> I wound up with a dual width, definitely fan-equipped card that wasn't dirt cheap. For some time I've been grumpy about this, and sort of wondering where they went.
The answer is twofold:
- AMD and Nvidia are charging through the nose for small die graphics cards. Mostly because they can.
- They are also clocking them to the stratosphere, in an extremely inefficient power band where dropping performance ~15% would massively cut power.
There are other factors like drastically changing silicon scaling and capital costs, and better IGPs obsoleting really small GPUs, but those are the big ones.
If you are the writer, you should keep an eye out for AMD Strix Halo, which could put your cooling assertions to the test. Technically the Mezzanine GPU cards (and the actual MI300A CPU) you see in servers use CPU-like heatsinks as well.
During AMDs long dark age Intel shipped an iGPU on basically every mainstream desktop and laptop CPU which obliterated the cheapo bottom end of the market and as others have mentioned killed off any remaining GPU makers as their niches got even smaller.
The only remaining markets for low end discrete GPUs I can think of are people who need basic video out for servers or people who have exhausted all display out options for multiple monitors and need a card to add more.
Another limit, now that I look, is the amount of power available to a PCIe card
CPUs can draw 200W or so which already seems extreme for a small rectangle of a few centimeters. But also consider that they run at low voltage, typically 1.1-1.2V or whatever. That means those tiny BGA/wires are carrying 100+ amps, similar to what a car battery delivers on starting. PCIe tracks are much bigger.
I measured that some years ago. It was ~60W or so pulled through the slot for a card with two extra cables running to it. Specifically I measured the current down the wires and subtraction gave me that as a reasonable estimate for the slot. Won't be perfect but pretty sure it's non-zero through the slot.
Don’t think even DDR5 gets anywhere close to gddr6x on dedicated GPUs. Plus gpu side ram tech is due for a refresh soon too - it’s been on gddr6x for a while
OTOH, there might be a niche for a GPU that has a lower bandwidth access to a larger pool of memory, because inference might consume a lot of data and not that much number crunching.
With a wide enough memory bus, it’d be possible to make a CPU and a GPU happy at the same time with the same memory system, but we’d need manage their expectations about how much they can hog the bus.
I'm hoping that Intel's integration of HBM on CPUs trickles down into something that doesn't cost $10k. It puts CPU memory bandwidth in the same ballpark as GPUs. And honestly at 16-32GB, it's enough for some whole systems, without needing DDR5 DIMMs.
The flipside of that is capacity though - consumer GPUs cap out at 24GB, while consumer CPUs can now support up to 192GB, and the delta is similar for professional/enterprise hardware. If iGPUs are scaled up to higher performance tiers then they could have a niche in crunching very large datasets that would have to go out-of-core on a dedicated GPU.
I think the coolest thing in iGPUs is that they have the possibility of becoming almost part of the ISA.
A compiler targeting the M[1|2|3] processors, for instance, can always rely on having the GPU (and NPU, and video engine, and so on) to offload work, the same way we can now almost rely on any x86 having AVX available.
It’d be interesting to see what an OS that could deal with multiple processors of different instruction sets sharing the same memory image would look like from a user’s perspective.
Interesting idea, however in practice: https://pcpartpicker.com/products/cpu/#g=0 (779 products) vs https://pcpartpicker.com/products/cpu/ (1369 products), and this is mostly just desktop/consumer CPUs - servers are much less likely to have one, and I don't even know where to start researching timesharing GPUs on VMs - most people just do passthrough, usually as a part of a dedicated gaming or transcoding build.
One thing I've observed, is that Electron apps all tend to ship libEGL.so, libGLESv2.so, libvulkan.so, etc rather than depending on system Mesa, so the part where the application tries to take advantage of a GPU by any means necessary (if possible) is already happening, it's just it appears to be Chromium-shaped rather than GCC-shaped.
Intel actually had GVT-g which let you do 7 time slices of the iGPU and give them to VMs. They abandoned it after 10th gen however it was never available on any of their dedicated GPUs.
That’s a chicken and egg problem: there is no server software for it because there is no server hardware to run it on and there is no incentive to build the hardware because the software doesn’t exist.
But that’s an x86 thing. The IBM Telum processor is as server-grade as it gets and has a huge fraction of its real estate dedicated to an inference accelerator. Of course, it’s not a GPU because mainframes tend to not have those, but it shows that if there’s a need (real-time fraud detection) someone will build a product.
