AMD's Big HPC Swing: Exascale Heterogeneous Processor [Updated]

[Notional]

Don't xeons have hyper threading? therefore the new 28 core skylake is really 46 cores? or is that just wrong logic?

and don't thew new skylake xeons have octacpu for mobos support?

 

Hyperthreading allows for a single core to handle two threads (data from a program). They still share resources, so you still have a primary thread, and whatever resources are leftover could be used by another thread. So a 28 core is still 28 physical cores, but can handle 46 threads at best.

[Arty]

Hyperthreading allows for a single core to handle two threads (data from a program). They still share resources, so you still have a primary thread, and whatever resources are leftover could be used by another thread. So a 28 core is still 28 physical cores, but can handle 46 threads at best.

i wonder how that would compare to amds 32 core.

hopefully someone benches them 

[dalekphalm]

i wonder how that would compare to amds 32 core.

hopefully someone benches them 

AMD Zen Cores will work in a similar fashion to Intel Xeon cores.

 

We won't know specifics or performance metrics until they are both in market though.

[HalGameGuru]

I second Dalek's post, almost literally.

 

 

 

 

Deprecated:

The Zen cores are going to hyperthread similarly to Intel's.

[Ekpyrosis]

Hyperthreading is just the name for Intel's implementation of simultanous multi-threading.

That doesn't mean other companies can't implement their version of SMT.

[Emanuel_M]

Don't get mad, why do I always see AMD making CPU's for business first instead of consumer?

[dalekphalm]

Don't get mad, why do I always see AMD making CPU's for business first instead of consumer?

Because that's where the money is, frankly.

 

Those business class CPU's are what's paying for the R&D for your sick FX-9001 Leet Edition 25 GHz CPU.

[Emanuel_M]

Because that's where the money is, frankly.

Those business class CPU's are what's paying for the R&D for your sick FX-9001 Leet Edition 25 GHz CPU.

But then they spend all their budget for business, then they have to kill themselves to make a consumer product.

[dalekphalm]

But then they spend all their budget for business, then they have to kill themselves to make a consumer product.

The fact of the matter is that AMD doesn't have enough money right now. Business class gives them more money then consumer class. This much is very clear.

 

If AMD had focused on the consumer class market right now, they wouldn't be here to give you that awesome consumer CPU.

 

By focusing on the business class first, they can generate the income needed to develop the consumer class CPU's that we all want.

[Slinkey]

Because that's where the money is, frankly.

 

Those business class CPU's are what's paying for the R&D for your sick FX-9001 Leet Edition 25 GHz CPU.

Well it ain't paying much since their R&D is at a low!

[Notional]

But then they spend all their budget for business, then they have to kill themselves to make a consumer product.

 

Most consumer products are cut down business parts anyways, so hardly an issue.

[patrickjp93]

Don't xeons have hyper threading? therefore the new 28 core skylake is really 46 cores? or is that just wrong logic?

and don't thew new skylake xeons have octacpu for mobos support?

 

Hyperthreading allows for a single core to handle two threads (data from a program). They still share resources, so you still have a primary thread, and whatever resources are leftover could be used by another thread. So a 28 core is still 28 physical cores, but can handle 46 threads at best.

I hate to be a nitpicking douche, but 2 * 28 is 56. And furthermore, if the cores have duplicate resources, or the threads are carefully crafted to use different resources at the same time, then you can get big performance gains. In fact IBM takes this to an extreme with 8-Way SMT on the Power 8 CPUs. The flagship is 12 cores at 4.3GHz, but each chip handles up to 96 threads, and in fact for a core to become active, all 8 threads must be in use (in order to be power-efficient).

[Notional]

I hate to be a nitpicking douche, but 2 * 28 is 56. And furthermore, if the cores have duplicate resources, or the threads are carefully crafted to use different resources at the same time, then you can get big performance gains. In fact IBM takes this to an extreme with 8-Way SMT on the Power 8 CPUs. The flagship is 12 cores at 4.3GHz, but each chip handles up to 96 threads, and in fact for a core to become active, all 8 threads must be in use (in order to be power-efficient).

 

Lol yeah. it was like 4 am, I just repeated the numbers, but yeah you're right on that one. As for Intel's SMT, they themselves claim you can only gain a max of 30% performance with hyperthreading (and at worst a little negative performance), so in Intel's implementation, the threads very much shares resources.

[patrickjp93]

Lol yeah. it was like 4 am, I just repeated the numbers, but yeah you're right on that one. As for Intel's SMT, they themselves claim you can only gain a max of 30% performance with hyperthreading (and at worst a little negative performance), so in Intel's implementation, the threads very much shares resources.

I think you misunderstand. If both threads use AVX at the same time, bad. If one is using AVX while another is calculating a bunch of branch decision variables, good. The use of hyperthreading does split core resources, but as long as the same 2 are not in use at the same time, you can get far better results.

[Notional]

I think you misunderstand. If both threads use AVX at the same time, bad. If one is using AVX while another is calculating a bunch of branch decision variables, good. The use of hyperthreading does split core resources, but as long as the same 2 are not in use at the same time, you can get far better results.

 

Indeed, but Intel still only states a max of 30% performance gain. And the negative gain, that is achievable is probably the reason you can manually disable hyperthreading in bios.

[Ohlyver]

8000$ a chip... thats a good deal. 32gb's of cache? Thats insane.

But how many PCI lanes?

[Deletive]

But how many PCI lanes?

For this type for this type of exascale processing, PCI lanes, isn't much of the biggest factor, Don't many many PCI lanes for a NIC. 

[Ohlyver]

For this type for this type of exascale processing, PCI lanes, isn't much of the biggest factor, Don't many many PCI lanes for a NIC.

Was expecting someone to answer with :Over 9000!!!

And shouldn't we expect the PCI standard to evolve a bit before this chip is released? ( in 2020)

[Deletive]

Was expecting someone to answer with :Over 9000!!!

And shouldn't we expect the PCI standard to evolve a bit before this chip is released? ( in 2020)

 

Yeah, PCI will change over the next 5 years, probably more uses for it such as Storage and such. Nothing too major, might even see a replacement in the server space (like how NVlink is becoming more of a full image on PowerPC line.)

[patrickjp93]

Indeed, but Intel still only states a max of 30% performance gain. And the negative gain, that is achievable is probably the reason you can manually disable hyperthreading in bios.

Intel's projected maximum is for a mixed workload. In HPC they can extract more with careful crafting. If you throw heavy branching in one thread and data streaming and crunching in the other, all the pipeline stalls and branch misses get covered up. The gained performance can be much greater.