Is the GPU in Intel Kaby Lake-G More Polaris than Vega?

Subject: Graphics Cards, Processors | April 9, 2018 - 04:25 PM |
Tagged: Vega, Polaris, kaby lake-g, Intel, amd

Over the weekend, some interesting information has surfaced surrounding the new Kaby Lake-G hardware from Intel. A product that is officially called the “8th Generation Intel Core Processors with Radeon RX Vega M Graphics” is now looking like it might be more of a Polaris-based GPU than a Vega-based one. This creates an interesting marketing and technology capability discussion for the community, and both Intel and AMD, that is worth diving into.

PCWorld first posted the question this weekend, using some interesting data points as backup that Kaby Lake-G may in fact be based on Polaris. In Gordon’s story he notes that in AIDA64 the GPU is identified as “Polaris 22” while the Raven Ridge-based APUs from AMD show up as “Raven Ridge.” Obviously the device identification of a third party piece of software is a suspect credential in any situation, but the second point provided is more salient: based on the DXDiag information, the GPU on the Hades Canyon NUC powered by Kaby Lake-G does not support DirectX 12.1.

View Full Size

Image source: PCWorld

AMD clearly stated in its launch of the Vega architecture last year that the new GPUs supported DX 12.1, among other features. The fact that the KBL-G part does NOT include support for it is compelling evidence that the GPU might be more similar to Polaris than Vega.

Tom’s Hardware did some more digging that was posted this morning, using a SiSoft Sandra test that can measure performance of FP16 math and FP32. For both the Radeon RX Vega 64 and 56 discrete graphics cards, running the test with FP16 math results in a score that is 65% faster than the FP32 results. With a Polaris-based graphics card, an RX 470, the scores between FP32 and FP16 were identical as the architecture can support FP16 math functions but doesn’t accelerate it with AMD’s “rapid packed math” feature (that was a part of the Vega launch).

View Full Size

Image source: Tom's Hardware

And you guessed it, the Kaby Lake-G part only runs essentially even in the FP16 mode. (Also note that AMD’s Raven Ridge APU that integrated Vega graphics does get accelerated by 61% using FP16.)

What Kaby Lake-G does have that leans toward Vega is support for HBM2 memory (which none of the Polaris cards have) and “high bandwidth memory cache controller and enhanced compute units with additional ROPs” according to the statement from Intel given to Tom’s Hardware.

It should be noted that just because the benchmarks and games that can support rapid packed math don’t take advantage of that capability today, does not mean they won’t have the capability to do so after a driver or firmware update. That being said, if that’s the plan, and even if it’s not, Intel should come out and tell the consumers and media.

The debate and accusations of conspiracy are running rampant again today with this news. Is Intel trying to pull one over on us by telling the community that this is a Vega-based product when it is in fact based on Polaris? Why would AMD allow and promote the Vega branding with a part that it knows didn’t meet the standards it created to be called a Vega architecture solution?

Another interesting thought comes when analyzing this debate with the Ryzen 7 2400G and Ryzen 5 2200G products, both of which claim to use Vega GPUs as a portion of the APU. However, without support for HBM2 or the high-bandwidth cache controller, does that somehow shortchange the branding for it? Or are the memory features of the GPU considered secondary to its design?

This is the very reason why companies hate labels, hate specifications, and hate having all of this tracked by a competent and technical media. Basically every company in the tech industry is guilty of this practice: Intel has 2-3 architectures running as “8th Generation” in the market, AMD is selling RX 500 cards that were once RX 400 cards, and NVIDIA has changed performance capabilities of the MX 150 at least once or twice.

The nature of semi-custom chips designs is that they are custom. Are the GPUs used in the PS4 and Xbox One or Xbox One X called Polaris, Vega, or something else? It would be safer for AMD and its partners to give each new product its own name, its own brand—but then the enthusiasts would want to know what it was most like, and how did it compare to Polaris, or Vega, etc.? It’s also possible that AMD was only willing to sell this product to Intel if it included some of these feature restrictions. In complicated negotiations like this one surely was, anything is feasible.

