Meet Sunny Cove: Intel's architecture that finally will improve IPC

[Greg S]

Something to look forward to: For the first time in several years, Intel is going to improve the instructions per clock performance of its CPU lineup. Architecture changes allow for more parallel instructions to be executed per cycle, but single-threaded tasks will also become faster across the board.

Looking ahead to 2019, Intel has revealed its plans for Sunny Cove architecture. The next-generation CPU will be built on a 10nm process and is designed to focus improving instruction execution parallelization as well as efficiency. Both Intel Xeon and Intel Core series processors will be making use of Sunny Cove.

To improve performance, Sunny Cove brings support for AVX-512 instructions. Accelerated vector operations will speed up neural networks as well as help with encryption and compression. Up to 75 percent more performance can be expected in these specific workloads.

Attempting to handle more instructions per clock cycle requires more cache space. First level caches has been expanded by 50 percent. Second level cache has also been expanded by 50 percent, but will vary based on the target market of specific CPUs.

To further improve parallel execution, Intel has added additional reservation stations that pass along instructions, doubling the number from two to four. Now up to five instructions can be sent per cycle as opposed to four on Skylake. Extra execution units have also been added for LEA instructions and vector shuffling.

Skylake CPUs can manage two loads and one store operation per clock cycle, while Sunny Cove architecture will be able to handle two loads and two stores. Scheduling of out-of-order executions will be much more flexible as a result of being able to run more memory operations per cycle.

Perhaps one of the greatest changes comes as a result of memory addressing. Both AMD and Intel switched to 64-bit addresses back in 2003, but only 48 bits have been used. Now, more significant bits are being put to work. Up to 4PB of memory is possible via 52-bit addresses. Virtual memory addresses expand up to 57 usable bits, allowing for 128PB. For high end consumer CPUs, this is completely unusable, but enterprise users with deep pockets will certainly find a way to take advantage of the new capabilities.

Consumer CPUs carrying the Intel Core branding using Sunny Cove architecture are expected to become available during the second half of 2019. On the enterprise front, Intel has not directly given a clear answer. Intel will be offering Cascade Lake in the first half of 2019 followed up by Cooper Lake, but Sunny Cove may not be realized until it carries Ice Lake nomenclature for Xeon processors.

Permalink to story.

https://www.techspot.com/news/77828-meet-sunny-cove-intel-architecture-finally-improve-ipc.html

 

[Puiu]

So... should I assume that it will become available to us mortals in early 2020? I'm assuming that they want to launch at least a few months earlier than Ryzen 4000.

 

[wiyosaya]

The bit about PB of memory sounds to me as if it is marketing I'm not saying that the added address space will not be there, but that I would be extremely surprised if there would be any MBs out that that would support anywhere near that amount of physical memory.

Virtual memory, that's another story. I work on a FEA package, and I note that in Windows 10, the software is successfully able to off-load requests for memory beyond what is physically in the PC to virtual memory - which is a vast improvement in memory management. However, I don't see putting 128 PB of disk space into anything other than a HPC environment. Most applications will fall into that "mere mortals" category, and sIntel parading this as a feature of this generation of processors is, to me, like I said above - pure marketing

 

[RebelFlag]

But will it run Crysis???

 

[ShagnWagn]

CPUs still seem to be increasing in processing power. Isn't Moore's Law still working? I'm not referring to core processing, but the entire CPU. Each core may not be increasing in speed any more, but if you can fit 128 cores on a CPU chip and more, then I still consider it increasing.

 

[FliGuyRyan]

But will it run Crysis???

Classic.

It just never gets old.

 

[Vulcanproject]

CPUs still seem to be increasing in processing power. Isn't Moore's Law still working? I'm not referring to core processing, but the entire CPU. Each core may not be increasing in speed any more, but if you can fit 128 cores on a CPU chip and more, then I still consider it increasing.

Moore's law referred to transistors doubling on a monolithic chip every two years. It's been more like two and a half since 22nm, and 3 years since 14nm. It'll only get ever slower beyond this point, unless some unforeseen breakthrough happens.

In the strictest sense, chiplets (daisy chaining individual dies together) aren't really doubling transistors on the chip. They are just packaging up more chips together.

Whether or not you count the future with chiplets as part of Moore's law I guess is up to you. Moore's law as such is only a very loose observation that the industry has since strived to remain in step with.

 

[Dimitrios]

Just ex AMD workers doing their magic since INTEL hired them almost a year ago and also Jim Keller. Yes I understand Jim Keller gets contracts from many tech companies then leaves when his job is complete.

 

[Nocturne]

The bit about PB of memory sounds to me as if it is marketing I'm not saying that the added address space will not be there, but that I would be extremely surprised if there would be any MBs out that that would support anywhere near that amount of physical memory.

