Intel launches five Kaby Lake G CPUs with Radeon RX Vega graphics built-in

[William Gayde]

Traditionally there's been a sizable performance gap between systems using integrated graphics compared to systems with a dedicated GPU, especially on Intel's Core CPU lineup which has been dominating desktops and laptops for a long time. With Intel's newest launch, built-in graphics performance is about to get a lot faster and they've partnered with rival AMD to do it.

The new Kaby Lake G processors are built on the 14nm process and essentially carry three separate chip dies on a single package. Each processor will have an 8th-gen CPU, a custom RX Vega M GPU, and 4GB of HBM2 memory. The three components are directly linked together to deliver improved performance and all models will come with 20 PCIe lanes. Of these, four will be used to connect the CPU to the chipset, eight will be dedicated to linking the CPU and the GPU, and the remaining 8 are available to the user for additional peripherals.

The new EMIB interconnect allows for a reduced footprint. This helps keep the chips both smaller and thinner. Rather than using older GDDR5 memory, Kaby Lake G will use HBM2 for increased power efficiency and throughput.

Moving over to GPU specifications, the Vega chip will come with up 24 compute units and up to 16 render back end units. This allows for calculations of up to 64 pixels per clock cycle. The CPU, GPU, and HBM can all be overclocked independently and can share power with Intel's "Dynamic Tuning" capability. Intel lists support for 9 display output as well as 6 displays, so we'll have to wait for clarification on that regard.

The new CPUs will be available on both i5 and i7 lines and all will be 4 core, 8 thread models. The GL variant is the lower tier and is aimed at light gaming and content creation. It will have 20 compute units, up to 8MB of cache, a memory bandwidth of 179GB/s, and a total TDP of 65 watts.

The more powerful GH line is centered on gaming and will feature 24 compute units 8MB of cache, a memory bandwidth of 204GB/s, and come in at a higher TDP of 100 watts.

For performance numbers, the GL line will generally come in slightly faster than a GTX 1050 and the GH line will be just slightly above a GTX 1060 which is damn impressive.

There are five planned SKUs for this initial launch as shown below. In addition to the Vega chip, the CPUs will still feature Intel's own HD Graphics 630.

This launch is based on the older Kaby Lake architecture rather than Coffee Lake, but it still looks to offer tremendous value. While they don't replace a separate, dedicated GPU, there will definitely be a market for these chips if the price point is right.

On paper these appear to be perfect for light gaming and would perform very well in 1080p eSports titles. From the preliminary benchmark numbers Intel has provided, it appears that they offer performance comparable to a $300-500 CPU and GPU combo depending on the configuration. If they can deliver Kaby Lake G at a discount over a similar traditional setup, it looks like they'll have a real winner here.

Permalink to story.

https://www.techspot.com/news/72628-intel-launches-kaby-lake-g-processors-radeon-rx.html

 

[Bao Nguyen]

Quite disappointed that H-series doesn't offer 6 core, should I wait for Coffeelake?

 

[cliffordcooley]

Impressive! I'd also like to know if these CPUs are shipped with immunity to Meltdown and Spectre. After all Intel has known about the threat for awhile and could have engineered a fix for them.

 

[Puiu]

Impressive! I'd also like to know if these CPUs are shipped with immunity to Meltdown and Spectre. After all Intel has known about the threat for awhile and could have engineered a fix for them.

The usual microcode update and windows patch. I don't this having any significang chances to the Kaby Lake architecture.

 

[cliffordcooley]

You're probably right.

 

[ZackL04]

Is it April 1st? An APU with not only 1050 performance but....1060??? My god, and I just bought a wii U to simplifly my emulation needs and add portability....looks like I’ll be building one of these bad boys in the near future.

Very exciting stuff

Edit: wait...are these BGA?

 

[Cryio]

Quite disappointed that H-series doesn't offer 6 core, should I wait for Coffeelake?

Kaby Lake G/R *is Coffe Lake*.

Intel hasn't officially announced 45W CPUs for Coffe Lake on laptops yet.

 

[Vulcanproject]

Superb news. For years I have been waiting for notebooks with decent integrated graphics. You either give up a lot of size and weight for a discrete GPU, battery performance, or simply a lot more cost.

Yet now in 2018 we'll not only see good Ryzen mobile parts with decent Vega graphics, we have good Intel parts with better graphics that might actually play games well!

Even the 20 Compute Unit model with 2.6 TFLOPS on the i5 should be able to throw up minimum PS4 level visuals and resolutions.

Radeon RX560, but with more memory bandwidth.

 

[Amet Monegro]

Superb news. For years I have been waiting for notebooks with decent integrated graphics. You either give up a lot of size and weight for a discrete GPU, battery performance, or simply a lot more cost.

Yet now in 2018 we'll not only see good Ryzen mobile parts with decent Vega graphics, we have good Intel parts with better graphics that might actually play games well!

