Why are iGPUs no longer done this way?

[da na]

I work on a good deal of old laptops, and I've noticed that a few particular AMD northbridge chipsets (which contained the iGPU in that era) have a single RAM chip set aside strictly for graphics use.

Spoiler

And to me, this seems like a win-win! GPU does not have to compete for system memory bandwidth, doesn't steal 512mb-2GB of your RAM, and if that one small VRAM chip fills up the GPU also has a range of system memory which it can allocate.

So what's the downside? With integrated graphics being more prevalent and powerful today, why abandon the idea of half-discrete VRAM for them? I could imagine with the northbridge being on the CPU these days, it'd be a little harder to execute, but I would imagine the performance benefits would still be there.

[Eigenvektor]

7 minutes ago, da na said:

So what's the downside?

I would assume cost.

[GuiltySpark_]

Cheaper to integrate it into the CPU not that its more feasible. 

[MrSimplicity]

With the speed of ram these days there's no need to do so. Especially since integrated graphics are not powerful enough to run more than heavier CPU reliant games AKA basic esports titles. As well as those that purchase machines that don't have discrete graphics more than likely aren't gaming on them anyways. So the iGPU doesn't need it's own separate vram. Essentially what the other guy said... it's just a cost saving measure.

[adm0n]

14 minutes ago, da na said:

I work on a good deal of old laptops, and I've noticed that a few particular AMD northbridge chipsets (which contained the iGPU in that era) have a single RAM chip set aside strictly for graphics use.

  Reveal hidden contents

And to me, this seems like a win-win! GPU does not have to compete for system memory bandwidth, doesn't steal 512mb-2GB of your RAM, and if that one small VRAM chip fills up the GPU also has a range of system memory which it can allocate.

So what's the downside? With integrated graphics being more prevalent and powerful today, why abandon the idea of half-discrete VRAM for them? I could imagine with the northbridge being on the CPU these days, it'd be a little harder to execute, but I would imagine the performance benefits would still be there.

There are multiple reasons I'd assume.

Some I can think of are

• Having a dGPU like that, would take up actual PCIe lanes. So you are either stuck with -8 lanes permanently or you need a PCIe switch
• Board space for fitting all the components of an actual GPU
• Cooling and extra VRM
• Extra Cost
• It would likely only be viable for lower end GPUs, so it would only make sense if the whole contraption was somehow cheaper than just buying a motherboard and a GPU. But you'll hardly save anything, especially if you don't want to compromise on the motherboard functionality

[YoungBlade]

Intel actually did one better with this. Unfortunately, it was in the short-lived and oft-forgotten 5th gen. The Core i5 5675C and Core i7 5775C actually have 128MB of RAM on package for the iGPU to use. This solution was so effective that it actually made their iGPU often better than the best AMD APUs of the day. If Intel hadn't charged through the nose for them, or had released a Core i3 5115C or something at $100, they probably could have killed off AMD right then and there in the CPU space. Those APUs were all AMD had going from them at that point.

 

But 5th gen was often slower than the Devil's Canyon refresh chips while costing way more than AMD APUs, meaning there was no market for them as CPUs nor as APUs. But for a brief time, Intel technically had the best APU chips on the market.

[Eigenvektor]

1 minute ago, adm0n said:

Having a dGPU like that, would take up actual PCIe lanes. So you are either stuck with -8 lanes permanently or you need a PCIe switch

That, or they would have to add additional connections on the model with an iGPU, which means they'd have to create a separate design. Which comes back to cost prohibitive.

 

Looking at the reviews of the Ryzen 8xxxG, APUs already tend to be more expensive than some lower end CPU+GPU combinations that outperform them. So the only advantage left is smaller size. And I think AMD also don't want their APUs to become too powerful, or they'd eat into GPU sales.

 

3 minutes ago, YoungBlade said:

The Core i5 5675C and Core i7 5775C actually have 128MB of RAM on package for the iGPU to use.

