TR Forums

So of course there's been a big deal about Intel putting Radeons into mobile chips

Well, here's a link to photographs of the actual chip in a real motherboard.

As you can see, it's quite a large package by mobile standards although it's small when you consider the space required by a traditional CPU + separate discrete GPU.

Incidentally, given the known size of 122.4mm^2 for the CPU die (on the left) I get a rough estimate of 188mm^2 for the GPU die on the right. You also see the single HBM2 stack on the right of the GPU. That maps pretty well to a Vega with 24 CUs since the biggest Vega 64 with 64 CUs weighs in at 484 mm^2, while this part has 37.5% of the CUs and about 38.8% of the die size.

I have a request. Can we please create a convention of calling it KLG/Kaby Lake G?

edit: I'm not criticizing your title Chuck. I just mean so many names are going to attach to this thing over the next 6 months it'll be very confusing I imagine.

DancinJack wrote:

I have a request. Can we please create a convention of calling it KLG/Kaby Lake G?

edit: I'm not criticizing your title Chuck. I just mean so many names are going to attach to this thing over the next 6 months it'll be very confusing I imagine.

You are right that Kaby Lake-G is the real name.
I just like having a little fun.

Look at all the power circuitry!!! That raises a lot of questions for me

TwistedKestrel wrote:

Look at all the power circuitry!!! That raises a lot of questions for me

There certainly is quite a bit since it is driving both a CPU and GPU. I'm pretty sure there are almost completely separate sets of regulators for the two pieces of silicon just as if this were a cpu + discrete GPU solution. That might change in the future if the level of integration increases.

Yeah, instead of being the kind of 6+2, or 8+2 power phase setup we're used to seeing, I guess this board would be something like 6+6+2? Just speculating, but there does seem to be a lot of components around that socket especially considering that this isn't a high end desktop board we're looking at.

Redocbew wrote:

Yeah, instead of being the kind of 6+2, or 8+2 power phase setup we're used to seeing, I guess this board would be something like 6+6+2? Just speculating, but there does seem to be a lot of components around that socket especially considering that this isn't a high end desktop board we're looking at.

I would assume that it is a 5 (core) +7 (GPU)+1 (HBM), but I can only guess without the board and multimeter in hand.

Man, that sure looks like a NUC-type board. Gimme gimme gimme.

DancinJack wrote:

Man, that sure looks like a NUC-type board. Gimme gimme gimme.

I'm guessing it's a test board and not a full commercial product, but the size is definitely in the M-ITX or maybe even a little smaller range (no PCIe slot).

I'm still uncertain as to how much of a benefit this is, as opposed to a discrete CPU/GPU. The biggest I see is that you are freeing up space on the motherboard; but does reducing that space (and assuming whatever wiring/connections go away as a result) translate to significant power savings? I can't imagine there are performance gains, given that it's still same CPU + GPU essentially. Is this the way of the future or is it more of a niche/novelty creation?

Space savings alone would be the payoff. I guess it's possible they could have tweaked the PCIe connection to save more power, but I don't know if that would gain much.

SecretMaster wrote:

Is this the way of the future or is it more of a niche/novelty creation?

I think this product will help answer that question. Hey, it's working well - we need more integration. Or, hey we threw a real GPU and HBM2 on there, and APUs still suck!

chuckula wrote:

DancinJack wrote:

Man, that sure looks like a NUC-type board. Gimme gimme gimme.

I'm guessing it's a test board and not a full commercial product, but the size is definitely in the M-ITX or maybe even a little smaller range (no PCIe slot).

Looks like a 5x5 (Mini-STX) board, perhaps?

derFunkenstein wrote:

Looks like a 5x5 (Mini-STX) board, perhaps?

I thought maybe that too, but from what I read most 5x5 stuff is LGA? Not that Intel couldn't have just BGA'd it for dev purposes.

I'm actually most interested to see how much the onboard HBM2 boosts the CPU performance. Could be even more than the Broadwell i7-5775c.

The Egg wrote:

I'm actually most interested to see how much the onboard HBM2 boosts the CPU performance. Could be even more than the Broadwell i7-5775c.

It won't boost the CPU performance at all outside of very specific situations where the regular Kaby Lake part would suffer because the IGP was sucking up most of the memory bandwidth for graphics and therefore the CPU cores would be bandwidth starved. In other words, the discrete GPU and the HBM2 remove standard CPU bottlenecks that commonly occur when the IGP sucks up resources, but there will be no actual performance boost to the CPU above & beyond what a normal Kaby Lake chip would do when it is not otherwise being bottlenecked.

The Egg wrote:

I'm actually most interested to see how much the onboard HBM2 boosts the CPU performance. Could be even more than the Broadwell i7-5775c.

I thought the HBM2 was only connected to the GPU and the CPU will still use regular DDR4 SDRAM? (which is in the photo)

derFunkenstein wrote:

The Egg wrote:

I'm actually most interested to see how much the onboard HBM2 boosts the CPU performance. Could be even more than the Broadwell i7-5775c.

