T4C Fantasytoo much horsepower in a Vega M GH to be 29 vs 24 against a mx150 same settings

Yeah, MX150 goes from 15fps(ULV intel) to 21FPS(i7 HQ) at very high on that particular game. Though not knowing how notebookcheck is testing that game it's apple's to oranges comparison anyway(i.e. in their XPS 15 with gtx1050 review very high fps for RoTR is 32.9).

ValantarPulling ARK list prices when talking about OEM-made products is rather meaningless. Besides, Dell has officially stated that they can't cool 65W in the XPS 15 2-in-1. To quote Tom's Hardware:
That's pretty darn close to the 45W of a 30W MX150 and a 15W U-series CPU, and far more performance than the ~10W differential would suggest. Thus, the comparison is very interesting for me at least (I wouldn't really consider a laptop with a 1050/Ti unless it was very small and thin, while this is right up my wheelhouse). Getting this kind of performance at this kind of power in this kind of chassis? That's impressive.

Well, actually, we do, given thermal dissipation limitations on the designs. The Swift 3 is a semi-slim 15" in the semi-budget range. Probably doesn't have the best cooling, but no reason it should be throttling given the size of the chassis. Dell has, as quoted above, stated how much heat the XPS is able to dissipate. Does maximum heat dissipation equal maximum power draw? Not at every moment, but over time, yes. According to Notebookcheck, the Swift 3 (whole system) draws ~55W under load. It's also worth adding that the Swift is thicker than the XPS, weighs roughly the same, and has a larger chassis. In other words: there's no reason it shouldn't cool as well as the XPS (given the need to, at least, and thus the addition of a matching HSF). What Dell are demonstrating here is their success in stuffing more power and cooling than ever before into a really small chassis - and a 2-in-1 to boot.

Is it more expensive? Sure. But so is a slim GTX 1050 design, and you don't get those this thin and light. The Acer Spin 5 is a 15" 8550U+GTX 1050 2-in-1, but comes in at 2.3kg and is probably far larger (at least it's far larger than the Yoga 720 2-in-1, which has slim bezels like the XPS), and it draws ~75W under load. It does 31.2 FPS in NotebookCheck's RotR 1080p high test (4xAF, FXAA), but of course we don't know how this compares here.

Would this have been a "more fair" comparison? Depends what you're looking for. Absolute performance? Sure. Performance per weight or size? No. Performance per watt? Possibly, but not necessarily.

@Valantar This is what DELL allowed users to compare against during CES 2018. From btarunr's article he claims these are "similarly priced" items, which they are not. Period. You cannot possibly "guesstimate" the configuration's TDP on the other devices and honestly conclude they are in some similar class. Sorry. I understand TPU's need to get this out there before everyone else, but I would've been much more careful about the phrasing (especially when claiming this obliterates the competition). The Swift 3 model which sports an MX150 costs $800, this Dell 2-in-1 "starts at $1299" so we're hardly comparing similar configurations. I won't even comment about the comparison to the HD620 IGP... that's hilarious!

I like that AMD and Intel found common ground to collaborate. I dislike the fact that websites are talking "obliteration" when comparing shitty-specced machines to Intel/AMD's "latest and greatest". You either compare machines of similar price (and your deltas are performance, battery life and overall portability), or you compare hardware of similar generations and specs (where your deltas are price and battery life). These are neither.

LemmingOverlord@Valantar This is what DELL allowed users to compare against during CES 2018. From btarunr's article he claims these are "similarly priced" items, which they are not. Period. You cannot possibly "guesstimate" the configuration's TDP on the other devices and honestly conclude they are in some similar class. Sorry. I understand TPU's need to get this out there before everyone else, but I would've been much more careful about the phrasing (especially when claiming this obliterates the competition). The Swift 3 model which sports an MX150 costs $800, this Dell 2-in-1 "starts at $1299" so we're hardly comparing similar configurations. I won't even comment about the comparison to the HD620 IGP... that's hilarious!

I like that AMD and Intel found common ground to collaborate. I dislike the fact that websites are talking "obliteration" when comparing shitty-specced machines to Intel/AMD's "latest and greatest". You either compare machines of similar price (and your deltas are performance, battery life and overall portability), or you compare hardware of similar generations and specs (where your deltas are price and battery life). These are neither.

I'm not saying that the use of words like "obliterate" isn't rather dumb, nor that the comparisons here are the best possible, just to get that right. Also, the "similarly priced" claim is indeed shaky, if not entirely false (you can easily find laptops with those specs that exceed $1000, after all). Still, your two modes of comparison are rather narrow for my tastes. Sure, those are common starting points. But for many users, the starting point can be "what kind of performance can I get in a size/weight that I can carry with me every day", a question that requires consideration of both size and weight - and thus power and cooling. "What kind of performance can I get in a given form factor" is also a valid starting point. Hence my comparison to the GTX 1050-equipped Spin 5, which checks the boxes for comparable performance (at least on paper) and form factor, but loses out in size and weight by a significant margin. The Lenovo Yoga 720 15" is also a good comparison, with strikingly similar specs. It's a bit thicker, but not by much, and has a H-series CPU and a GTX 1050. As such, it'll be a great point of comparison once reviews start arriving, and it'll tell us something about the veracity of Intel's space/power saving claims and Dell's implementation.

As for the "This is what Dell allowed users to compare against" - that's a weird claim. Did Dell have competing PCs from other brands lined up in their booth? And even so, how difficult would it be for whoever ran the tests to simply take a note of the settings and have a comparison run on whatever hardware would be suitable? It's not like Dell can police journalists' benchmarking practices, after all. The biggest challenge there would be access to comparable hardware, which I assume to be the reasoning behind the rather lopsided comparison here. Of course, the transparent and right thing to do here would be to also report detailed settings for the benchmark, which as I see it is the only real error here outside of some over-the-top, semi-clickbaity phrasing.