These are tough choices for companies to make. AMD loves having the Vega branding in more products as it gives weight to the development cost and time it spent on the design. Having Vega associated with more high-end consumer products, including those sold by Intel, give them leverage for other products down the road. From Intel’s vantage point using the Vega brand makes it looks like it has the very latest technology in its new processor and it can benefit from any cross-promotion that occurs around the Vega brand from AMD or its partners.

Unfortunately, it means that the devil is in the details, and the details are something that no one appears to be willing to share. Does it change the performance we saw in our recent Hades Canyon NUC review or our perspective on it as a product? It does not. But as features like Rapid Packed Math or the new geometry shader accelerate in adoption, the capability for Kaby Lake-G to utilize them is going to be scrutinized more heavily.

Source: Various

Video News

April 9, 2018 | 04:34 PM - Posted by Joseph Taylor (not verified)

The most puzzling thing about this though is that RPM is a feature that needs to have a critical mass of GPU's in the wild before it will be widely utilized so withholding that feature on any product going forward seems hard to justify.

April 9, 2018 | 05:55 PM - Posted by BrandingThreeCardMonteObfuscations (not verified)

Can someone get Intel/AMD to get some whitepaper quality information on this SKU? I'd like to Know if the HBCC/HBC(HBM2 Used as Cache) IP from Vega actually wotks for these Intel Kaby Lake-G SKUs.

I'd like to Look at maybe getting one(That Intel NUC) for Blender Rendering, gaming also, but I want that Vega Like HBCC/HBC(HBM2) IP to be working for very large Texture/Mesh Model sizes that can easily top 4GB or even 8GB in size on Blender for super high resoultion textures and High Polygon count mesh models based scenes.

Both Intel and AMD better realize that some Folks need the proper data in order to make informed consumer decisions and maybe the FTC needs to be looking at these sorts of branding obfuscations as prime targets for some enforcment.

I think that proper and full computing/processor data sheets should be required by Law and that Intel, AMD and others need to be more forthcoming with that necessary information or risk violations of consumer Laws!

April 10, 2018 | 12:10 PM - Posted by Randy (not verified)

Government agencies doing their jobs to regulate/enforce to protect consumers under this administration? Your joking right? Pretty much every person leading those types of agencies are on record previously saying they should be shut down.....

April 10, 2018 | 01:39 PM - Posted by ThatGoesForAnyOfThoseElectedCashDrops (not verified)

Consumer protection is as much a State Government responsibility as a federal government responsibility so it's The FTC for federal and whatever state agencies do the same.

There are different adminstrations in each state but looking at some states in the northeast US that you would think should be doing more to protect consumers they have some of the weakest consumer protection laws on the books.

And that: "this administration" maybe even less hope there, but even under the Opposing Party's adminstration the Big Monopoly interests still had the same influnce with the high technology campanies being as bad for fair and equal competition than any of the Abusive Monopoly Trusts like the Standard Oil trust of the late 19th early 20th century.

That Citizens United POTUS decision gave all the political power to the corporations be they Google, Amazon, Intel, Nvidia or any Big Oil and Coal interests that are likewise abusive monopoly interests!

One of the supposed Bluest of the Blue states In the NE US does not even have laws on the books requiring any Businsses to give Reciepts for any business transactions, and that includes Landlords not legally required to give reciepts for payment of monthly rent! So how is that for the little citizen without enough cash to matter to any politions!

You would be shocked at what's not codified in most state laws on the books that even explicitelly target consumer protection. I mean most businesses will give reciepts but in some states it's not legally required.

Landlords in the norteast have loads of cash and the political power that can be purchased with that cash, be they red or blue! And there's more than a few red and blue politicians only concerned with that color of money!

April 10, 2018 | 02:16 PM - Posted by ThatGoesForAnyOfThoseElectedCashDrops (not verified)


they do rhyme though and not in good ways sometimes!

April 14, 2018 | 01:58 PM - Posted by Anonyninymous (not verified)


For once, you actually had something insightful to offer. It was still a semi-wall-of-text in broken English (or just bad), but it was spot on about the myth of "consumer protection" as it pertains to the two largest political parties in the U.S.

