The end of the world as we know it?
Intel unveils the details surrounding its new G-series processor that combines Kaby Lake and Radeon Vega!
A surprise to most in the industry that such a thing would really occur, AMD and Intel announced in November a partnership that would bring Radeon graphics to Intel processors in 2018. The details were minimal at the time, and only told us specifics of the business relationship: this was a product purchase and not a license, no IP was changing hands, this was considered a semi-custom design for the AMD group, Intel was handling all the integration and packaging. Though we knew that the product would use HBM2 memory, the same utilized on the RX Vega products released last year, it was possible that the “custom” part was a Polaris architecture that had been retrofitted. Also, details of the processor side of this technology was left a mystery.
Today we have our answers and our first hands-on with systems utilizing what was previously known as Kaby Lake-G and what is now officially titled the “8th Generation Intel Core Processors with Radeon RX Vega M Graphics.” I’m serious.
For what I still call Kaby Lake-G, as it easier to type and understand, it introduces a new product line that we have not seen addressed in a very long time – high performance processors with high performance integrated graphics. Even though the combined part is not a single piece of silicon but instead a multi-chip package, it serves the same purpose in the eyes of the consumer and the OEM. The marriage of Intel’s highest performance mobile processor cores, the 8th Generation H-series, and one of, if not THE fastest mobile graphics core in a reasonable thermal envelope, the Vega M, is incredibly intriguing for all kinds of reasons. Even the currently announced AMD APUs and those in the public roadmaps don’t offer a combined performance package as impressive as this. Ryzen Mobile is interesting in its own right, but Kaby Lake-G is on a different level.
From a business standpoint, KBL-G is a design meant to attack NVIDIA. The green giant has become one of the most important computing companies on the planet in the last couple of years, leaning into its graphics processor dominance and turning it into cash and mindshare in the world of machine learning and AI. More than any other company, Intel is worried about the growth and capability of NVIDIA. Though not as sexy as “machine learning”, NVIDIA has dominated the mobile graphics markets as well, offering discrete GPU solutions to pair with Intel processor notebooks. In turn, NVIDIA eats up much of the margin and profitability that these mainstream gaming and content creation machines can generate. Productization of things like Max-Q give the market reason to believe that NVIDIA is the true innovator in the space, regardless of the legitimate answer to that question. Intel see that as no bueno – it wants to remain the leader in the market completely.
But it can’t do that with its own graphics designs – it has tried for too long. Intel integrated graphics continues to fall behind both NVIDIA and AMD in this space and though it does just fine for media and general-purpose compute scenarios, for gaming and content creation, it needed help.
Kaby Lake-G will definitely make for interesting discussion as we head through 2018 and the roadmap for AMD mobile processors becomes clearer. If and when the company decides it also wants to make a 65- to 100-watt part for mobile and SFF configurations, the agreement with Intel will look curiouser and curiouser. Why would AMD enable its primary competitor with its biggest IP advantage, even if it means short term gains with a new semi-custom design win? That’s a question for another time.
What is a Kaby Lake-G processor?
So, what is Kaby Lake-G and the 8th Generation Intel Core Processors with Radeon RX Vega M Graphics? On some documentation we received it is called the G-series, and that seems fitting. This release is the first 8th gen H-series processor, with 4-cores and 8-threads, pasted onto a substrate with a custom Vega graphics processor and 4GB of HBM2 memory. There will be more H-series parts released later in 2018 that aren’t tied to the Vega graphics integration, but this is all Intel is pushing out today.
The Vega M GPU is a discrete die and connects to the processor through 8 lanes of PCI Express 3.0 electrically, like how you would see it in a desktop system or a notebook with discrete graphics. However, the PCIe connection is instead routed through a shared substrate. The GPU has as many as 24 CUs (compute units) and 1536 stream processors, though there will be a version with 20 CUs. This graphics chip enables the same features that you would find on a modern Radeon graphics card including FreeSync, async compute, video encode and decode, and the Radeon display engine.
Next to the Vega die is a 4GB stack of HBM2 with a 1024-bit memory bus. HBM2 was always touted as the lower power, lower footprint solution for graphics memory and this implementation demonstrates that perfectly. A GDDR5-based solution would be significantly larger and would require more power, minimizing the value of the integration in the first place. This memory is connected to the Vega GPU via an EMIB, embedded multi-die interconnect bridge, that enables the package to have a lower z-height. More on that below.