Going from the HotHardware source article, we know the tests were run at 1080p high, with 29.69fps reported for 1080p very high (no word on AF or AA settings, but no reason to assume they're not stock). Assuming that NotebookCheck's RotR benchmark is the built-in one, that's a bit below the Yoga 720s result of 32 fps and the Spin 5's result of 31.2.

As for "guesstimating" power draw, it's clear you didn't actually read what I posted. Every number I posted outside of Dell's claims for the XPS 15's cooling capacity isn't TDP, but rather actual measured power draw under real-world loads by a trustworthy third party. Of course, Dell is talking CPU/GPU cooling capacity, while the review numbers are for the full system, which probably adds 10-15W to that depending on the configuration. If Dell says they can cool 55W, I assume the whole laptop will draw ~70W, if not closer to 75 with the 4k display. For good measure, let's add 20W to that for peak power draw for short periods of time. More than that is extremely unlikely, unless Dell manages to bork this design beyond belief. Comparing that to real-world numbers from the Yoga 720, it peaks at almost 120W during FurMark+Prime95, and hovers just below 100W playing The Witcher 3. That's a testament to how well-built its cooling system is (even if the CPU throttles to 800MHz during the power virus test), but it also shows the real-world power draw of something that we expect KBL-G to be comparable to. Are my XPS power draw numbers guesstimates? Sure. But they're well founded, and you don't seem to say otherwise. Even if I'm 10% off, we're still seeing significantly less power draw than the Yoga 720 for performance that's almost within margin of error.

To me, it seems like KBL-G makes a pretty good middle ground between low-end KBL-R+MX150 systems and bigger, bulkier systems using the GTX 1050, while being closer to the latter in performance (and essentially the same price). If that's how a power-throttled pre-production KBL-G performs, this is looking promising.

While the comparisons in the original HotHardware article are less than ideal (the HD 620? that's ridiculous (but probably also what they had access to)), the wording rather dumb, and TPU re-using that wording possibly even worse, that's not the most interesting thing here. Sure, we can talk about lack of journalistic integrity, which is a valuable topic to discuss (especially in the time of clickbait deluges and fake news). In this specific case, I'm more interested in discussing the actual news, regardless of its packaging, as it is rather significant.

LemmingOverlordNo, they're not. You have no idea about the price or power consumption on each one of those devices. If they'd stated the retail price on each configuration *AND* performed battery tests, I'd even accept your reply... but no.

@btarunr article states they are similarly priced, but I find that hard to believe. Core i7 vs. Core i5 vs. Core i3 vs. iGPU vs. dGPU vs. whatever. Is TPU the source for this actual article?

You're right. I have no idea. But then I never claimed I was a source. I'm only going by what the article states. It seems some here lack reading comprehension to do the same. Or they're merely trying to move the goalpost away from what the article talks about.

ValantarAs for "guesstimating" power draw, it's clear you didn't actually read what I posted. Every number I posted outside of Dell's claims for the XPS 15's cooling capacity isn't TDP, but rather actual measured power draw under real-world loads by a trustworthy third party. Of course, Dell is talking CPU/GPU cooling capacity, while the review numbers are for the full system, which probably adds 10-15W to that depending on the configuration. If Dell says they can cool 55W, I assume the whole laptop will draw ~70W, if not closer to 75 with the 4k display. For good measure, let's add 20W to that for peak power draw for short periods of time. More than that is extremely unlikely, unless Dell manages to bork this design beyond belief. Comparing that to real-world numbers from the Yoga 720, it peaks at almost 120W during FurMark+Prime95, and hovers just below 100W playing The Witcher 3. That's a testament to how well-built its cooling system is (even if the CPU throttles to 800MHz during the power virus test), but it also shows the real-world power draw of something that we expect KBL-G to be comparable to. Are my XPS power draw numbers guesstimates? Sure. But they're well founded, and you don't seem to say otherwise. Even if I'm 10% off, we're still seeing significantly less power draw than the Yoga 720 for performance that's almost within margin of error.
.

Hi,

Just found this and you seem to know what you're talking about. maybe you could illuminate something i've been wondering about some time.

It's the TDP.

The early bench shows around 1050 TI, but I'm not even convinced that they'll do 1050 or even maxQ 1050 (if such a thing exists).
Many sources claim that KBL-G have the 'H' series processor integrated with the GPU unit. 1050 alone has TDP of 60 something Watts, give or take a few. The whole thing is rated at 60W.

Unless the EIMD bridge thing breaks physics and creates more power than it consumes, this can't be done.

Now, some articles do say that the 60W that this thing consumes have intelligent controller that directs power to where it needs to go between the CPU and GPU.

if it is indeed the H series it's 60W between 45+60 (or more on GPU).

how does this thing work?

turns out the Vega M is polaris 22 and not vega at all, ID is polaris and the codename is NDA polaris

T4C Fantasyturns out the Vega M is polaris 22 and not vega at all, ID is polaris and the codename is NDA polaris

what the hell? so at best we are looking at macbook pro 15 specs at this point?
I was literally holding off buying a mbp15 because I thought the H series will be coming out next month.

beerandchipswhat the hell? so at best we are looking at macbook pro 15 specs at this point?
I was literally holding off buying a mbp15 because I thought the H series will be coming out next month.

its still modified with HBM2 but its Polaris, Vega M is gfx804, real vega is gfx900

also a RX550 is gfx804, which means Polaris 12, this NDA codename for Vega M is literally Polaris 22

beerandchipsHi,

Just found this and you seem to know what you're talking about. maybe you could illuminate something i've been wondering about some time.

It's the TDP.