I thought the HBM2 was only connected to the GPU and the CPU will still use regular DDR4 SDRAM? (which is in the photo)

I'm not sure, TBH. Being on the same interposer, I would've expected to CPU to have some degree of access (and that even if reduced compared to the GPU, it would still be at a much higher bandwidth than the main DDR4).

The gap between the CPU and GPU is rather wide. Is this the MCIG, or Minimum Competitive Integration Gap?

The Egg wrote:

derFunkenstein wrote:

The Egg wrote:

I'm actually most interested to see how much the onboard HBM2 boosts the CPU performance. Could be even more than the Broadwell i7-5775c.

I thought the HBM2 was only connected to the GPU and the CPU will still use regular DDR4 SDRAM? (which is in the photo)

I'm not sure, TBH. Being on the same interposer, I would've expected to CPU to have some degree of access (and that even if reduced compared to the GPU, it would still be at a much higher bandwidth than the main DDR4).

While things may very well change in the future, the first iteration of the product works this way: [Kaby Lake H] <-- Regular PCIe connection through the PCB --> [VegaLite GPU] <-- EMIB high-speed connection --> [HBM2 memory]. In other words, the layout of the CPU and GPU is basically identical to any desktop Kaby Lake part that has a discrete GPU plugged in over PCI express. There's no special access for the CPU to the HBM2 memory in these parts.

SecretMaster wrote:

I'm still uncertain as to how much of a benefit this is, as opposed to a discrete CPU/GPU. The biggest I see is that you are freeing up space on the motherboard; but does reducing that space (and assuming whatever wiring/connections go away as a result) translate to significant power savings? I can't imagine there are performance gains, given that it's still same CPU + GPU essentially. Is this the way of the future or is it more of a niche/novelty creation?

A few things. It should be the wave of the future, though it won't trickle down into cheap laptops anytime remotely soon. Keep in mind this setup does replace GDDR5 with HBM2, so it will reduce power draw over discrete mobile GPUs by a significant amount.

Going by a GamerNexus article Vega: Frontier Edition’s 16GB HBM2 pulls 20W, and RX480 uses 40-50W for an 8GB config. Considering KLG has a single HBM2 stack, power draw for the memory subsystem is probably around 10w I'm guessing, making the draw one third of a discrete mobile GPU with GDDR5. If not more. That's pretty significant savings for a SFF or laptop right there.

I'm pretty sure the HBM2 stack is only connected to the Vega GPU core.

The Egg wrote:

I'm actually most interested to see how much the onboard HBM2 boosts the CPU performance. Could be even more than the Broadwell i7-5775c.

It won't help directly as the CPU side has no reason to access the HBM memory connected to the GPU. The HBM isn't a cache level for the CPU (though the HBM may be configured as a cache for the GPU).

We're a few generations from the big game changer where Intel/AMD/the's massive ego/nVidia releases a SoC that uses HBM/WideIO for main memory while sacrificing all external memory to the SoC package. The result is a reduction in main memory latency while increasing bandwidth and while keeping the simple memory topology integrated graphics currently enjoy. Win, win, win except for your pocket book which now belongs to Intel/AMD/the's massive ego/nVidia. The usage of EMIB for HBM here means that you only get to loose your pocket book and not a kidney in the future.

Reading on a phone is weird. No way for the longer posts to show while displaying the author's name I was like "why is this guy bagging on the" and then I scrolled back up to see who wrote the post. Oh.

DancinJack wrote:

I have a request. Can we please create a convention of calling it KLG/Kaby Lake G?

edit: I'm not criticizing your title Chuck. I just mean so many names are going to attach to this thing over the next 6 months it'll be very confusing I imagine.

Wirko in another thread called it RAJACHIP, that has a real nice ring to it...................................

HERETIC wrote:

DancinJack wrote:

I have a request. Can we please create a convention of calling it KLG/Kaby Lake G?

edit: I'm not criticizing your title Chuck. I just mean so many names are going to attach to this thing over the next 6 months it'll be very confusing I imagine.

Wirko in another thread called it RAJACHIP, that has a real nice ring to it...................................

ATi 3D Raja Pro

TwistedKestrel wrote:

Look at all the power circuitry!!! That raises a lot of questions for me

I'm moderately surprised that we haven't seen CPU-socketed GPUs yet. The power and heat dissipation would be there for a flexible-socket design that could take either another CPU or a GPU.

With all the MCM designs and the abandoning of multi-socket motherboards that ship might have sailed though.

DancinJack wrote:

Man, that sure looks like a NUC-type board. Gimme gimme gimme.

Hey I might have been right! https://www.pcper.com/news/General-Tech ... cs-Spotted

Hades Canyon because, you know, hell froze over.

derFunkenstein wrote:

Hades Canyon because, you know, hell froze over.

Being somewhat serious it would not surprise me if that is the real reason for the codename.

derFunkenstein wrote:

Hades Canyon because, you know, hell froze over.

Occasionally somebody at Intel sneaks a funny name through.