> It’d be interesting to see what an OS that could deal with multiple processors of different instruction sets sharing the same memory image would look like from a user’s perspective.
I might at some point build machine with a high-end GPU for AI stuff but so far I don't do that and I do not game at all (except on my vintage arcade cab once every blue moon) so I love CPUs with an iGPU.
> These days these cards are mostly extinct
Precisely. It's really hard to find a GPU that is passively cooled and typically GPUs have fans that are ultra-noisy when under load.
I had an AMD 3700X CPU (no iGPU on that one) and I really didn't like it.
So one year ago I built myself a little rig: AMD 7700X CPU, Be Quiet! tower, Noctua fan, 32 GB of DDR5 and that series of CPU does have an iGPU. I put it in eco mode (I think at peak it'll draw 95W instead of 105W: no discernible difference which I tested with, for example, compiling the Linux kernel or compiling Emacs) in the UEFI/BIOS and when I use it it's often running in powersaving mode (for example when I switch to a browser, the CPU automatically goes to powersaving mode but if I switch to Emacs or an IDE, it goes to performance mode).
The thing is exceptionally quiet. When it's compiling on all cores I can slightly hear the CPU's fan (a Noctua one) ramping up and that's how I like it: it's very muffled, very subtle, but I can hear it. When I do something else than compiling, it's inaudible.
If you want a really quiet desktop and don't game, a CPU with an iGPU is the way to go IMO.
Just my 2c. I daily drive a GT710 passively cooled from MSI, it has a HDMI DVI-D and a VGA port. One of the newer GT710 models from ASUS has 4xHDMI.
Displayport quality is miles ahead compared to the above on the same monitor. (I daily drive 2x Dell P2419H)
If you are going for a multi monitor build then the GT710 is a bad choice if you care about quality.
I would suggest something along the lines of a proper GPU which has a silent mode in which fans do not run unless they are under load. IIRC Asus STRIX cards and Zotac cards had the feature and I am sure other vendors also have something similar.
Ofcourse the best option would be to find a motherboard with a ton of displayport connections, but I did not find any!
You used to be able to passively cool a high end card though. I remember having the Accelero S1 cooler, which supported the top of the line GeForce 9800 GTX. Though I think mine was on an overclocked 8600.
True. I'm using an old desktop (2009) as a little home server. The only gripe I had was the worn out GPU fan on the old radeon card. Picked up a fanless Nvidia quadro for $6.99 on ebay. Problem solved.
You can game with the AMD iGPUs. Especially if you are going for 1080p. AMD iGPUs are basically perfect for an under the TV console style gaming machine. Personally, I'm fine with 1080p on a TV.
Speaking as a Vega 8/4900HS owner, they are really not.
The new 8000/7000 series APUs may be a step up, but its still below the bar for even really old games, at least if you dont want any hitching at 60FPS.
Discrete GPUs have a thermal advantage, because they don't have to share heat-sinking capacity with the CPU.. My Graphics card certainly spews out _A_LOT_ more heated air than the CPU, and on full load, the entire system is practically a space-heater.. Having more physical room for heat-sinking is an advantage in this scenario.
I see the same advantage with power. A PCIe card can more or less ignore the supply from the socket and decide how many various molex plugs they need. There's already dedicated PCIe plugs, and nothing prevents the addition of more, except physical space on the wide edge of the card, the capacity of PSUs and the physical challenges of transporting 99.manynines of that power away as heat.
Good title. Putting the GPU on the CPU is evidently a thermodynamic mistake - they heat each other up and you generally get less surface area for the heatsink. But yes, sharing a heatsink does make geometric sense, and the CPU really might have a great heatsink on it.
I like passive computers, in a pretty with reduced lifespan sort of sense, and a separate GPU is a definite nuisance to cool passively.
I haven't done it myself so I'm mostly speculating but I would expect a separate GPU to be easier to cool passively if it is performing as poorly as an iGPU would
At the end of the day it's all watts and surface area isn't it? Dedicated gpus are intended to perform better and allowed to draw more watts to do so. If you lowered the clock speeds and power limits down to what igpus are allowed to use, you'd then have a much larger surface to dissipate heat from and not share it with the heat from a CPU, so it should be easier to cool.
But id also expect anyone wanting to passively cool their system to also care about things like cost, size constraints, weight, number of parts that can fail etc so in practice just using an iGPU will probably make more sense.