The early bench shows around 1050 TI, but I'm not even convinced that they'll do 1050 or even maxQ 1050 (if such a thing exists).
Many sources claim that KBL-G have the 'H' series processor integrated with the GPU unit. 1050 alone has TDP of 60 something Watts, give or take a few. The whole thing is rated at 60W.

Unless the EIMD bridge thing breaks physics and creates more power than it consumes, this can't be done.

Now, some articles do say that the 60W that this thing consumes have intelligent controller that directs power to where it needs to go between the CPU and GPU.

if it is indeed the H series it's 60W between 45+60 (or more on GPU).

how does this thing work?

Now, I don't claim to understand this fully, let alone know anything for sure, but the answer is probably as "simple" as optimization and aggressive power budgeting between the CPU and GPU with some throttling when both are under load, plus some hardware advantages. Also, some of your numbers are a bit off. Remember that there are 65W and 100W versions of these chips, with varying CU counts, core clocks, GPU clocks, and more.

Firstly, for CPUs the TDP isn't very relevant for gaming loads. 45W mobile CPUs don't actually consume 45W while gaming, as gaming is rarely a consistent 100% load across all cores. As such, the performance loss of limiting the CPU to, say, 25-30W (dynamically) while the GPU is under load is probably barely noticeable. As an example, look at the measured power draw of this laptop with a 7300HQ (45W) and a 1050Ti vs this one with an 8550U (@25W) and a 1050Ti. They're essentially the same, and a 7700HQ doesn't increase that by all that much.

The GPU goes for a wide-and-slow approach too, which is always more efficient than pushing clocks high (though it has the downside of higher fab costs with larger die sizes - which Intel compensates for by pricing this kind into oblivion). The KBL-G chips have 20 or 24 CUs enabled. For comparison, at regular clocks, the 16-CU RX 560 is comparable to the 1050Ti desktop. Adding 50% more CUs and aiming for the same performance window would allow for a serious downclock - which the .9-1.2GHz specs corroborate. Vega at 1-1.2GHz is scary efficient, at times outpacing Pascal for perf/W in similar low-clocked scenarios. Now, the people above might be right that Vega M is actually Polaris (although that sounds unlikely to me, at least unless it's some in-between arch without Rapid Packed Math or other compute-focused functionality), but for now, I prefer to think that Vega=Vega.

Then there's GDDR5 vs HBM, which is a major advantage in small power envelopes like this. I can't remember the source or the exact numbers for this at the moment, but I've read a HBM power analysis that compared Vega to Pascal at ~250W (I believe it was Vega FE vs Titan X(p)). That analysis estimated a ~50W power draw from the GDDR5X on the Titan, vs. <20W for the HBM2 (Edit: it might have been thisGamersNexus article, which clearly states Vega FE memory consumes <20W). GDDR5X is more efficient at high clocks than GDDR5, so it's not unreasonable to expect the 4GB on a mobile 1050 to consume around 10-15W despite lower clocks. The GN article estimates the 8GB of 8GBPS GDDR5 on an RX 480 on a 256-bit bus to consume 40-50W, so RAM could easily be a larger portion of the ~50W TDP of a mobile 1050Ti (the 1050Ti clocks the ram at 7GBPS on a 128-bit bus, so it should consume less than half of the RAM of the 480 - but not far less). A single 4GB stack of lower-than-Vega FE clocked HBM2 should more than halve that power draw, with theoretical numbers (according to specs) for a 4GB single stacks running around 3.75W. If we assume 15W vs 5W - which seems reasonable for real-world numbers to me, if not overly nice to Nvidia - that's 10W saved right there.

So, let's say we start out with the 45W + ~70W (according to NotebookCheck) combination of a H-series i7 and a mobile 1050Ti. That's 115W. Subtract 20W for the dynamic CPU limit, another 10-15 for the RAM swap, and you're around 80-85W. Then add the efficiency of a wide-and-slow Vega setup, and you won't have to reduce performance much at all to hit a 65W goal. The 931-1101MHz clocks of the 65W KBL-G chips is what makes this believable to me, as that should be an incredibly efficient clock range for Vega.

Again: all of this is speculation. But I don't doubt that it's possible, given the money and resources (and lack of qualms about selling expensive chips).

ValantarNow, I don't claim to understand this fully, let alone know anything for sure, but the answer is probably as "simple" as optimization and aggressive power budgeting between the CPU and GPU with some throttling when both are under load, plus some hardware advantages. Also, some of your numbers are a bit off. Remember that there are 65W and 100W versions of these chips, with varying CU counts, core clocks, GPU clocks, and more.

Firstly, for CPUs the TDP isn't very relevant for gaming loads. 45W mobile CPUs don't actually consume 45W while gaming, as gaming is rarely a consistent 100% load across all cores. As such, the performance loss of limiting the CPU to, say, 25-30W (dynamically) while the GPU is under load is probably barely noticeable. As an example, look at the measured power draw of this laptop with a 7300HQ (45W) and a 1050Ti vs this one with an 8550U (@25W) and a 1050Ti. They're essentially the same, and a 7700HQ doesn't increase that by all that much.

The GPU goes for a wide-and-slow approach too, which is always more efficient than pushing clocks high (though it has the downside of higher fab costs with larger die sizes - which Intel compensates for by pricing this kind into oblivion). The KBL-G chips have 20 or 24 CUs enabled. For comparison, at regular clocks, the 16-CU RX 560 is comparable to the 1050Ti desktop. Adding 50% more CUs and aiming for the same performance window would allow for a serious downclock - which the .9-1.2GHz specs corroborate. Vega at 1-1.2GHz is scary efficient, at times outpacing Pascal for perf/W in similar low-clocked scenarios. Now, the people above might be right that Vega M is actually Polaris (although that sounds unlikely to me, at least unless it's some in-between arch without Rapid Packed Math or other compute-focused functionality), but for now, I prefer to think that Vega=Vega.