This processor will still include the integrated graphics supplied by the Intel H-series processor, along with support for QuickSync and its display engine. In total, this processor could support as many as 9 display outputs, though integration and desire for something like that is going to be complicated. Don’t be surprised if you find that as an edge case for a unique OEM design.
Let’s dive into the specifications of these new processors. It’s a complicated set of data though – there are five different processors and two different GPU options to look at.
|Vega M GH||Vega M GL|
|Architecture||Vega M||Vega M|
|Base Clock||1063 MHz||931 MHz|
|Boost Clock||1190 MHz||1011 MHz|
|Memory Bandwidth||204.8 GB/s||179.2 GB/s|
|Peak FP32 Perf||3.7 TFLOPS||2.6 TFLOPS|
|ROPs||64 pix/clock||32 pix/clock|
|High Bandwidth Cache||4GB HBM2||4GB HBM2|
|Architecture||Kaby Lake||Kaby Lake||Kaby Lake||Kaby Lake||Kaby Lake|
|Base Clock||3.1 GHz||3.1 GHz||3.1 GHz||3.1 GHz||2.8 GHz|
|Max Turbo Clock||4.2 GHz||4.1 GHz||4.1 GHz||4.1 GHz||3.8 GHz|
|Discrete Graphics||Vega M GH||Vega M GH||Vega M GL||Vega M GL||Vega M GL|
|Intel HD Graphics||630||630||630||630||630|
|TDP||100 watts||100 watts||65 watts||65 watts||65 watts|
The Vega graphics implementation is broken up into two options: GH and GL. Or “graphics high” and “graphics low” if you will. The GH integration is the higher performance, higher power draw option and includes the full 24 CUs and 1536 stream processors. It has a base clock speed of 1063 MHz, a boost clock of 1190 MHz, peak theoretical compute of 3.7 TFLOPS. Memory bandwidth is 204.8 GB/s. The GL integration runs with 1280 stream processors and clock speeds that are decreased by 130-180 MHz, respectively. Memory bandwidth also drops to 179.2 GB/s. Both have 4GB of HBM2 but the GH version has twice the ROP throughput at 64 pixels/clock.
The processors that integrate these graphics solutions use the “G” designation at the end of the model name to tip us off. Four of the five being announced today are Core i7 derivatives with just a single Core i5 option, though all five integrate a quad-core configuration with HyperThreading enabled. The Core i7-8809G is the highest performing option and has a base of 3.1 GHz, Turbo lock of 4.2 GHz, and 8MB of cache. It’s also the only unlocked version of the G-series, allowing OEMs and partners to integrate overclocking capabilities.
You can see from the table above that the 8709G, 8706G, 8705G and 8305G all have slightly decreasing clock rates from the 8809G, and the Core i5 lowers the cache size to 6MB. All have the same official memory speed support.
There were early rumors that this might be a socketed processor, but Intel does not have plans for that today, as all the Kaby Lake-G options being launched are BGA only.
Finally, the TDP of these processors is important to note. The Core i7-8809G and 8709G that integrate the Vega M GH solution will have 100-watt TDPs while the other three models that use the lower performing GPU solution will have 65-watt TDPs. Combining a traditional 35-watt H-series CPU with what is essentially a very well-integrated discrete GPU of 60+ watts means a lot of power will be needed for these systems. Ryzen Mobile, the current highest performance CPU+GPU option for the mobile space, has a TDP of around 30-40 watts under a full load, so you can see why I said earlier that this launch put KBL-G in a very unique class. Solutions based on the G-series will have more in common with notebooks that use the GTX 960M, GTX 1050, or even GTX 1060.
It is worth noting here that OEMs will have the ability to limit the TDP under what you see here. The Dell XPS 15 2-in-1 for example, will have a 45-watt TDP. Dell and Intel adjust for this with intelligent power technology, which we will touch on below.
How Intel made it all work
These are truly impressive specifications and Intel has done an amazing job to engineer the G-series to what we see today. The first trick was to get all this technology on a single package. Using a substrate for the GPU and CPU made sense but it was the invention of the EMIB (embedded multi-chip interconnect bridge) that allows for low z-height integration of a high speed interconnect for the HBM2 memory on package. Z-height is critical to being able to design a thin notebook or SFF PC and at just 1.7mm tall, the Intel G-series processors could find their way into very unique implementations.