I love iGPUs too, but they are great until they're not...
In my recent experience on an Intel z790/i7-14700K system, getting a signal during boot (e.g., when you want to go into BIOS) was not reliable (usually worked, but sometimes not). When connected to a portable 1280x720 monitor I use when setting up systems, parts of the BIOS screens were a garbled mess. Plugging in a discrete GPU fixed that.
I've also encountered weirdness on the same system where I couldn't disable secure boot until I booted from a discrete GPU (and then after that, on iGPU, I could enable/disable at will). Plus CSM is not an option when booting from certain Intel iGPUs: https://www.asus.com/me-en/support/FAQ/1045467/
Chris may have missed what may be the most probable reason.
Integrated graphics onboard the CPU's eviscerated the sales of "low end graphics cards" to the point that few to no makers are building and selling "low end cards" because there simply is no market left for basic graphics cards.
For the user that is satisfied with their integrated graphics, they have no reason at all to buy a separate card, and for the users who do want a separate card, most want the high power GPU cards, not a card that is little more than the onboard GPU but on a plug in card.
It is the same reason why parallel/serial I/O cards disappeared (when most motherboards began integrating onboard parallel/serial ports, sales of standalone cards would have collapsed) or sound cards (same reason, if onboard is "good enough", no additional sale of a separate card occurs).
I disagree almost wholeheartedly. One of the easiest ways to help cooling is to spread out the heat load, then you need to sink away less thermal power from any specific area to keep the max temperature under whatever limit.
The article seems to imply that iGPUs are more efficient. This is kind of true, but mostly because they have to be, because they need to share the capabilities of their cooler. Most CPUs and GPUs nowadays are clocked for max performance, not efficiency. Pulling them back just a little bit from the red-line and their efficiency skyrockets, but does the user want to pay more for less?
The downside of dGPUs nowadays is that the PCIe form factor was not designed for triple+ slot 10 kg AIBs. Jostle an assembled computer with a RTX 4090 in it and you're probably lucky that it doesn't rip the PCIe connector off the motherboard. This could be made into a cooling argument by pointing out that (consumer) cases just blow the heat around, and don't really exhaust it. (GPUs for servers have passive heatsinks that have front-to-back channels that the jet-engine server fans blow straight through).
I'm almost positive some iGPUs were tested and found that when loaded they caused the CPU portion to throttle. Which is logical because the heat is all right there. Generally it wouldn't be a problem as you wouldn't be loading them at the same time but it could come into play in games.
The right place for gaming and mass-parallel workload GPUs is outside of the case, on a flexible cable carrying 16 channels of PCIe v4 or 5 or whatever and a good ground. Selling the GPU in a case with appropriate power supply and cooling adds relatively little to the cost of a high-end GPU, and the marketing is obviously all about maximum performance unconstrained by your PC's form factor, without having to make any decisions or tradeoffs except money for speed.
> The CPU is probably the best cooled thing in a typical desktop (and it needs to be).
What? Surely it is discrete graphics cards which have the most sophisticated cooling systems, since they consume the most power.
> [...] the amount of power available to a PCIe card, especially one that uses fewer than 16 PCIe lanes; apparently a x4 or x8 card may be limited to 25W total
Maybe through the PCIe connector that is true, but discrete graphics cards have dedicated power connections.
> What? Surely it is discrete graphics cards which have the most sophisticated cooling systems, since they consume the most power.
dGPUs have the "advantage" of a relatively huge die, so per watt they are easier to cool than a small, much mode power dense CPU. Its why they can get away with kind of modest looking heatsinks while burning 350W or whatever.
On his point about multi-monitor setups is there any reason a usbc connection couldnt be used for each monitor? I've used a usbc-hdmi connector for years for a single monitor. Seems like it would just be a matter of having enough ports/dongles to fan-out the connection?
Unfortunately physics is merciless, ANY really interest GPU, MUST have significant consumption, even if it is some moderate level player.
Real numbers, ~15..25W for moderate modern GPU, even without considering other disadvantages, like high RAM bandwidth requirement (and sure, usage).
If you talking about low end CPU, for them 45W TDP usually already a lot, and unfortunately, this is not include iGPU consumption.
Even for top CPUs, where normal 65..150W, additional 25W is significant.
For example, I tested things on now considered old, core i7-7700, which claimed 65W TDP, and it's iGPU, which is around 1/10 of old nvidia 1060, consume ~15W, and these numbers looking plausible, as nvidia 1060 claimed consumption 120W.
