a lot of share of 4P servers and 4P servers cannibalized a lot of share of big iron servers.

 

[Fjodor2001]

Regarding 8 core Zen, has there been any mentioning of an 8 core Zen APU with decent iGPU? I'm not talking about some gaming-capable monster with HBM and a huge number of GPU cores. But something close to Intel HD530 or AMD Kaveri iGPU performance.

Wouldn't that be a useful SKU for companies demanding high end desktop CPUs? There are lots of workplaces where the employees just need a CPU with excellent MT performance and decent GPU performance. It would be inefficient for them to have to add a discrete GPU to an 8 core Zen CPU. Don't you think?

 

[ShintaiDK]

has there been any mentioning of an 8 core Zen APU

No, its 2 and 4 from the info given.

 

[Fjodor2001]

No, its 2 and 4 from the info given.

Ok, I hope there will be a Zen SKU like the one I mentioned eventually. Strange that Intel doesn't have one either. Seems unnecessary to pair an 8 core CPU with a discrete GPU for workstation use.

 

[Arachnotronic]

Ok, sad. I hope there will be a Zen SKU like the one I mentioned eventually. Strange that Intel doesn't have one either. Seems unnecessary to pair an 8 core CPU with a discrete GPU for workstation use.

APUs are designed mainly for laptops, and an 8 core CPU + iGPU doesn't make sense for laptops, even high performance ones, since you would basically be giving up so much frequency at full 8 core load that you might as well go with 4 higher clocked CPU cores. The market for such a SKU would be so small that I doubt it's worth the time for either Intel or AMD to build one.

 

[cbn]

APUs are designed mainly for laptops, and an 8 core CPU + iGPU doesn't make sense for laptops, even high performance ones, since you would basically be giving up so much frequency at full 8 core load that you might as well go with 4 higher clocked CPU cores. The market for such a SKU would be so small that I doubt it's worth the time for either Intel or AMD to build one.

Intel has 25W Skylake quad cores.....and the clocks are pretty good on those.

So eight cores for a laptop isn't out of reach provided the TDP is at least 45W.

P.S. I was actually thinking a 16C 14nm APU with eight cherry picked cores might do pretty well against a Intel Cannonlake 8C laptop (re: what it loses in process tech, it kinda makes up with better binning)

 

[Fjodor2001]

APUs are designed mainly for laptops, and an 8 core CPU + iGPU doesn't make sense for laptops, even high performance ones, since you would basically be giving up so much frequency at full 8 core load that you might as well go with 4 higher clocked CPU cores. The market for such a SKU would be so small that I doubt it's worth the time for either Intel or AMD to build one.

An 8 core high frequency APU/CPU is certainly not designed for laptop usage, but workstation usage. No way you could fit a ~95W TDP SKU in a laptop. And having to add a discrete GPU is unnecessary for a lot or workstation tasks. A HD530 class iGPU will do fine in most cases.

 

[Arachnotronic]

An 8 core high frequency APU/CPU is certainly not designed for laptop usage, but workstation usage. No way you could fit a ~95W TDP SKU in a laptop. And having to add a discrete GPU is unnecessary for a lot or workstation tasks. A HD530 class iGPU will do fine in most cases.

You're not thinking clearly. The workstation market itself is fairly small from a unit perspective and I would imagine that the vast majority of workstation workloads require some sort of reasonable GPU power.

So you are suggesting that Intel/AMD build specialized SKUs targeted at one small-sub segment of the already small workstation market?

Notice how today Intel doesn't even have a specialized set of workstation processors; Xeon E3 are rebranded consumer desktop chips (which are actually based on dies designed first for laptops) and Xeon E5 which servers the server market.

tl;dr -- market's not big enough to warrant such a SKU.

 

[ShintaiDK]

So now the "wasted" IGP die is good enough anyway?

 

[cbn]

tl;dr -- market's not big enough to warrant such a SKU.

It would have to come from a cut down Server APU.

And 8C/16 Zen + dGPU would probably be a lot cheaper.

 

[itsmydamnation]

I find this thread hilarious its like no one knows what data center processors are used for today. NUMA performance for the vast majority is Irrelevant, what matter is latency for memory access for each core.