And it also appears you're starting to get the hang of paragraphs! Huzzah!

April 9, 2018 | 09:13 PM - Posted by Anonymouse (not verified)

Isn't it possible that the driver intel uses doesn't have the RPM feature enabled (yet)?

April 10, 2018 | 12:45 AM - Posted by Anony mouse (not verified)

Ryan Shrout your getting sloppy AF. Read your own sources

We reached out to Intel with our findings. The company’s response echoed what we were told at CES in January, when we first looked into this matter. “This is a custom Radeon graphics solution built for Intel,” the statement began. “It is similar to the desktop Radeon RX Vega solution with a high bandwidth memory cache controller and enhanced compute units with additional ROPs (Render Output Units).”

April 10, 2018 | 09:30 AM - Posted by AMDsLackOfROPsIsWhatLostAMDTheGamingGPUMarket (not verified)

This is proof positive that AMD can in fact engineer a GPU product that is rather ROP heavy in design.

So Why can't AMD take the Vega 56(Binned Vega 10 base die tapeout based) SKU as a starting point that has the exact same numbers of Shader cores and TMUs as the GTX 1080Ti. And it's only that the GTX1080Ti has more ROP's(88 ROPs) compared to Vega 56's 64 ROPs(The maximum number of ROPs available from that Vega 10 base die tapout allows).

AMD is perfictly capable of doing a New Gaming Focused Vega base die Tapeout that's similar to Nvidia's GP102 that lacks the total shader cores compared to Vega 10 but that GP102 tapout offers up to 96 ROPs for Nvidia to create GPUs that have way higher pixel fill rates than Vega 10 with its maximum available allotment of ROPs of only 64.

All AMD would have to do is tapeout a Vega 56 new tapeout like design and Keep the numbers of Shaders and TMUs but increase the maximum numbers of ROPs to 96 and that would be AMD's more gaming oriented base die Tapeout for AMD to at least compete against any GP102 based designs/bins.

Folks ROPs are what gives that Pixel fill rate(Gpixels/s) that is directly related to the FPS metric that everybody uses do judge a GPUs gaming performance(Justly or Not).

I sure hope that AMD's Epyc Sales pick up over the next year or two and that Navi is in fact made up of Modular/Scalable GPU Die Chiplets. And this is so AMD can with Navi quickly create a full lineup of GPUs low end to mid range and Flagship in a single release year instead of AMD having to do Mainstream/below one year and Flagship the next. It's apparent for all see how well that Zen/Zeppelin modular/scalable die design worked out for AMD with its CPU market offerings and the ability to create Threadripper almost as an after thought because of that Zen/Zeppelin Modular/Scalable die design/tapeout and that Infinity Fabric IP that's also scalable/modular in nature.

Nvidia is not beating AMD in gaming with Better GPU technology Nvidia is beating AMD with Money, enough money to complete 5 different base die tapeouts every new generation(GP100, GP102, GP104, GP106, G108) compared to AMD's Vega 10(Only One Discrete GPU Base Die Tapeout).

GP102 has still has excess ROPs above what is used(Fused) to give the GTX 1080Ti its 88 ROPs and commanding pixel fill rates.

AMD should work to get the loscks a little higher also but any lower clock rates can be compensated for by adding extra ROPs compared to the competition.

I think that Raja left AMD because AMD's management would not let him compete with Nvidia in ROP counts and that AMD is a more compute CPU/GPU oriented company. And that's not a bad idea for AMD if you look at the Markups for CPUs and GPU accelerator Products in the Server/HPC/Workstation markets and that new AI/GPU Accelerator market. Vega 10 gave up ROPs for shaders and AMD got lucky this time around with the miners loving Vega 10's shader/compute heavy designs.

Vega is a great GPU micro-arch and a GPU's tapeout(TMU, ROP numbers) especially ROP's have more to do with a GPUs gaming performance than having an excess number of shader cores for current and past GPUs.