What all that mean, real TDP of core i7-7700, when high iGPU usage, far exceeding claimed 65W (easy to calculate, 65W + 15W = 80W).
Unfortunately, I have not equipment to make accurate enough measurements, and not existing any reliable source of such measurements, but from open approximations, and rumors, and experience, I could say, when use iGPU only for trivial office graphics (HN page and some text blogs), and limit CPU freq to official 3.6GHz, looks like it really spent 65W and official boxed cooler is enough.
When I tried to use TurboBoost, also official 4.2GHz, and run intensive for iGPU graphics (or some GPU math), in about minute got throttle. Rumors said, i7 more then doubles it's consumption in TurboBoost, so need much more powerful cooler, considered ~150W, much more than claimed 65W.
But with better iGPU, really added consumption will be 45..80W (considering, iGPU usually using more energy effective tech than discrete GPU). I don't think, you will be happy, even if CPU alone TDP around 250..300W.
Sure, these all disadvantages, are not end of world, but hard to name them advantage..
I've checked. Unfortunately, some things are not easy to found.
So, for i7-7700, iGPU is HD 630, have RAM bandwidth 25.6Gb/s.
Also on later generations, appear Iris Plus Graphics, and Iris Xe Graphics, which doubled or even quadrupled number of Execution Units comparable to top HD line, but RAM bandwidth remain same.
Even exist Intel CPU with AMD iGPU and HBM, and it is really good for iGPU, but around 1060 on some benchmarks.
But if you look on Arc line of GPUs, they are totally other beasts, have RAM bandwidth from 124GB/s (looks like it is iGPU), and consumption from 75W.
- Sure discrete 770 have RAM bandwidth 560GB/s and 225W TDP.
60 comments
[ 3.1 ms ] story [ 94.7 ms ] threadThe answer is twofold:
- AMD and Nvidia are charging through the nose for small die graphics cards. Mostly because they can.
- They are also clocking them to the stratosphere, in an extremely inefficient power band where dropping performance ~15% would massively cut power.
There are other factors like drastically changing silicon scaling and capital costs, and better IGPs obsoleting really small GPUs, but those are the big ones.
If you are the writer, you should keep an eye out for AMD Strix Halo, which could put your cooling assertions to the test. Technically the Mezzanine GPU cards (and the actual MI300A CPU) you see in servers use CPU-like heatsinks as well.
And one of the main groups of holdouts still using desktop computers is presumably PC gamers - who don't buy passively cooled graphics cards.
I'm sure there's some market for low-end desktop PCs - like traditionalists who just don't want a laptop - but I don't think it's a growing market.
The only remaining markets for low end discrete GPUs I can think of are people who need basic video out for servers or people who have exhausted all display out options for multiple monitors and need a card to add more.
PCIe traces however run 10x (or more) the distance, so heat dissipation is more of a concern there.
The formula for resistive wattage dissipation does not include "distance" in it.
The formula is i^2r (i squared times r) where i is current and r is resistance.
So 10 amps flowing through 2 ohms dissipates 200 watts, and it does not matter if the "2 ohms" is 1mm long or 1m long.
So what matters is the net resistance of that short distance, not the distance itself.
Don’t think even DDR5 gets anywhere close to gddr6x on dedicated GPUs. Plus gpu side ram tech is due for a refresh soon too - it’s been on gddr6x for a while
The low end GPU market has become non-existent today. At a minimum, every GPU, even the cheapest, has large amounts of specialized high bandwidth RAM.
We can, on the very high end.
Plus, graphics-oriented RAM is just optimized differently: for bandwidth at the expense of latency.
LPDDR5X on a wide bus (and a healthy LLC) is fine, it just needs implementations beyond Apple.
A compiler targeting the M[1|2|3] processors, for instance, can always rely on having the GPU (and NPU, and video engine, and so on) to offload work, the same way we can now almost rely on any x86 having AVX available.
It’d be interesting to see what an OS that could deal with multiple processors of different instruction sets sharing the same memory image would look like from a user’s perspective.
One thing I've observed, is that Electron apps all tend to ship libEGL.so, libGLESv2.so, libvulkan.so, etc rather than depending on system Mesa, so the part where the application tries to take advantage of a GPU by any means necessary (if possible) is already happening, it's just it appears to be Chromium-shaped rather than GCC-shaped.
