Intel Skylake / Kaby Lake

[ShintaiDK]

Again, same old tune, if they made a mainstream hex core, would not have to worry about any of this.

Then people would just complain about 8 and 10 cores and how their hexcore was uncore starved.

 

[imported_ats]

Server uncore huh. Do the HEDT platforms have the server uncore on their CPUs?

Yes and have for quite a while. They are all currently effectively die chops of the server parts. Though it probably makes sense for them to switch HEDT over to chops of Xeon-D instead of Xeon-E.

 

[ShintaiDK]

Yes and have for quite a while. They are all currently effectively die chops of the server parts. Though it probably makes sense for them to switch HEDT over to chops of Xeon-D instead of Xeon-E.

Xeon-D is a limited SoC type. It would not fit the HEDT role.

 

[imported_ats]

Xeon-D is a limited SoC type. It would not fit the HEDT role.

Xeon-D could easily fit in the HEDT role. Current Xeon-D frequencies are primarily because of its power/thermal limitations, there is no reason that it couldn't be clocked significantly higher given a larger power budget. It also has ample PCIe capacity for a HEDT and if more is needed it can easily be added via a PCIe switch considering it doesn't require a southbridge.

It is also designed as a single socket solution, something the current HEDT parts are not which generally means lower latencies. From a design perspective, it would be much easier to combine HEDT with Xeon-D vs Xeon-E because Xeon-E will continously push up the performance stack on the server side adding significantly more cores which will impact the chop complexity for HEDT.

In fact I would be surprised if Intel hasn't already switched HEDT from Xeon-E to Xeon-D internally.

 

[ShintaiDK]

Remember dual channel memory too.

 

[HiroThreading]

Remember dual channel memory too.

Not to mention a lot of wasted silicon. A lot of the embedded features won't be used by enthusiasts in the HEDT segment.

 

[imported_ats]

Remember dual channel memory too.

Meh, not really going to make much difference. DDR4-3200 provides enough bandwidth.

Not to mention a lot of wasted silicon. A lot of the embedded features won't be used by enthusiasts in the HEDT segment.

what features do you considered wasted silicon in Xeon-D for HEDT?

 

[LTC8K6]

Not sure if this is new info or not?

https://www.yahoo.com/tech/intel-outs-three-processors-crystal-115449763.html

Intel has reportedly updated its price list with three new “Skylake-H” processors: the Core i7-6785R, the Core i5-6685R, and the Core i5-6585R.

...

 

[Burpo]

Intel Adds Crystal Well-based Skylake-R Processors: 65W with 128MB eDRAM

http://www.anandtech.com/show/10281/intel-adds-crystal-well-skylake-processors-65w-edram

" The chips that Intel has added to its price list are the Core i7-6785R, the Core i5-6685R, and the Core i5-6585R. The new processors are designed to fit in all-in-one PCs, small form-factor and other types of highly-integrated PCs that can satisfy the 65W TDP over the mobile Crystal Well variants that run at 45W. The new desktop chips from Intel are based on the Skylake-H silicon in its most advanced configuration: with four general-purpose cores as well as the GT4e integrated graphics. Not all the specifications of the processors are known at this point, but we are talking about quad-core processors with Generation 9 Iris Pro graphics and 72 execution units (as well as 128 MB of eDRAM), a dual-channel DDR4-2133 memory controller, a PCI Express 3.0 interface and three display outputs. The power consumption of Intel’s new embedded products for desktops does not exceed 65 W, which is typical for R-series offerings, and offer a potential upgrade path by OEMs for any equivalent systems that used an equivalent Broadwell-based R-series processor.
The new parts feature higher clock rates compared to the Broadwell processors, although slightly lower than their K series counterparts. The processors have a number of important architectural improvements which will affect the performance of these CPUs in real-world applications. It is interesting to note that Intel retained the full L3 cache size in its new R-series CPUs: in the Broadwell models part of the L3 was used for eDRAM tags, reducing their capacity by 2 MB, but the new Skylake parts are now in line with their i7 and i5 naming due to the way the eDRAM is implemented. This means that the i7-6785R has 8 MB of L3, similar to the i7-6700K, and the i5-6685R/6585R has 6 MB of L3, similar to the i5-6600K."

.." The new arrangement for the eDRAM in these Skylake processors has placed the eDRAM in a different part of the chain, between the System Agent and the DDR memory. This means that the eDRAM acts as a DRAM buffer, with 50 GBps bandwidth in each direction to the LLC, but is also accessible for early reads/writes by any device that needs memory access through the system agent (i.e. anything through PCIe). Previously this was not possible, but now it means that Skylake's eDRAM implementation should offer a speedup in many more scenarios than before.

The final noteworthy improvement of the Skylake processors compared to previous-generation offerings is revamped graphics core as well as increased amount of execution units. Based on our findings last year, real-world performance of Intel's high-end Iris Pro 6200 graphics core (Broadwell’s top iGPU) is higher than that of entry-level discrete graphics cards. Meanwhile, the highest-performing GT4e graphics core of Intel Skylake contains 72 EUs, up from 48 in the case of the Broadwell. Compute performance of Intel's contemporary top-of-the-range iGPU (Iris Pro 580) is around 1.1 TFLOPS depending on its frequency, so, this one should be tangibly faster than its predecessor. Last, but not least, Skylake’s iGPU has a revamped multimedia engine, which supports hardware decoding and encoding of UHD videos using HEVC or VP9 codecs.

Intel’s new Core i7-6785R, Core i5-6685R, and Core i5-6585R are already available at a tray price of $370, $288 and $255 respectively."

Nothing boring about these quad cores

 

[DrMrLordX]

R, but no C? There needs to be an unlocked 6785R!