Then there's GDDR5 vs HBM, which is a major advantage in small power envelopes like this. I can't remember the source or the exact numbers for this at the moment, but I've read a HBM power analysis that compared Vega to Pascal at ~250W (I believe it was Vega FE vs Titan X(p)). That analysis estimated a ~50W power draw from the GDDR5X on the Titan, vs. <20W for the HBM2 (Edit: it might have been thisGamersNexus article, which clearly states Vega FE memory consumes <20W). GDDR5X is more efficient at high clocks than GDDR5, so it's not unreasonable to expect the 4GB on a mobile 1050 to consume around 10-15W despite lower clocks. The GN article estimates the 8GB of 8GBPS GDDR5 on an RX 480 on a 256-bit bus to consume 40-50W, so RAM could easily be a larger portion of the ~50W TDP of a mobile 1050Ti (the 1050Ti clocks the ram at 7GBPS on a 128-bit bus, so it should consume less than half of the RAM of the 480 - but not far less). A single 4GB stack of lower-than-Vega FE clocked HBM2 should more than halve that power draw, with theoretical numbers (according to specs) for a 4GB single stacks running around 3.75W. If we assume 15W vs 5W - which seems reasonable for real-world numbers to me, if not overly nice to Nvidia - that's 10W saved right there.

So, let's say we start out with the 45W + ~70W (according to NotebookCheck) combination of a H-series i7 and a mobile 1050Ti. That's 115W. Subtract 20W for the dynamic CPU limit, another 10-15 for the RAM swap, and you're around 80-85W. Then add the efficiency of a wide-and-slow Vega setup, and you won't have to reduce performance much at all to hit a 65W goal. The 931-1101MHz clocks of the 65W KBL-G chips is what makes this believable to me, as that should be an incredibly efficient clock range for Vega.

Again: all of this is speculation. But I don't doubt that it's possible, given the money and resources (and lack of qualms about selling expensive chips).

Too much proof points to vega m being polaris, calling it vega is factually incorrect, it has the architecture id of polaris 12 and its not wrong

T4C FantasyToo much proof points to vega m being polaris, calling it vega is factually incorrect, it has the architecture id of polaris 12 and its not wrong

Call me too lazy to Google, but care to show us any of this evidence? I'd be especially interested in this considering we've seen photos of both the Vega M dGPU module and the custom Intel Vega M (which, btw, don't look to be the same chips - the proportions don't quite match), and thus know for a fact that Vega M is a significantly larger die than Polaris 12. The HBM2 stack makes it rather easy to estimate die size, as its measurements are standardized. Also, Polaris 12 is one specific Polaris die, the one used in the 550 and 560, which lacks the CUs (16, not 24) and thus die size to be Vega M. Not to mention Polaris 12 lacks a HBM controller. So, Vega M - whether dGPU or Intel semi custom - is not Polaris 12. Might Vega M still be some sort of Polaris 12-adjacent half-Polaris, half-Vega arch with the compute-specific parts of Vega stripped out? Sure. Why not, especially if they could get Intel to pay for the base R&D for this amalgamation? But saying "it's Polaris 12" due to some 'leaks' and rumors seems like a stretch to me.

For all we know, the architecture ID might be misidentified by software unable to correctly read unreleased hardware, or just a placeholder based on the closest released hardware resource-wise for drivers to work. I'd be happy to be proved wrong, though, if you can provide some sources.

ValantarCall me too lazy to Google, but care to show us any of this evidence? I'd be especially interested in this considering we've seen photos of both the Vega M dGPU module and the custom Intel Vega M (which, btw, don't look to be the same chips - the proportions don't quite match), and thus know for a fact that Vega M is a significantly larger die than Polaris 12. The HBM2 stack makes it rather easy to estimate die size, as its measurements are standardized. Also, Polaris 12 is one specific Polaris die, the one used in the 550 and 560, which lacks the CUs (16, not 24) and thus die size to be Vega M. Not to mention Polaris 12 lacks a HBM controller. So, Vega M - whether dGPU or Intel semi custom - is not Polaris 12. Might Vega M still be some sort of Polaris 12-adjacent half-Polaris, half-Vega arch with the compute-specific parts of Vega stripped out? Sure. Why not, especially if they could get Intel to pay for the base R&D for this amalgamation? But saying "it's Polaris 12" due to some 'leaks' and rumors seems like a stretch to me.

For all we know, the architecture ID might be misidentified by software unable to correctly read unreleased hardware, or just a placeholder based on the closest released hardware resource-wise for drivers to work. I'd be happy to be proved wrong, though, if you can provide some sources.

they dont need to fit the dimensions,
AMD GPUs are basically made like LEGO structures... AMD can play with the CUs almost any way they want
if they have a Polaris CU (with 64 shaders), they can pack any number of it into a GPU, and pair it up with virtually any memory subsystem
that's why they were hired to do the console GPUs... very flexible, not like Intel's iGPUs

gfxbench.com/device.jsp?benchmark=gfx40&os=Windows&api=gl&D=AMD+694C:C0&testgroup=info

browser.geekbench.com/v4/compute/811174

its Polaris 22 i said its architecture was 12

Amd/comments/7ahb6z
no evidence its vega xD

i say it again... its not Vega.
Polaris = gfx800+
Vega = gfx900, 901 (901 is in Ryzen 2000G series)

T4C Fantasythey dont need to fit the dimensions,
AMD GPUs are basically made like LEGO structures... AMD can play with the CUs almost any way they want
if they have a Polaris CU (with 64 shaders), they can pack any number of it into a GPU, and pair it up with virtually any memory subsystem
that's why they were hired to do the console GPUs... very flexible, not like Intel's iGPUs

gfxbench.com/device.jsp?benchmark=gfx40&os=Windows&api=gl&D=AMD+694C:C0&testgroup=info

browser.geekbench.com/v4/compute/811174

its Polaris 22 i said its architecture was 12

Amd/comments/7ahb6z
no evidence its vega xD

i say it again... its not Vega.
Polaris = gfx800+
Vega = gfx900, 901 (901 is in Ryzen 2000G series)