Now for Future GPUs and gaming with that New DXR(Ray Tracing) API from Microsoft and AMD's and Khronos' Vulkan equlivent for Ray Tracing acceleration on the GPUs shader cores, there really will be a need and good use of any excess shader cores on gaming GPUs. Volta(GV100) sure has a high number of shader cores compared to Nvidia's previous generation of GPU offerings. AMD's got Vega 20 and This time around AMD has not yet produced any dual Vega GPUs on a single PCIe card offerings and Navi is a wildcard at the moment, that AMD is much too closed mouthed about.

April 10, 2018 | 09:34 AM - Posted by AMDsLackOfROPsIsWhatLostAMDTheGamingGPUMarket (not verified)

Edit: loscks a little
To: clocks a Little

May 2, 2018 | 04:31 AM - Posted by Photonboy


You are only justifying the reason for the Article. What you have in BOLD basically just says it's similar to the normal Vega architecture.

The point is that there are DIFFERENCES, or did you even read the article?

April 10, 2018 | 02:56 AM - Posted by WhyMe (not verified)

IIRC wasn't there talk of Kaby Lake-G using Polaris when it was announced back in November?

Did some digging and it was you guys who said...

The GPU looks to be based on the Polaris architecture which is a slight step back from AMD’s cutting edge Vega architecture. Polaris does not implement the Infinity Fabric component that Vega does.

April 10, 2018 | 03:19 AM - Posted by Anonymouse2 (not verified)

board names: 694C:C0 gfx804 = Polaris
old news..

Selling Polaris instead of VEGA to Intel makes much more sense. It's the IP-stack they are experienced in selling to OEMs, which this Intel deal clearly is.

Polaris, current North star, and VEGA, future North star, clearly suggest an old and a new route for AMD GPU. Since both IP-stacks run with IF, HBCC and are ported to newest nodes, we probably see some more interesting Polaris deals coming.

Since Polaris also uses HBCC, it's easy to see the next update to Polaris: GDDR6. Apple and Intel will surely be interested. So much power (G4560 + 1050 ti) in such 50% of the footprint, should interest others as well. Can't wait on those AMD Polaris Chromebooks.

April 10, 2018 | 03:28 AM - Posted by Anonymouse2 (not verified)

Main differences between the IP-stacks are their media blocks and NextCU. Those darn media codec patents.

However, some important patents ended and amd now has 2 GPU IP stacks. It's waiting for Coreboot support..

April 10, 2018 | 05:34 AM - Posted by DIMITRI JOSEPH

I could be wrong about this.

But from what I remember AMD was gonna sell Polaris to Intel for them to use as integrated graphics.

But AMD were gonna use Vega integrated graphics for themselves.

April 10, 2018 | 05:38 AM - Posted by Prodeous13 (not verified)

So basically they took VEGA, ripped out things that are not yet massively supported like packed math, and probably other features, and done.

As long as the GCN core is from VEGA and not Polaris, then fine. And if it outperforms the previous Polaris, the in the end that is what matters.

April 10, 2018 | 09:22 AM - Posted by Anonymouse2 (not verified)

They took Polaris and added HBCC. Then Intel uses EMIB to tie everything together.

April 10, 2018 | 09:53 AM - Posted by AMDsLackOfROPsIsWhatLostAMDTheGamingGPUMarket (not verified)

It would sure be nice if the Rumored RX 580X/other refresh SKUs had that HBCC/HBC IP and could make use of any GDDR5 as a last level VRAM Cache in a similar manner to how Vega Makes use of HBM2 sa the HBC(High Bandwidth Cache). GDDR5 is High Bandwidth also realative to regular system DDR4 DRAM.

GPUs are great at hiding any memory's latency what wih all those GPU threads available(thousands of shader cores/threads) and the more cache the better for bandwidth and latency hiding improvments.

May 2, 2018 | 04:40 AM - Posted by Photonboy

What about people who would actually USE the stuff that's missing or at least plan to in the near future?

Where do you draw the line in naming if you change things?

What if I bought an i7-8000U mobile CPU then found out the AVX processing was 4X slower than a desktop i7-8000 series at the same frequency?

FP16 is a big function of VEGA whether many people use it or not, but for those people a 65% difference is huge.

It's also somewhat of a premium product. If it was a cheap-ass $300 laptop I'd be less annoyed.