Part of the custom design of this Vega GPU, according to Intel, was its request for hooks in the chip to enable Intel to intelligently manage power consumption across both the CPU and GPU on the same package. By using Intel Dynamic Tuning, at the hardware level, the processor can monitor the status of the graphics and primary processing systems on the G-series part to get information about power draw and load levels. With that info Intel can adjust the power being delivered to each part and thus control performance capabilities more granularly than you could do with a traditional CPU and discrete GPU solution. The result is that more power within that TDP limit can be given to the CPU when it needs it and more to the GPU when it needs it.
One of the net results of this technology is gaming efficiency, which Intel measured by frames per watt. With the ability to intelligently balance power and performance between the two components, rather than depending on more simplistic heuristics, OEMs can offer similar performance at a lower TDP. Intel gives an example of a 45-watt solution offering the same performance as a 62.5-watt solution without Dynamic Tuning (up to 18% better), resulting in more “frames per watt” gaming efficiency. This is what allows Dell to have the confidence it does in releasing a 45-watt TDP-down version of Kaby Lake-G.
By combining the smaller footprint of the EMIB and the dynamic power capability that Intel has created, we should see some very unique, slim form factors for gaming machines of this performance caliber.
Though we are still waiting to get our hands-on systems using Kaby Lake-G, Intel did provide some metrics to look at. The processor itself is well known and understood; it’s a KBL part with 4 cores at solid clock speeds. Intel has metrics in its deck that compare it to a three-year-old gaming system, which is reasonable, but I don’t think it something that our readers would find incredibly interesting. The graphics and gaming performance is what we are after.
Compared to a GeForce GTX 950M based system, the Vega M GL is 2.2-3.0x faster in graphics based workloads. Comparing the GL to the GTX 1050, Intel shows a few results that are 10-40% faster.
The Vega M GH results are more indicative of SFF PCs (like a new NUC) than the first couple of notebooks announced, but performance still looks impressive. Compared to a GTX 960M, it comes in 2.0-2.7x faster. And compared to the GTX 1060 Max-Q solution, 7-13% faster.
These are not earth-shattering numbers, but it puts an Intel-only solution on the table to compete against what previously required an NVIDIA chip to do in a notebook for gaming and content creation.
A big question mark for us was the software side of this. While we can wrap our heads around the idea of AMD selling Intel a semi-custom Radeon GPU to use for this project, who would manage the software? To be blunt, Intel has had a poor track record of keeping up with the gaming market requirements for software and drivers, even if it has seen improvements in the last couple of quarters. If Intel is asking gamers and OEMs to put their faith in this new G-series line, it would have to make a strong statement about its intentions.
And they have done so. Both the Intel and Radeon driver will exist on the system and both will be distributed by Intel directly. Intel gets the core of the Radeon driver stack from AMD as a part of the business arrangement, and Intel is responsible for re-branding and distribution. (I am sure there are some other minor tweaks occurring.) One of the selling points for this processor is support for all the current and upcoming features that Radeon Vega graphics can provide and in my talks with Intel leading up to this launch they have iterated over and over that they will release drivers day and date with AMD, supporting games on day one, etc. It is not an easy task and even with the majority of the optimization and work done by AMD on the driver side, will mean Intel needs to have the systems and processes in order to deliver.
There will be much skepticism from consumers; Intel needs to be sure there are no stumbles along the way.
Both Dell and HP have systems based on Kaby Lake-G showing at CES this week and both plan availability in the spring time. We’ll have separate news stories from those parties once their official launches occur, so check pcper.com for more on that.
The 8th Generation Core Processors with Radeon RX Vega M Graphics marks an interesting shift for Intel in many ways. It indicates that it sees the value both in the gaming market as well as the need to have a more powerful graphics architecture paired with its CPUs than it can reasonably develop. It also means that Intel is working with AMD to provide a solution that neither could do on its own today. (Interestingly, the letters “AMD” appear nowhere in the press briefing and instead this was always a “Radeon” product.) With the hiring of Raja Koduri and the commitment to build GPUs from the ground up in the future, it also means that Intel might being using the Radeon solution merely as a stopgap, to fill in this void until its own roadmap can catch up.
Either way, I’m exited to see what this partnership and this new processor can offer gamers. Bring on the reviews!!
them benches are quite
them benches are quite impressive for a 100w tdp.