You seem to be missing a key point here: there's no such thing as a "Polaris 12 architecture". Polaris is the architecture, Polaris 12 is a specific chip found in the RX 550 and 560 with 16 CUs and a fixed feature set. There's no architectural difference between Polaris 10, 11, or 12 beyond minor tweaks and optimizations that don't amount to anything near an architectural revision. While there are revisions of the Polaris architecture between chips, there are no named revisions. As such, there's no way of identifying revisions based on hardware ID numbers - those numbers identify specific chip designs. Any new chip with more or less CUs and other features changed would necessarily have a new hardware ID, as that would be necessary for drivers to utilize the correct paths and optimizations.

As for the "lego-like" nature of AMD GPUs, you're not telling anyone here anything we don't know - but you're not answering my questions either. See above. Given that GFX804 is Polaris 12, how can it have 20-24 CUs as per the official spec sheets of KBL-G? That's impossible, as Polaris 12 has 16 CUs and a far smaller die size. In other words: this is not Polaris 12. Which, again, tells us that the GFX804 Hardware ID is erroneous in some way.

Now, the GFXBench link above is the only one actually saying anything linking GFX804 with 694C/694E, the rest are just benchmarks with varying degrees of misidientification of the CPU, GPU or both. Which is to be expected with pre-release hardware. If you see the others as evidence of any link between Vega M and the Polaris architecture, you'll have to spell that out for us, 'cause all the rest of the links are just saying "Vega M is 694C/694E" which tells us nothing at all.

As for the GFXBench thing, it's also interesting to note that only the OpenCL part identifies it as GFX804.



Now, I actually have a couple of hypotheses as to why we're seeing this weirdness, which are far friendlier with Ockham's razor than "'Vega M' is Polaris, AMD is lying":

First, some base facts:

• KBL-G is a fully Intel product, with the entire driver stack provided by Intel (see launch information for confirmation of this from AMD)
• AMD then must provide Intel with base driver code to integrate into the Intel driver stack for KBL-G - Intel doesn't write drivers for AMD hardware, but probably optimizes it
• the KBL-G "pGPU" is a semi-custom chip, not entirely matching any existing amd GPU and thus requiring specific driver optimizations
• Support for new GPU designs rarely, if ever, appears in drivers until launch day or close to it
• Intel is selling this as a gaming chip, and has as of now not even hinted at pro-level aspirations for this design.

None of these things are new assumptions. Combining points 3 and 4 gives us two options for pre-release testing: either tweak existing drivers to misidentify the GPU to enable not-yet-finalized optimizations from another not-that-dissimilar chip (similar CU count/compute resources), or use not-yet-ready pre-release driver code lacking all these optimizations. They're probably doing both, but the first option would definitely be done as a step to seeing how the drivers would need to be tweaked compared to what's already released.

Then there's the fact that Intel is doing the testing, not AMD. They're on a tight schedule, requiring them to get this working quickly, while working with relatively unknown code and unknown hardware. Could this lead to the use of incorrect Hardware IDs as placeholders? I don't think that's unlikely - and certainly not as unlikely as AMD blurring the distinction between Polaris and Vega (which are separate architectures in very significant ways that don't relate to marketing whatsoever). Even though this project has been in the works for 2+ years, Intel hasn't had access to fully-functioning hardware or drivers until very recently.

Then there's the gaming-focused nature of the chip. If Intel has to pay AMD for every part of their hardware design AND software stack, and the OpenCL part is barely goint to be used at all (outside of professional applications, which isn't gaming, so not what this chip is marketed for). Why pay extra for a newer, more optimized version of the OpenCL driver stack if you don't need it, or could get an older one for cheaper? Or what if AMD is unwilling to part with it, to not cannibalize sales of Radeon Pro parts like in the MBP range and upcoming Vega Pro cards? Considering the compute-focused nature of the Vega architecture, it makes very much sense for AMD to not give Intel full access to the driver stack here, limiting the access to new, optimized code to the graphics/gaming parts.

In other words, this could be

• Placeholder IDs for early hardware testing, implemented to get drivers to work before significant rewriting can be done, or
• The result of AMD not giving Intel their newest OpenCL drivers, with Intel implementing the code "as-is" in pre-release testing

Neither of these hypotheses require any new assumptions - such as AMD calling a Polaris arch Vega for ... marketing purposes? - and as such fit well with Ockham's razor. Your hypothesis requires assuming that both AMD and Intel are lying or at best twisting the truth, which is a far bigger leap to make, and thus requires more evidence. What you've presented so far does not hold up to my standards, sadly. Heck, outside of the GFXBench listings, there's no link there at all. If you have more, feel free to post it here, but for now I'll chalk it up to speculation with pretty shaky foundations.

ValantarYou seem to be missing a key point here: there's no such thing as a "Polaris 12 architecture". Polaris is the architecture, Polaris 12 is a specific chip found in the RX 550 and 560 with 16 CUs and a fixed feature set. There's no architectural difference between Polaris 10, 11, or 12 beyond minor tweaks and optimizations that don't amount to anything near an architectural revision. While there are revisions of the Polaris architecture between chips, there are no named revisions. As such, there's no way of identifying revisions based on hardware ID numbers - those numbers identify specific chip designs. Any new chip with more or less CUs and other features changed would necessarily have a new hardware ID, as that would be necessary for drivers to utilize the correct paths and optimizations.