That’s a chicken and egg problem: there is no server software for it because there is no server hardware to run it on and there is no incentive to build the hardware because the software doesn’t exist.
But that’s an x86 thing. The IBM Telum processor is as server-grade as it gets and has a huge fraction of its real estate dedicated to an inference accelerator. Of course, it’s not a GPU because mainframes tend to not have those, but it shows that if there’s a need (real-time fraud detection) someone will build a product.
Isn't that AMD's long-standing HSA dream?
> These days these cards are mostly extinct
Precisely. It's really hard to find a GPU that is passively cooled and typically GPUs have fans that are ultra-noisy when under load.
I had an AMD 3700X CPU (no iGPU on that one) and I really didn't like it.
So one year ago I built myself a little rig: AMD 7700X CPU, Be Quiet! tower, Noctua fan, 32 GB of DDR5 and that series of CPU does have an iGPU. I put it in eco mode (I think at peak it'll draw 95W instead of 105W: no discernible difference which I tested with, for example, compiling the Linux kernel or compiling Emacs) in the UEFI/BIOS and when I use it it's often running in powersaving mode (for example when I switch to a browser, the CPU automatically goes to powersaving mode but if I switch to Emacs or an IDE, it goes to performance mode).
The thing is exceptionally quiet. When it's compiling on all cores I can slightly hear the CPU's fan (a Noctua one) ramping up and that's how I like it: it's very muffled, very subtle, but I can hear it. When I do something else than compiling, it's inaudible.
If you want a really quiet desktop and don't game, a CPU with an iGPU is the way to go IMO.
Or a modern CPU and a really old PCIe GPU.
You can, however, buy something like a passively cooled Nvidia GT 710 for really cheap, which I think is what you're getting at.
Displayport quality is miles ahead compared to the above on the same monitor. (I daily drive 2x Dell P2419H)
If you are going for a multi monitor build then the GT710 is a bad choice if you care about quality.
I would suggest something along the lines of a proper GPU which has a silent mode in which fans do not run unless they are under load. IIRC Asus STRIX cards and Zotac cards had the feature and I am sure other vendors also have something similar.
Ofcourse the best option would be to find a motherboard with a ton of displayport connections, but I did not find any!
https://www.quietpc.com/ac-acc-s1rev2
Back when a high end gaming GPU was $300
The new 8000/7000 series APUs may be a step up, but its still below the bar for even really old games, at least if you dont want any hitching at 60FPS.
Discrete GPUs have a thermal advantage, because they don't have to share heat-sinking capacity with the CPU.. My Graphics card certainly spews out _A_LOT_ more heated air than the CPU, and on full load, the entire system is practically a space-heater.. Having more physical room for heat-sinking is an advantage in this scenario.
I see the same advantage with power. A PCIe card can more or less ignore the supply from the socket and decide how many various molex plugs they need. There's already dedicated PCIe plugs, and nothing prevents the addition of more, except physical space on the wide edge of the card, the capacity of PSUs and the physical challenges of transporting 99.manynines of that power away as heat.
I like passive computers, in a pretty with reduced lifespan sort of sense, and a separate GPU is a definite nuisance to cool passively.
At the end of the day it's all watts and surface area isn't it? Dedicated gpus are intended to perform better and allowed to draw more watts to do so. If you lowered the clock speeds and power limits down to what igpus are allowed to use, you'd then have a much larger surface to dissipate heat from and not share it with the heat from a CPU, so it should be easier to cool.
But id also expect anyone wanting to passively cool their system to also care about things like cost, size constraints, weight, number of parts that can fail etc so in practice just using an iGPU will probably make more sense.
In my recent experience on an Intel z790/i7-14700K system, getting a signal during boot (e.g., when you want to go into BIOS) was not reliable (usually worked, but sometimes not). When connected to a portable 1280x720 monitor I use when setting up systems, parts of the BIOS screens were a garbled mess. Plugging in a discrete GPU fixed that.
I've also encountered weirdness on the same system where I couldn't disable secure boot until I booted from a discrete GPU (and then after that, on iGPU, I could enable/disable at will). Plus CSM is not an option when booting from certain Intel iGPUs: https://www.asus.com/me-en/support/FAQ/1045467/
Or even better, with AMD APUs 4 years ago, if you used a discrete Radeon GPU instead, you'd lose sound over HDMI: https://www.reddit.com/r/Amd/comments/d5emmk/agesa_bug_no_so...