April 10, 2018 | 12:46 PM - Posted by psuedonymous

The other semi-custom GPUs from AMD (those used in the most recent generation of consoles from Sony and Microsoft) have both been 'Polaris, with some bits that later were folded into Vega added in' (as they preceded Vega rather than followed it). The most unique part of the Kaby Lake-G part would be the HBM interface, as that would be an EMIB-specific interface rather than the interposer-based one used for 'big Vega'.

April 10, 2018 | 02:05 PM - Posted by ThatGoesForAnyOfThoseElectedCashDrops (not verified)

Embedded Multi-Die Interconnect Bridge (EMIB) IP still makes use of an Small silicon Interposer embedded down in the EMIB's organic substrate.

The Vega GPU to HBM2 stack interface is still silicon based on Intel's EMIB/MCM module. There is a smaller silicon interposer Interconnect Bridge that only has traces for the Vega GPU to HBM2 stack (1024 bit) connection and the Intel SOC on the other end of the EMIB/MCM is still wired to the Vega GPU by only a less dense x8 PCIe 3.0 connection that's routed through the EMIB's organic(Non Silicon) Substrate the same way its done on any PCB based module.

Don't get wrong that EMIB is a nice solution but AMD is going to be at some poit in time making use of active interposers with the whole interface fabric, traces and logic circuitry/transistors and all, etched into the Interposer's silicon substrate and you can not etch that using any organic substrate type of materials the way that it can be done with silicon processes.

There will be whole Networks on an Silicon Interposer Chip of the active Silicon interposer design where processors begin to become 3D stacked rather than just 2.5D stacked like they are currently.

April 10, 2018 | 03:12 PM - Posted by Anonymouse2 (not verified)

I think the HBM is exclusively connected with IF to the GPU and mem+gpu are packaged with MCM. Then comes EMIB to connect to Intel CPU.

April 11, 2018 | 02:04 AM - Posted by EducateUSelfAndReadingIsFundamentalFoThat (not verified)

Nope, Nope, Nope-A-Dope! Go and read the HBM/HBM2 JEDEC Standards and the Infinity Fabric is a bunch of different Protocols and is actually has seperate Data and Control Fabric PHYs.

The Infinity Fabric Protocol can have Other Protocol traffic support also via packet encapsulation sorts of IP and it's the GPUs memory controller that speaks to the JEDEC standard HBM/HBM2 stacks via 8, 128 bit independently operating channels per HBM/HBM2 stack(1024 bits wide per stack). The newer HBM2 standard also supports a 64 bit pseudo mode where each of those 8, 128 bit channels can be split into 2, 64 bit pseudo channels for finer grained memory access.

Go and Read Charlie's of ant S/A article about AMD's Infinity Fabric(1).

Also wikichip's Infinity Fabric entry is very informative(2).


"AMD Infinity Fabric underpins everything they will make"


"Infinity Fabric (IF) - AMD"

April 10, 2018 | 03:38 PM - Posted by Anonymouse (not verified)

Guys nobody wants this because after all its subpar. /s

April 10, 2018 | 09:19 PM - Posted by Anony mouse (not verified)

Well in fairness. They didn't blog about how transparent and beneficial it is to the consumer.

Thus it warrants Ryan scorn.

April 11, 2018 | 02:17 AM - Posted by EducateUSelfAndReadingIsFundamentalFoThat (not verified)

Not with all those Extra ROPs that AMD baked into that semi-custom design. Let's get Shrout to shout out to Intel and get the ROP counts and the Pixel Fill rates on those semi-custom AMD Vega die variants/bins that Intel is using on its G series SKUs.

I for one think that the Press needs to tell the GPU makers that if there is no Shader/TMU/ROP counts along with GTexel and GPixel fill rates then the product can not be reported on. That Shader Flops metric is not enough and there needs to be complete specification sheets provided on these processors and that needs to be reguired by LAW in order for cosumers to make informed decisions!

Proper specifications is a must, and that includes making with the whitepapers also, do you hear me AMD!

ROPs are what give the Better FPS metrics no matter the quality of the frames flung out there!

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

This question is for testing whether you are a human visitor and to prevent automated spam submissions.