As for the "lego-like" nature of AMD GPUs, you're not telling anyone here anything we don't know - but you're not answering my questions either. See above. Given that GFX804 is Polaris 12, how can it have 20-24 CUs as per the official spec sheets of KBL-G? That's impossible, as Polaris 12 has 16 CUs and a far smaller die size. In other words: this is not Polaris 12. Which, again, tells us that the GFX804 Hardware ID is erroneous in some way.

Now, the GFXBench link above is the only one actually saying anything linking GFX804 with 694C/694E, the rest are just benchmarks with varying degrees of misidientification of the CPU, GPU or both. Which is to be expected with pre-release hardware. If you see the others as evidence of any link between Vega M and the Polaris architecture, you'll have to spell that out for us, 'cause all the rest of the links are just saying "Vega M is 694C/694E" which tells us nothing at all.

As for the GFXBench thing, it's also interesting to note that only the OpenCL part identifies it as GFX804.



Now, I actually have a couple of hypotheses as to why we're seeing this weirdness, which are far friendlier with Ockham's razor than "'Vega M' is Polaris, AMD is lying":

First, some base facts:

• KBL-G is a fully Intel product, with the entire driver stack provided by Intel (see launch information for confirmation of this from AMD)
• AMD then must provide Intel with base driver code to integrate into the Intel driver stack for KBL-G - Intel doesn't write drivers for AMD hardware, but probably optimizes it
• the KBL-G "pGPU" is a semi-custom chip, not entirely matching any existing amd GPU and thus requiring specific driver optimizations
• Support for new GPU designs rarely, if ever, appears in drivers until launch day or close to it
• Intel is selling this as a gaming chip, and has as of now not even hinted at pro-level aspirations for this design.

None of these things are new assumptions. Combining points 3 and 4 gives us two options for pre-release testing: either tweak existing drivers to misidentify the GPU to enable not-yet-finalized optimizations from another not-that-dissimilar chip (similar CU count/compute resources), or use not-yet-ready pre-release driver code lacking all these optimizations. They're probably doing both, but the first option would definitely be done as a step to seeing how the drivers would need to be tweaked compared to what's already released.

Then there's the fact that Intel is doing the testing, not AMD. They're on a tight schedule, requiring them to get this working quickly, while working with relatively unknown code and unknown hardware. Could this lead to the use of incorrect Hardware IDs as placeholders? I don't think that's unlikely - and certainly not as unlikely as AMD blurring the distinction between Polaris and Vega (which are separate architectures in very significant ways that don't relate to marketing whatsoever). Even though this project has been in the works for 2+ years, Intel hasn't had access to fully-functioning hardware or drivers until very recently.

Then there's the gaming-focused nature of the chip. If Intel has to pay AMD for every part of their hardware design AND software stack, and the OpenCL part is barely goint to be used at all (outside of professional applications, which isn't gaming, so not what this chip is marketed for). Why pay extra for a newer, more optimized version of the OpenCL driver stack if you don't need it, or could get an older one for cheaper? Or what if AMD is unwilling to part with it, to not cannibalize sales of Radeon Pro parts like in the MBP range and upcoming Vega Pro cards? Considering the compute-focused nature of the Vega architecture, it makes very much sense for AMD to not give Intel full access to the driver stack here, limiting the access to new, optimized code to the graphics/gaming parts.

In other words, this could be

• Placeholder IDs for early hardware testing, implemented to get drivers to work before significant rewriting can be done, or
• The result of AMD not giving Intel their newest OpenCL drivers, with Intel implementing the code "as-is" in pre-release testing

Neither of these hypotheses require any new assumptions - such as AMD calling a Polaris arch Vega for ... marketing purposes? - and as such fit well with Ockham's razor. Your hypothesis requires assuming that both AMD and Intel are lying or at best twisting the truth, which is a far bigger leap to make, and thus requires more evidence. What you've presented so far does not hold up to my standards, sadly. Heck, outside of the GFXBench listings, there's no link there at all. If you have more, feel free to post it here, but for now I'll chalk it up to speculation with pretty shaky foundations.

The clear evidence of all is the permanent device id placement, maybe saying architecture of polaris 12 was a bit too much but it is more polaris than vega. I know my ids, have you seen the gpudb? all are linked to device ids. I know placeholders and this isnt one of them.

On my free time i maintain the largest device id database aswel
pci-ids.ucw.cz/read/PC/1002
pci-ids.ucw.cz/read/PC/10de

T4C FantasyThe clear evidence of all is the permanent device id placement, maybe saying architecture of polaris 12 was a bit too much but it is more polaris than vega. I know my ids, have you seen the gpudb? all are linked to device ids. I know placeholders and this isnt one of them.

On my free time i maintain the largest device id database aswel
pci-ids.ucw.cz/read/PC/1002
pci-ids.ucw.cz/read/PC/10de

Well, I still find it hard to believe that AMD has decided to re-use a hardware ID across different dice for the first time ever. Not to mention using the same name for fundamentally different architectures for the first time. There are enough firsts for this die (first significant consumer PC deployment of semi-custom GPU silicon, first mobile HBM2, first pGPU/EMIB, first AMD GPU with drivers not distributed by AMD, first AMD GPU not sold by AMD or their AIB partners) to make other explanations sound reasonable to me, at least. There are plenty of steps along the way where a placeholder might have been implemented or a naming error might have stuck in ES silicon. For all we know, it might be a way to flag the GPUs so that drivers limit them to Polaris-level functionality (no RPM, no HBCC, and so on), to avoid end users "unlocking" unsupported (read: not licensed by Intel) functionality through AMD drivers. I suppose we'll see once retail hardware starts appearing.


Also: where do you get this "permanent device ID" stuff from? All we've seen is pre-release/ES hardware running pre-release software. A lot of things can change between ES and retail.