Integrated graphics onboard the CPU's eviscerated the sales of "low end graphics cards" to the point that few to no makers are building and selling "low end cards" because there simply is no market left for basic graphics cards.
For the user that is satisfied with their integrated graphics, they have no reason at all to buy a separate card, and for the users who do want a separate card, most want the high power GPU cards, not a card that is little more than the onboard GPU but on a plug in card.
It is the same reason why parallel/serial I/O cards disappeared (when most motherboards began integrating onboard parallel/serial ports, sales of standalone cards would have collapsed) or sound cards (same reason, if onboard is "good enough", no additional sale of a separate card occurs).
The article seems to imply that iGPUs are more efficient. This is kind of true, but mostly because they have to be, because they need to share the capabilities of their cooler. Most CPUs and GPUs nowadays are clocked for max performance, not efficiency. Pulling them back just a little bit from the red-line and their efficiency skyrockets, but does the user want to pay more for less?
The downside of dGPUs nowadays is that the PCIe form factor was not designed for triple+ slot 10 kg AIBs. Jostle an assembled computer with a RTX 4090 in it and you're probably lucky that it doesn't rip the PCIe connector off the motherboard. This could be made into a cooling argument by pointing out that (consumer) cases just blow the heat around, and don't really exhaust it. (GPUs for servers have passive heatsinks that have front-to-back channels that the jet-engine server fans blow straight through).
What? Surely it is discrete graphics cards which have the most sophisticated cooling systems, since they consume the most power.
> [...] the amount of power available to a PCIe card, especially one that uses fewer than 16 PCIe lanes; apparently a x4 or x8 card may be limited to 25W total
Maybe through the PCIe connector that is true, but discrete graphics cards have dedicated power connections.
dGPUs have the "advantage" of a relatively huge die, so per watt they are easier to cool than a small, much mode power dense CPU. Its why they can get away with kind of modest looking heatsinks while burning 350W or whatever.
Unfortunately physics is merciless, ANY really interest GPU, MUST have significant consumption, even if it is some moderate level player. Real numbers, ~15..25W for moderate modern GPU, even without considering other disadvantages, like high RAM bandwidth requirement (and sure, usage).
If you talking about low end CPU, for them 45W TDP usually already a lot, and unfortunately, this is not include iGPU consumption.
Even for top CPUs, where normal 65..150W, additional 25W is significant.
For example, I tested things on now considered old, core i7-7700, which claimed 65W TDP, and it's iGPU, which is around 1/10 of old nvidia 1060, consume ~15W, and these numbers looking plausible, as nvidia 1060 claimed consumption 120W.
What all that mean, real TDP of core i7-7700, when high iGPU usage, far exceeding claimed 65W (easy to calculate, 65W + 15W = 80W).
Unfortunately, I have not equipment to make accurate enough measurements, and not existing any reliable source of such measurements, but from open approximations, and rumors, and experience, I could say, when use iGPU only for trivial office graphics (HN page and some text blogs), and limit CPU freq to official 3.6GHz, looks like it really spent 65W and official boxed cooler is enough.
When I tried to use TurboBoost, also official 4.2GHz, and run intensive for iGPU graphics (or some GPU math), in about minute got throttle. Rumors said, i7 more then doubles it's consumption in TurboBoost, so need much more powerful cooler, considered ~150W, much more than claimed 65W.
But with better iGPU, really added consumption will be 45..80W (considering, iGPU usually using more energy effective tech than discrete GPU). I don't think, you will be happy, even if CPU alone TDP around 250..300W.
Sure, these all disadvantages, are not end of world, but hard to name them advantage..
So, for i7-7700, iGPU is HD 630, have RAM bandwidth 25.6Gb/s.
Also on later generations, appear Iris Plus Graphics, and Iris Xe Graphics, which doubled or even quadrupled number of Execution Units comparable to top HD line, but RAM bandwidth remain same.
Even exist Intel CPU with AMD iGPU and HBM, and it is really good for iGPU, but around 1060 on some benchmarks.
But if you look on Arc line of GPUs, they are totally other beasts, have RAM bandwidth from 124GB/s (looks like it is iGPU), and consumption from 75W. - Sure discrete 770 have RAM bandwidth 560GB/s and 225W TDP.
https://en.wikipedia.org/wiki/List_of_Intel_graphics_process...