Around 1060 performance is
Around 1060 performance is actually very impressive. I wasn’t expecting that.
I guess it will all come down to price.
Nvidia has been doing whatever it wants, especially in the mobile space, with 1060 laptops being costing more than 970 ones. It really sucks for the consumer not to have any competition.
Nevermind, I just saw the
Nevermind, I just saw the $999 MSRP for the “NUC”. Preposterous.
Sounds perfect for a cheap
Sounds perfect for a cheap Steam box
At 799$ for barebones it
At 799$ for barebones it quite expensive but if the performance is there thats a perfect HTPC/Gaming box for most homes.
Looking at the Linus Video
Looking at the Linus Video and the leaked slide deck at VideoCardz on these Kaby Lake-G SKUs. The EMIB/MCM module in the Linus Video has A lot of BGA bumps on the bottom side surrounding the CPU and GPU and it’s very intresting how the Module will access the off module memory and any off module PCIe connectivity.
Linus has had a sample module and has shown the back side so hopefully there will be more information on just what the module has available for off EMIB/MCM communication. The on module CPU connects to the Radeon Vega-M GPU via an PCIe 3.0 x8 interface and I’d like some analysis of what resource/s the GPU will have for off MCM communication to regular DDR4 System memory and of course the Vega Graphics able to use its HBCC/HBC(4GB HBM) for Virtual VRAM paging to/from DDR4 system memory and even SSD/NVM-Optane/Xpoint.
That 8x on MCM PCIe looks narrow if it will be the only method that the Vega HBCC has for any off EMIB/MCM module Virtual VRAM Paging requests made by Vega’s HBCC to and from regular DDR4 based system RAM and any HBM2 acting as High Bandwidth Cache.
The leaked slide deck on VideoCardz website does mention the mobile semi-custom Vega GPU HBCC/HBC where Vega can turn the HBM2 into a L4 like GPU/VRAM cache for a larger amount of virtual VRAM paged out to system DRAM/SSD/NVM/XPoint(Optane ?). So that’s interesting for those that may want to work with graphics workloads where Textuer/Mesh-Models resources are well in excess of 4GB.
The slide deck(Leaked) states that there will be up to 16 Render Back Ends(ROPs) available via the Vega graphics so looking at the ROP to Shader ratios will be interesting going forward and comparing that to the ROP to Shader Rations on Desktop Vega 56/64. So how will that Pixel fill rate to shader resources measure up for gaming workloads at 100 Watts for the top end SKUs that will be in the Hades Canyon NUC. The price is going to be the biggest problem at first because the system is not shipping with an OS and Memory so hopefully the pricing will come down some after a while.
The other big question for DX12/Vulkan usage is how nicely will the Vega-M discret Graphics and the Intel Kaby Lake Integrated graphics play well with each other on any explicit multi-GPU-Adaptor enabled games where both the Vega and Intel Kaby Lake Integrated graphics can be accessed for games usage. Hopefully the Integrated Graphics can be accessed for maybe gaming physics acceleration while the Radeon Vega-M GPU works the gaming graphics.
Hopefully AMD will make a
Hopefully AMD will make a similar Vega HBM part for mobile. This Intel MCM isn’t really much different from placing a separate HBM GPU and cpu on a single board. I don’t care that much about the thickness, especially if it is a 65 to 100 W part. A full GPU on a silicon interposer may be a little taller than EMIB when mounted on a board, but it doesn’t seem like it would be that big of a difference.
Although, I have thought that it would be better to just place HBM cache directly connected to a powerful APU. The Xbox X uses 40 compute units, so about twice as powerful as this GPU. It has a much weaker cpu, but probably sufficient for most things. I don’t see why AMD wouldn’t make a similar part for their own use, unless they are just making a larger APU for their use. AMD has the advantage that they could make an APU of similar power with HBM memory and not need a separate cpu. The Xbox One X APU is considerably more powerful that these parts made for Intel, except it does not use HBM. It would be great to have such a powerful APU with some HBM. With 4 or 8 GB of HBM cache, system memory would be much less important. If enough HBM is used, they could probably get away without having any external DRAM. Some PCIe NVMe might be sufficient.