ValantarWell, I still find it hard to believe that AMD has decided to re-use a hardware ID across different dice for the first time ever. Not to mention using the same name for fundamentally different architectures for the first time. There are enough firsts for this die (first significant consumer PC deployment of semi-custom GPU silicon, first mobile HBM2, first pGPU/EMIB, first AMD GPU with drivers not distributed by AMD, first AMD GPU not sold by AMD or their AIB partners) to make other explanations sound reasonable to me, at least. There are plenty of steps along the way where a placeholder might have been implemented or a naming error might have stuck in ES silicon. For all we know, it might be a way to flag the GPUs so that drivers limit them to Polaris-level functionality (no RPM, no HBCC, and so on), to avoid end users "unlocking" unsupported (read: not licensed by Intel) functionality through AMD drivers. I suppose we'll see once retail hardware starts appearing.

well thats not a first time at all.. they do it every year for workstation, desktop, mobile, mobile workstation... IGPs.. they are known for this also i have inside info but ill share it.. i cant prove it as its inside info and noone else but inside people know it..
the graphics core is based on polaris via bios and register layout and the IMC registry is the same as FIJI (fury series)... this isnt a rumor or a grain of salt.. its fact.. but i know its not enough for you.. so i end this convo with... im sorry i cant prove it to you that its polaris.. but i know what it is.

T4C Fantasywell thats not a first time at all.. they do it every year and mostly for workstation, desktop, mobile, mobile workstation... IGPs.. they are known for this also i have inside info but ill share it.. i cant prove it as its inside info and noone else but inside people do..
the graphics core is based on polaris via bios and register layout and the IMC registry is the same as FIJI (fury series)... this isnt a rumor or a grain of salt.. its fact.. but i know its not enough for you.. so i end this convo with... im sorry i cant prove it to you that its polaris.. but i know what it is.

I mean it makes sense to use those parts. Why make something new when you have off the shelf parts.

T4C Fantasywell thats not a first time at all.. they do it every year for workstation, desktop, mobile, mobile workstation... IGPs.. they are known for this also i have inside info but ill share it.. i cant prove it as its inside info and noone else but inside people know it..
the graphics core is based on polaris via bios and register layout and the IMC registry is the same as FIJI (fury series)... this isnt a rumor or a grain of salt.. its fact.. but i know its not enough for you.. so i end this convo with... im sorry i cant prove it to you that its polaris.. but i know what it is.

I suppose it would be interesting to see when this hits retail. But to clarify: they regularly re-use hardware IDs across different dice, not different products based on the same die? Remember, any die - let's use Polaris 10 as an example - any Polaris 10 die is the same design regardless if it's in an RX 480, 470, Radeon Instinct accelerator or Radeon Pro workstation card. These cards sharing a hardware ID (but having different board IDs) is thus entirely natural - anything else would be very odd, given that they're based on the same silicon. Could you give an example of hardware IDs being shared across different dice with different specs (i.e. not just a new hardware revision or re-spin of an existing design)?

As for the insider info: it might very well be true. I know for sure that I don't have anything like that. But again, I'll believe it when it can actually be corroborated :)

cdawallI mean it makes sense to use those parts. Why make something new when you have off the shelf parts.

That's what semi-custom means, more or less. The customer goes to AMD, says "I want a chip with this part, this part, and four of these parts", and AMD makes the chip for them. Unless the customer has crazy amounts of money and very specific needs, they use "off-the-shelf" component designs, as anything else would be wildly expensive and time consuming. I suppose a relevant question here is whether Vega was available for the development of semi-custom designs 2+ years ago when this collaboration reportedly began. Given that retail Vega launched half a year ago, I don't see that as unlikely, especially given that retail Vega was delayed. After all, there's no reason why this design had to be finalized and initial fab test runs started until after the retail launch, and reportedly both the PS4 Pro and Xbox One X have Vega elements in their GPU cores. This is also something that makes me doubt that this is a Polaris design - by now, it's simply too old. Might it be Polaris-adjacent, or Vega with features disabled to approach Polaris feature parity? Sure. Would probably be cheaper for Intel.


Another part of what makes me doubt that this could be Polaris (i.e. GCN 1.4 CUs, and not Vega NCUs, mainly) is power consumption: the Radeon RX 470D, with 28 Polaris CUs, has a Total Board Power of 150W at 1266MHz. Even subtracting the full ~30W for the 4GB of GDDR5 on there, 4/28ths to bring it to CU parity, and another 5% or so for power delivery losses that are externalized by the on-package GPU, you're still pushing the power envelope of the Vega M GH (~98W, according to my way oversimplified math). Which, of course, doesn't include the CPU part attached to the Vega M. A 76MHz downclock to the 1190MHz boost spec of the i7-8709G and 8809G doesn't account for 25W+ of power, nor should a downclock to the base speed of 1063MHz do so either. In other words: a pure Polaris part would need some serious efficiency tuning to reach these power numbers.

But again: I don't have any industry sources telling me anything, and there's every possibility I'm wrong. It just doesn't seem likely AMD would call this Vega unless it was significantly Vega based.

ValantarAnother part of what makes me doubt that this could be Polaris (i.e. GCN 1.4 CUs, and not Vega NCUs, mainly) is power consumption: the Radeon RX 470D, with 28 Polaris CUs, has a Total Board Power of 150W at 1266MHz.

The e9550 is a full 32 polaris CU part with a total board power of "up to" 95w at 1.12ghz So a well binned chip with 16 polaris CU's (an RX560) could quite easily be in the 40w range.

cdawallThe e9550 is a full 32 polaris CU part with a total board power of "up to" 95w at 1.12ghz So a well binned chip with 16 polaris CU's (an RX560) could quite easily be in the 40w range.