Virtual memory, that's another story. I work on a FEA package, and I note that in Windows 10, the software is successfully able to off-load requests for memory beyond what is physically in the PC to virtual memory - which is a vast improvement in memory management. However, I don't see putting 128 PB of disk space into anything other than a HPC environment. Most applications will fall into that "mere mortals" category, and sIntel parading this as a feature of this generation of processors is, to me, like I said above - pure marketing

This is merely for Enterprise and upper workstation computing, the Z series where 75% of the consumer market lies will only support a base amount of RAM as the mobos will only have 2-4dimm slots regardless due to constraint costs this is probably for forward product support as Samsung had announced rediculous mega density sticks, and of course maybe optane dimms.

Either way I doubt the regular consumer is getting anything more than a marginal ipc boost, thanks to Intels pay wall taxes on PCIe Lanes and such. Either way honestly not returning my streaming gaming rig back to Intel until they decide to be competitive and do something that matters for consumers,

IF* AMD can truely give me a 8c16t 5.0ghz chip with near relative performance, AT A DECENT TDP, the verdict is simple they win. Features don't matter if 90% of people on the platform never see it being ideal or used to their benefit for the adopted amount for the platform, that is why X series was always a money gouge as a regular consumer product.

 

[Rees]

CPUs still seem to be increasing in processing power. Isn't Moore's Law still working? I'm not referring to core processing, but the entire CPU. Each core may not be increasing in speed any more, but if you can fit 128 cores on a CPU chip and more, then I still consider it increasing.

Moore's Law is about transistor density in an area not to be confused with cores per chip. For instance Threadripper has more cores then ryzen 7, but the area is also bigger. Moore's Law observes that the # of transistor density will double every 2 years. With rumors of Zen 2 and Sunny Cove, Moore's law could be still intact.

 

[bexwhitt]

Sounds like spin to try and counter the amd epyc threat to Intel in enterprise.

 

[cliffordcooley]

Sounds like spin to try and counter the amd epyc threat to Intel in enterprise.

I wouldn't go so far as to call it a counter. Intel has likely been sitting on this for years, but never felt threatened enough to bring it forward. If they haven't been sitting on it, then how could they bring it to the table so quickly. Look how long it took AMD to bring Ryzen forward.

Ever since AMD took a back burner and stated they were no longer competing with Intel, I've been waiting for this day. I knew then that AMD needed time to change architectures. And when they did start competing again, I also figured it would take Intel a year or two to adjust. We are now fixing to see what Intel has been working on for the last 8 years, while waiting on AMD to return to the table.

 

[EClyde]

An article on why this matters, sure but it does not matter to me. If they want to improve the PC stop thinking of it as computer because 99% are game and writing machines. To do 95% of the computing that goes does go on this isn't needed either. But I am reminded of the fellow that wrote "A gig is not enough"

 

[meric]

I think this is why the "leaks" about zen 2 better be true

 

[wiyosaya]

This is merely for Enterprise and upper workstation computing, the Z series where 75% of the consumer market lies will only support a base amount of RAM as the mobos will only have 2-4dimm slots regardless due to constraint costs this is probably for forward product support as Samsung had announced rediculous mega density sticks, and of course maybe optane dimms.

Either way I doubt the regular consumer is getting anything more than a marginal ipc boost, thanks to Intels pay wall taxes on PCIe Lanes and such. Either way honestly not returning my streaming gaming rig back to Intel until they decide to be competitive and do something that matters for consumers,

IF* AMD can truely give me a 8c16t 5.0ghz chip with near relative performance, AT A DECENT TDP, the verdict is simple they win. Features don't matter if 90% of people on the platform never see it being ideal or used to their benefit for the adopted amount for the platform, that is why X series was always a money gouge as a regular consumer product.

We're on the same page - but a Petabyte, in other words, 1,000 Terabytes or 1,000,000 Gigabytes ? Even in the "enterprise" market, this is so far beyond current hardware that I don't see it happening any time soon. That is why I said marketing

 

[wiyosaya]

CPUs still seem to be increasing in processing power. Isn't Moore's Law still working? I'm not referring to core processing, but the entire CPU. Each core may not be increasing in speed any more, but if you can fit 128 cores on a CPU chip and more, then I still consider it increasing.

A single-atom transistor has been developed, however, it has yet to be commercialized. - https://en.wikipedia.org/wiki/Single-atom_transistor

 

[fktech]

Intel disappoints again!

 

[cliffordcooley]

Intel disappoints again!

How so?

 

[kmo911]

Nxt gen mb
128 tb to 1024 tb ssd hdd hybrid
128 pb ram to 1024 pb ram
sound card that support atmos 64 speakers
vega series pcie 3.0-5.0
nvidia gtx rtx ati rx series 3000
mb that support all pcie1.0-5.0 all gpu
pci still needed to power up a brikked pcie gpu
agp 16x arrives
matrox cards s3 cards gets 128 gb + cpu help to process movies made.
100 mp camera
144gb lan- 200gb lines 16 optical lines in one
sound card can now hanndle cpu gpu sound at 1 pb big files.

 

[fktech]

How so?

Every way possible!

 

[cliffordcooley]

Every way possible!

It is good to see you are reserving your opinion until after launch.

 

[fktech]

It is good to see you are reserving your opinion until after launch.

LOL.