 

[Head1985]

I hate intel for this..Why they didnt release 6700k with Edram?Or atleast 7700k with edram.

 

[jpiniero]

I hate intel for this..Why they didnt release 6700k with Edram?Or atleast 7700k with edram.

So they can sell Skylake-C when it comes out.

 

[DrMrLordX]

So they can sell Skylake-C when it comes out.

Then they should bloody well come out with it already.

 

[jpiniero]

Then they should bloody well come out with it already.

They have to sell Skylake-K first. But not before Kabylake-K.

 

[DrMrLordX]

They have to sell Skylake-K first. But not before Kabylake-K.

Considering how much trouble they've had with volume, you'd think they could clear out inventory quickly.

 

[frozentundra123456]

Considering how much trouble they've had with volume, you'd think they could clear out inventory quickly.

Yea, they stubbornly refuse to go all in on giving better value, but insist on segmenting the market to the nth degree to try to maximize margins.

Cue the usual suspect(s) defending this, and I understand that sentiment. But dammit, sometimes in a flat/declining market a company just needs to shake things up. The two things that come immediately to mind is bringing out a mainstream hex core and making edram much more widely available.

 

[DrMrLordX]

Maybe they're worried that they don't have anything else compelling after launching those two products. Except . . . consumer hexcores with eDRAM!

 

[frozentundra123456]

Maybe they're worried that they don't have anything else compelling after launching those two products. Except . . . consumer hexcores with eDRAM!

Maybe in about 15 years, at 5nm.

But no they will still be selling quad cores so small you have to us a magnifying glass to see them.

 

[jj109]

So they can sell Skylake-C when it comes out.

The eDRAM interface is on the long axis. This setup won't fit in LGA1151.

A Skylake-C version is going to need a bigger socket or a new GT2/GT3 die without the hilariously long GPU.

And then maybe 100 people will buy it, so it'll be $100K USD per CPU.

 

[AtenRa]

The eDRAM interface is on the long axis. This setup won't fit in LGA1151.

A Skylake-C version is going to need a bigger socket or a new GT2/GT3 die without the hilariously long GPU.

And then maybe 100 people will buy it, so it'll be $100K USD per CPU.

Hell that die is huge, add the eDRAM and selling this SKU for only $370 will have very small margins.

I really like to see how it performs in DX-12 games.

 

[IntelUser2000]

Hell that die is huge, add the eDRAM and selling this SKU for only $370 will have very small margins.

I really like to see how it performs in DX-12 games.

That chip has 72EUs, and its very similar in die size to the 48EU equipped 5775C. On the same process too. So perf/process/mm2 likely went up nearly 50%.

5775C is about 200mm2, and this chip is at ~215mm2 or so. Perhaps the difference is even less, at 10mm2. It opens up the possibility of a 96EU or perhaps 128EU GPU if they want to go all the way to 300mm2.

 

[AtenRa]

That chip has 72EUs, and its very similar in die size to the 48EU equipped 5775C. On the same process too. So perf/process/mm2 likely went up nearly 50%.

5775C is about 200mm2, and this chip is at ~215mm2 or so. Perhaps the difference is even less, at 10mm2. It opens up the possibility of a 96EU or perhaps 128EU GPU if they want to go all the way to 300mm2.

Where does it say it is only ~215mm2 ??

Edit. Broadwell-H is 168mm2. So yes another 24CUs (~41mm2) will make the die close to 210-220mm2

That is a huge die for 14nm. Almost 120-130mm2 is dedicated to the iGPU alone.

Broadwell-H die size on page 8 link bellow.
http://www.intel.com/content/www/us/en/embedded/products/broadwell-u/5th-gen-core-mobile-embedded-thermal-mech-guide.html

 

[SAAA]

That chip has 72EUs, and its very similar in die size to the 48EU equipped 5775C. On the same process too. So perf/process/mm2 likely went up nearly 50%.

5775C is about 200mm2, and this chip is at ~215mm2 or so. Perhaps the difference is even less, at 10mm2. It opens up the possibility of a 96EU or perhaps 128EU GPU if they want to go all the way to 300mm2.

Makes you wonder what kind of standalone GPU they could release if they made a 500-600mm2 die.
Over 300EUs with 8-16GB HBM and they could challenge AMD/Nvidia high end!
Price? No idea

Also I remember that old slide (3-4 years ago?) about a generation x gpu with 96EUs, I'm starting to think that's really Cannonlake: 10nm should be fine for +33%EUs and new arch, gen 10 = x in other words.
So the long term plan of IGPs went this far... nice to know.

 

[AtenRa]

Now that AMD, Intel and NVIDIA are at very similar process (14/16nm), it will be very informative to see how this 120mm2 14nm Intel iGPU will perform vs AMD and NVIDIA 110-120mm2 Polaris and Pascal.

 

[SAAA]

Now that AMD, Intel and NVIDIA are at very similar process (14/16nm), it will be very informative to see how this 120mm2 14nm Intel iGPU will perform vs AMD and NVIDIA 110-120mm2 Polaris and Pascal.

Still not fair: give them the same drivers developers and we'll talk. :sneaky:

Beside the supposed density of those processes is so different...
Yeah I know you can count transitors in many ways but GPUs have been constantly twice as dense if not more than processors on each node.

Check out Wikipedia, 45nm parts onward and you'll see what I'm talking about: there seem to be a constant advantage for GPUs density that doesn't happen on Intel's parts.

Some of their IGP have actually lower density than the CPU portion if you divide partial die size and transitors count... look at Haswell quad core GT2 and dual core GT3 to see that.

Maybe on 14nm Skylake GT4 is better but the point still stands: they are using a process that isn't optimized for density alone, plus they need an external eDRAM or they are limited to dual channel DDR4 against GDDR5.