ESXi, Hyper V, KVM , Zen are all NUMA aware hypervisors, So even if the interconnect sucks for Zen (which i dont expect it to) So long as each core has suitable access to its local memory everyone is happy.

as a percentage
1.how many guests have 2 threads assigned
2.how many guests have 4 threads assigned
3.how many guests have 8 threads assigned
4.how many guests have 16 threads assigned

i've done design in many data centres for many clients. In my experience option 2 is by far the most used in my experience. The current Devops mantra (which i hate, or the abuse of it) leads to more guests of less resources even on high throughput applications.

 

[Fjodor2001]

You're not thinking clearly. The workstation market itself is fairly small from a unit perspective and I would imagine that the vast majority of workstation workloads require some sort of reasonable GPU power.

So you are suggesting that Intel/AMD build specialized SKUs targeted at one small-sub segment of the already small workstation market?

Notice how today Intel doesn't even have a specialized set of workstation processors; Xeon E3 are rebranded consumer desktop chips (which are actually based on dies designed first for laptops) and Xeon E5 which servers the server market.

tl;dr -- market's not big enough to warrant such a SKU.

What do you think most Intel HEDT CPUs are used for? Workstation use. That's right. Compiling SW, editing videos and similar. You need excellent MT performance, and only a decent iGPU.

 

[Fjodor2001]

So now the "wasted" IGP die is good enough anyway?

For workstation use you just need a decent iGPU, not Iris Pro class iGPU. The latter is waste.

 

[Arachnotronic]

What do you think most Intel HEDT CPUs are used for mostly? Workstation use. That's right. Compiling SW, editing videos and similar. You need excellent MT performance, and only a decent iGPU.

You are not really getting my argument. Please re-read it.

 

[Fjodor2001]

You are not really getting my argument. Please re-read it.

I get your argument, but I disagree.

 

[ShintaiDK]

For workstation use you just need a decent iGPU, not Iris Pro class iGPU. The latter is waste.

What about all the stories that the die space could be used for "moar cores"?

Anyway, as Arachnotronic also says. You are trying to create a product that doesn't have its existential worth.

 

[Fjodor2001]

What about all the stories that the die space could be used for "moar cores"?

I've never said that no iGPU should be included. I'm just saying it doesn't have to occupy 50% of the die area like on Skylake.

Also, there can be different desktop SKUs with different size iGPU. Just like they have various SKUs for CPU performance. I'd reduce the number of CPU options and increase the iGPU options. Today we have tons of various SKUs differentiated only by CPU (core count, frequency, TDP).

Anyway, as Arachnotronic also says. You are trying to create a product that doesn't have its existential worth.

Again, this is where we disagree.

 

[dark zero]

APUs are designed mainly for laptops, and an 8 core CPU + iGPU doesn't make sense for laptops, even high performance ones, since you would basically be giving up so much frequency at full 8 core load that you might as well go with 4 higher clocked CPU cores. The market for such a SKU would be so small that I doubt it's worth the time for either Intel or AMD to build one.

Err... some gamers would allow that.... however the cost might be insane..

 

[PPB]

You're not thinking clearly. The workstation market itself is fairly small from a unit perspective and I would imagine that the vast majority of workstation workloads require some sort of reasonable GPU power.

So you are suggesting that Intel/AMD build specialized SKUs targeted at one small-sub segment of the already small workstation market?

Notice how today Intel doesn't even have a specialized set of workstation processors; Xeon E3 are rebranded consumer desktop chips (which are actually based on dies designed first for laptops) and Xeon E5 which servers the server market.

tl;dr -- market's not big enough to warrant such a SKU.

Nope, you dont need reasonable GPU power for a whole lot of workstation use cases. The loads that need GPU power in the other ones that probably do ,will need far more grunt that any iGP would give you ATM anyway. So it is better to fit more cores in that power and TDP budget that give you unwanted die space in a iGP.

 

[bystander36]

You are forgetting that you are comparing a two generations jump because between Nehalen and Sandy Bridge we had Westmere. If we allow Nehalen the same courtesy then you would have to compare it against Conroe Xeon.

If you look at clock for clock comparisons, they are very closely matched as well. It's only due to the very conservatively clocked i7 920's that there was a big gap.