Even the 20 Compute Unit model with 2.6 TFLOPS on the i5 should be able to throw up minimum PS4 level visuals and resolutions.

Radeon RX560, but with more memory bandwidth.

Not is RX560, 24 CU = 768 SP, RX560 is 896/1024 SP diferent ARKs

 

[psycros]

I wonder if we'll see Intel/Ryzen in smart phones and tablets eventually? In any case, this crazy new partnership will do more to advance developer support for Radeon GPUs than anything AMD has tried.

 

[CaptainTom]

If this 100w APU really performs around a desktop i5 + GTX 1060 it will be an insane product. They could justifiably charge $400 for it on desktop, and frankly it makes every other laptop combination obsolete unless you want one of those hilarious 20LB "desktop replacements."

Remember you can configure this down to a 65w TDP, and that would be able to fit in the slimmest of gaming notebooks and even some of the high-end ultrabooks. That 13" macbook pro with a 10 hour battery life now could game as well as a High-End desktop!

 

[CaptainTom]

Superb news. For years I have been waiting for notebooks with decent integrated graphics. You either give up a lot of size and weight for a discrete GPU, battery performance, or simply a lot more cost.

Yet now in 2018 we'll not only see good Ryzen mobile parts with decent Vega graphics, we have good Intel parts with better graphics that might actually play games well!

Even the 20 Compute Unit model with 2.6 TFLOPS on the i5 should be able to throw up minimum PS4 level visuals and resolutions.

Radeon RX560, but with more memory bandwidth.

Not is RX560, 24 CU = 768 SP, RX560 is 896/1024 SP diferent ARKs

24 x 64 = 1536SP's buddy. This is also Vega, not Polaris (It has an IPC advantage). It should perform as well as a stock RX 470 or a 1060.

 

[Vulcanproject]

Not is RX560, 24 CU = 768 SP, RX560 is 896/1024 SP diferent ARKs

Please read my post if you are going to reply to me. I specifically stated the 20CUs 2.6TFLOPS model and you even quoted me on it. 'Radeon RX Vega M GL' on the final chart. All the specs are right there on the article's final slide. Yet you're talking about the 24CU model and saying I'm wrong? Huh?

20CUs, 1280 streams, 32 ROPS, 1011MHz boost clock, 2.6TFLOPS. The core performance is then roughly equivalent to an RX560 (the original 1275MHz model):

16CUs, 1024 streams, 16 ROPS, 1275MHz boost clocks, 2.6TFLOPS......

RX560 has less CUs and ROPS, but 26 percent higher boost clocks. The end result is their peak theoretical compute is near identical @ 2.6TFLOPS. The advantage this integrated 20CU model has is the extra performance from ROPS (twice as many, but at slower clocks) and with HBM2 it should have more memory bandwidth than any RX560.

Performance of that model will probably be like an overclocked RX560 or close to a GTX1050ti at best, but definitely not remotely as fast as an RX470/570. Those are 5 TFLOP cards, 2048 streams >1200MHz boost clocks.

The 24CU model @ 3.7 TFLOPS will probably also fall a bit short of an RX470/570. It's not going to be near a GTX1060. You can overclock it though it seems, at least the top model.

 

[tipstir]

So this looks like the old Intel PII way plug in type with SoC/APU/GPU onto the PBC or is this interchangeable technology. Or is this mere just die cast technology like we have right now. To improve on speed bus and graphics would be nice tape into true pipeline with unlimited band width. So what do we have here still are they getting away from 32-bit/64-bit or 256/256 bit or 512/512 or 1024/1024-bit or higher. Where are we at time in this spectrum. In all specs impressive review just hope all what you have said actually gets release to the mainstream "US"!

 

[CaptainTom]

The 24CU model @ 3.7 TFLOPS will probably also fall a bit short of an RX470/570. It's not going to be near a GTX1060. You can overclock it though it seems, at least the top model.

Ok I delete the first 80% of your post since you were talking about the models I was clearly not talking about.

It will be around a 1060 pretty easily if it can keep those 1200MHz clocks, and that will be easy on the 100w model. In fact check their own benchmarks - they show it winning in multiple games over the (notebook) 1060.

But I am not sure why this is odd to you - it has more TFLOPS , more bandwidth, and significantly less latency communicating with its cpu than the 1060 does. On average it might end up being a tad weaker, but in general it should perform within 10-20% of a 1060. I would say that's close enough for a bloody APU!

 

[Vulcanproject]

Ok I delete the first 80% of your post since you were talking about the models I was clearly not talking about.

It will be around a 1060 pretty easily if it can keep those 1200MHz clocks, and that will be easy on the 100w model. In fact check their own benchmarks - they show it winning in multiple games over the (notebook) 1060.

But I am not sure why this is odd to you - it has more TFLOPS , more bandwidth, and significantly less latency communicating with its cpu than the 1060 does. On average it might end up being a tad weaker, but in general it should perform within 10-20% of a 1060. I would say that's close enough for a bloody APU!