True, but that is also an embedded part destined for multi-display video and digital signage, and would probably never see a sustained 100% GPU load. While the results of binning in this case are indeed impressive, I can't help but think it'd power throttle if faced with a truly taxing 3D workload. Also, I doubt they'd be able to produce sufficient amounts to satisfy Intel's needs. Not to mention the question of what to do with the more power hungry bins of this semi-custom chip? Give them to Intel for free, to use somewhere else?

Polaris is just as power efficient as vega, so the whole tdp thing isnt really valid

ValantarTrue, but that is also an embedded part destined for multi-display video and digital signage, and would probably never see a sustained 100% GPU load. While the results of binning in this case are indeed impressive, I can't help but think it'd power throttle if faced with a truly taxing 3D workload. Also, I doubt they'd be able to produce sufficient amounts to satisfy Intel's needs. Not to mention the question of what to do with the more power hungry bins of this semi-custom chip? Give them to Intel for free, to use somewhere else?

You mean a processing product wouldn't run at absolute maximum speeds 24/7? I take it you haven't used a processor or gpu with "turbo" before?

T4C FantasyPolaris is just as power efficient as vega, so the whole tdp thing isnt really valid

Really? Polaris hits 185W at 1340 MHz with 36 CUs (RX 580). Vega hits 210W at similar clocks (1410MHz boost, but that's not always sustainable) with 56 CUs. 25W for 20 CUs? Even subtracting the HBM advantage of ~40W (landing the 580 at ~145W, and the Vega 56 at ~205), that's noticeably less power per CU. Not accounting for VRM losses (which should be roughly linear, and thus safely ignored) and uncore (which is pretty much impossible to estimate), that's a difference of ~10% (3.66W/CU vs. 4.03W/CU). Not a massive difference, but definitely noticeable. Of course, the RX 580 is clocked around Polaris' upper limit, which skews the numbers, but every miner out there will tell you Vega scales very well for power when underclocked and undervolted, so the advantage should stick around when slowing down as well, even if it shrinks a bit.

cdawallYou mean a processing product wouldn't run at absolute maximum speeds 24/7? I take it you haven't used a processor or gpu with "turbo" before?

Clock speed isn't the only determinant of power draw, and remember, manufacturers are free to define specs however they like. "TDP" and "base clock" doesn't necessarily mean the same thing in the embedded/digital signage world as in consumer gaming - which would make sense, as gaming PCs and digital signage boxes run very different workloads in very different environments. If "base clock" meant "the GPU will never clock below this limit EVER under load", GPUs wouldn't power throttle in FurMark. While the e9550 is no doubt heavily binned, I sincerely doubt it could sustain its base clock under a 100% 3D load such as any modern gaming load. Of course, I might be wrong. But demanding 3D is not what those cards are used or made for, and thus not specced for.

ValantarReally? Polaris hits 185W at 1340 MHz with 36 CUs (RX 580). Vega hits 210W at similar clocks (1410MHz boost, but that's not always sustainable) with 56 CUs. 25W for 20 CUs? Even subtracting the HBM advantage of ~40W (landing the 580 at ~145W, and the Vega 56 at ~205), that's noticeably less power per CU. Not accounting for VRM losses (which should be roughly linear, and thus safely ignored) and uncore (which is pretty much impossible to estimate), that's a difference of ~10% (3.66W/CU vs. 4.03W/CU). Not a massive difference, but definitely noticeable. Of course, the RX 580 is clocked around Polaris' upper limit, which skews the numbers, but every miner out there will tell you Vega scales very well for power when underclocked and undervolted, so the advantage should stick around when slowing down as well, even if it shrinks a bit.


Clock speed isn't the only determinant of power draw, and remember, manufacturers are free to define specs however they like. "TDP" and "base clock" doesn't necessarily mean the same thing in the embedded/digital signage world as in consumer gaming - which would make sense, as gaming PCs and digital signage boxes run very different workloads in very different environments. If "base clock" meant "the GPU will never clock below this limit EVER under load", GPUs wouldn't power throttle in FurMark. While the e9550 is no doubt heavily binned, I sincerely doubt it could sustain its base clock under a 100% 3D load such as any modern gaming load. Of course, I might be wrong. But demanding 3D is not what those cards are used or made for, and thus not specced for.

Polaris is designed specifically to scale cores with power, where Vega is designed to scale clocks.

at 1.1ghz they would be about the same

ValantarReally? Polaris hits 185W at 1340 MHz with 36 CUs (RX 580). Vega hits 210W at similar clocks (1410MHz boost, but that's not always sustainable) with 56 CUs. 25W for 20 CUs? Even subtracting the HBM advantage of ~40W (landing the 580 at ~145W, and the Vega 56 at ~205), that's noticeably less power per CU. Not accounting for VRM losses (which should be roughly linear, and thus safely ignored) and uncore (which is pretty much impossible to estimate), that's a difference of ~10% (3.66W/CU vs. 4.03W/CU). Not a massive difference, but definitely noticeable. Of course, the RX 580 is clocked around Polaris' upper limit, which skews the numbers, but every miner out there will tell you Vega scales very well for power when underclocked and undervolted, so the advantage should stick around when slowing down as well, even if it shrinks a bit.


Clock speed isn't the only determinant of power draw, and remember, manufacturers are free to define specs however they like. "TDP" and "base clock" doesn't necessarily mean the same thing in the embedded/digital signage world as in consumer gaming - which would make sense, as gaming PCs and digital signage boxes run very different workloads in very different environments. If "base clock" meant "the GPU will never clock below this limit EVER under load", GPUs wouldn't power throttle in FurMark. While the e9550 is no doubt heavily binned, I sincerely doubt it could sustain its base clock under a 100% 3D load such as any modern gaming load. Of course, I might be wrong. But demanding 3D is not what those cards are used or made for, and thus not specced for.

Considering I mine with Polaris based cards I can tell you at the wall total board power can quite easily maintain 1100mhz under 100w.