I had considered HBM in general to probably be too expensive for mobile at the moment. I kind of doubt that EMIB is really that much cheaper, so these parts are going to be for high end ultra book type parts (or Apple). It is unclear exactly what AMD is going to do competition wise. It seems like they would have smaller Vega based parts coming out soon. It would be nice if HBM on silicon interposers comes down in price enough for use in midrange parts. They have Ryzen mobile APUs but these are only 8 or 10 CU, vs. 20 to 24 for this part. Integrating more than that with only dual channel DDR4 is probably pointless, so any higher end APU will need a faster memory system. The Xbox One X APU is 359 mm². That would be an expensive part if used on a full silicon interposer. So will AMD make a high end APU or just use a GPU similar to this part for intel?
I think that the first APU
I think that the first APU that AMD will make with it’s own HBM2 will be for the graphics workstation market and will probably be branded with the Epyc CPU and “WX” GPU Branding(Formally called FirePro). AMD did at one time make a Firepro branded APU for the overseas market but this new Zen/Vega/HBM2 workstation APU SKUs may make use of a separate Zen cores die, larger Vega GPU die and more HBM2 stacks all on an interposer for the workstation graphics market and portable workstation markets. So it will not be low priced like the consumer SKUs are.
Why in the world would the
Why in the world would the graphics workstations market care for integrated graphics?
mac mini and mbp has good
mac mini and mbp has good future on intel x amd partnership
mac mini and mbp has good
mac mini and mbp has good future on intel x amd partnership
is this for next gen console
is this for next gen console too?
I think an AMD Ryzen apu with
I think an AMD Ryzen apu with Vega would make a more appealing workstation? Also would Kaby Lake G fit into a Mac?
If they want to use Kaby
If they want to use Kaby Lake-G, I’m sure Apple of all companies would find a way to make it work.
Well, it seemed somewhat
Well, it seemed somewhat impressive at first, but now…
By the time there are actual shipping products, laptops with 6th-gen CPUs and GTX 1060 will be roughly a year and a half old. We know that IPC from 6th- to 8th-gen processors is the same, so the only speed increase on the CPU side will be from higher clocks.
From a GPU perspective, beating 1060 Max-Q performance means these will be about the same as the non Max-Q laptops.
From a power perspective, the 6700HQ is a 45W part (configurable down to 35W). The mobile GTX 1060 is an 80W part, while the Max-Q drops that to 60-70W. The parts in the graphs are the 100W versions of KLG. So, yes it is able to pull less power, but when I’m gaming I plug in and don’t really care.
So essentially, 16-20 months later (depending on the actual availability), Intel will be able to roughly match the performance (of non Max-Q) at roughly the same price. The only advantage seems to be lower power draw which will be effectively irrelevant on NUCs or laptop gaming (which most people with gaming laptops do plugged in).
That’s pretty sad.
Looks like the performance of 62FPS on Rise of the Tomb Raider by i7-8809G
according to notebookcheck.net/Nvidia-GeForce-GTX-1060-Desktop-Review.169419.0.html
is just slightly under the performance of a desktop GTX1060.
Will be very interesting to see how much this costs
and whether it comes to the desktop like the i7-5775C (which could run Rise of the Tomb Raider at 1080p 60+ fps with low quality – meaning that the i7-8809G … is only slightly better than i7-5775C when the only difference is low vs high quality details at the same resolution with the same fps).
i7-5775C review with Rise of
i7-5775C review with Rise of the Tomb Raider
Also a bunch of comments from
Also a bunch of comments from the left field, AMD FM2+ A10-7890K vs Intel 1150 i7-5775C:
12th February 2016, 11:41 AM
“Gamegpu has been testing games using APUs since 2015, they have a total of 31 different game benchmarks. I’ve picked the best settings above 30fps.
Comparison of A10-7870k vs i7 5775c (iris pro 6200):
Dragon’s Dogma: Dark Arisen (1080p ? low) –> 36 vs 33
Just Cause 3 (720p ? low) –> 27 vs 48
Rainbow Six Siege (1080p ? low) –> 38 vs 30
World of Tanks 9.13 (1080p ? med) –> 37 vs 31
Dirt Rally (1080p ? med) –> 35 vs 31
War Thunder 18.104.22.168 (1080p ? high) –> 38 vs 29
7870k was faster on 5/6 games.
On average 1.09 times faster.”
It’s so amazing that the
It’s so amazing that the company can continue selling the INTENDED faulty CPU chips designed basing on the specifications with the INTENDED flaw of not checking and not correctly handling the privilege levels, just amazing!