I compared them to desktop parts for performance. I assumed you did too, since you stated 'RX 470'.

There is no mobile RX 470, just the GPU with the same performance as the desktop GPU. There is also no mobile RX560 as I first mentioned. There is no mobile RX570. These desktop tier parts have been crammed into a scant few alienware notebooks, but they aren't special mobile parts, Alienware said to themselves.

There IS an R9 470X mobile GPU but we both did state the RX prefix denoting the desktop variant. A 24CU part here is definitely not as fast then as an RX 470, as I stated.

Intel compared this mobile part to a thin and light Max Q version of the mobile GTX1060, itself slower than the full fat mobile part, itself slower than the desktop GTX1060.

In theory then sure, it might just be as fast as the very slowest mobile GTX1060 with the name, but it won't be nearly as fast as the desktop GTX1060, nor as fast as the RX470.

 

[CaptainTom]

I compared them to desktop parts for performance. I assumed you did too, since you stated 'RX 470'.

There is no mobile RX 470, just the GPU with the same performance as the desktop GPU. There is also no mobile RX560 as I first mentioned. There is no mobile RX570. These desktop tier parts have been crammed into a scant few alienware notebooks, but they aren't special mobile parts, Alienware said to themselves.

There IS an R9 470X mobile GPU but we both did state the RX prefix denoting the desktop variant. A 24CU part here is definitely not as fast then as an RX 470, as I stated.

Intel compared this mobile part to a thin and light Max Q version of the mobile GTX1060, itself slower than the full fat mobile part, itself slower than the desktop GTX1060.

In theory then sure, it might just be as fast as the very slowest mobile GTX1060 with the name, but it won't be nearly as fast as the desktop GTX1060, nor as fast as the RX470.

My point was that while the 1060 for laptops is the same as the desktop version, it generally can't boost quite as high in most of the slim platforms it is usually sold in. As such, a 1536-SP Vega should be able to roughly match it.

I did say "roughly" and "within 10-20%", so I am unsure why you are becoming obsessed with the semantics. For all tense and purposes this APU will provide the exact same gaming experience as the 1060 while using half the energy. That's impressive.

 

[Vulcanproject]

My point was that while the 1060 for laptops is the same as the desktop version, it generally can't boost quite as high in most of the slim platforms it is usually sold in. As such, a 1536-SP Vega should be able to roughly match it.

I did say "roughly" and "within 10-20%", so I am unsure why you are becoming obsessed with the semantics. For all tense and purposes this APU will provide the exact same gaming experience as the 1060 while using half the energy. That's impressive.

It's not semantics. It's just you also compared it to a desktop part, the RX 470.

I merely don't believe that a 24CU 3.7 TFLOP AMD GPU (RX VEGA M GH) is going to be anything like a match for a 32CU 4.9TFLOP GPU (RX470)

Nothing surprising about that. Who can blame me? 24CUs @ 1190MHz is not very close to the performance of an RX470 on paper, or even half Vega 56 on paper (whereas RX470 IS), so I have severe doubts it will be in practice! I will wait and see what miracles will have happened for it to be that fast.

I'm not saying it'll be a bad GPU, it looks amazing performance for integrated graphics and compared to existing Intel solutions, ridiculously superior and capable. Couldn't be happier about that.

 

[CaptainTom]

It's not semantics. It's just you also compared it to a desktop part, the RX 470.

I merely don't believe that a 24CU 3.7 TFLOP AMD GPU (RX VEGA M GH) is going to be anything like a match for a 32CU 4.9TFLOP GPU (RX470)

Erhh you are quite misinformed. The RX 470 is less than half as strong as Vega 64:

https://tpucdn.com/reviews/EVGA/GTX_1070_Ti_FTW2/images/perfrel_3840_2160.png

^Now that summation of benchmarks isn't perfect, but I would (personally) say techpowerup averages are usually within 10% of whatever a perfect unbiased average is. Close enough.

Furthermore Vega becomes substantially more efficient in perf/watt as you clock it lower. Really, these slides aren't insane at all. Vega comes after Polaris. It is the newer architecture, and it does have higher IPC....

 

[Vulcanproject]

Erhh you are quite misinformed. The RX 470 is less than half as strong as Vega 64:

https://tpucdn.com/reviews/EVGA/GTX_1070_Ti_FTW2/images/perfrel_3840_2160.png

^Now that summation of benchmarks isn't perfect, but I would (personally) say techpowerup averages are usually within 10% of whatever a perfect unbiased average is. Close enough

Nothing surprising about that. Who can blame me? 24CUs @ 1190MHz is not very close to the performance of an RX470 on paper, or even half Vega 56 on paper (whereas RX470 IS).

RX 470 is almost exactly half the performance of Vega 56. the 24 CU mobile part shown here definitely isn't. I do suggest you read a little more carefully.