Thanks, I was trying to remember which CPU it was that had RAM on package

[da na]

9 minutes ago, YoungBlade said:

Intel actually did one better with this. Unfortunately, it was in the short-lived and oft-forgotten 5th gen. The Core i5 5675C and Core i7 5775C actually have 128MB of RAM on package for the iGPU to use. This solution was so effective that it actually made their iGPU often better than the best AMD APUs of the day. If Intel hadn't charged through the nose for them, or had released a Core i3 5115C or something at $100, they probably could have killed off AMD right then and there in the CPU space. Those APUs were all AMD had going from them at that point.

 

But 5th gen was often slower than the Devil's Canyon refresh chips while costing way more than AMD APUs, meaning there was no market for them as CPUs nor as APUs. But for a brief time, Intel technically had the best APU chips on the market.

Fascinating. There've been some GPUs with RAM on-package (many, many years before HBM - this is in like 2002) but had no clue some CPUs did that.

[Agall]

34 minutes ago, da na said:

I work on a good deal of old laptops, and I've noticed that a few particular AMD northbridge chipsets (which contained the iGPU in that era) have a single RAM chip set aside strictly for graphics use.

  Reveal hidden contents

And to me, this seems like a win-win! GPU does not have to compete for system memory bandwidth, doesn't steal 512mb-2GB of your RAM, and if that one small VRAM chip fills up the GPU also has a range of system memory which it can allocate.

So what's the downside? With integrated graphics being more prevalent and powerful today, why abandon the idea of half-discrete VRAM for them? I could imagine with the northbridge being on the CPU these days, it'd be a little harder to execute, but I would imagine the performance benefits would still be there.

Maybe if FM2+ ended up becoming FM3 and potentially FM4, where they could differentiate the motherboard layout to accommodate onboard DDR6 VRAM for an APU, but that's just not how it ended up. We still haven't seen MCM APUs since the 8700G is monolithic just like previous desktop APUs.

 

Maybe we'll get more complexity if we see a TR4 APU, where it could be an MCM+HBM2/3 APU, but they'd probably just solder the substrate than have it socketed on an existing platform.

 

I'm still waiting for MCM Zen4c and RDNA3 on the same substrate. I don't think AM5 is big enough for it though. Maybe we'll see FM3 after all these years with a TR4 like sized die with no RAM slots since the die also has HBM for the CPU and GPU. Very very wishful thinking though.

[porina]

19 minutes ago, YoungBlade said:

Intel actually did one better with this. Unfortunately, it was in the short-lived and oft-forgotten 5th gen. The Core i5 5675C and Core i7 5775C actually have 128MB of RAM on package for the iGPU to use.

It was rated at 50GB/s bandwidth, which is about double that of the standard DDR3-1600 ram of the time. It isn't clear to me if it worked both directions simultaneously, or not like normal ram.

 

Also note it was a victim cache. You don't access it like a separate pool. Recently used data fell into it when evicted from lower level caches. It really speeded up some CPU only workloads, in a broadly similar way to 3D cache.

 

19 minutes ago, YoungBlade said:

If Intel hadn't charged through the nose for them, or had released a Core i3 5115C or something at $100, they probably could have killed off AMD right then and there in the CPU space.

I think they were priced appropriately, comparable to the equivalent Skylake -K offerings. I wasn't up to speed when Haswell was current but I wouldn't be surprised if they were similarly priced too.

 

19 minutes ago, YoungBlade said:

But 5th gen was often slower than the Devil's Canyon refresh chips while costing way more than AMD APUs, meaning there was no market for them as CPUs nor as APUs. But for a brief time, Intel technically had the best APU chips on the market.

They were likely only released by Intel for them to say they released it at all. They weren't regular desktop CPUs but the existing mobile offering shoved onto desktop socket. As much fun as some make of Intel's 14nm, it was the first product on that node. It had problems and was very late. It's prime sales window was a matter of months before Skylake came along - May to August 2015. Even Rocket Lake had longer than that - March to Nov 2021.

 

It was only considered slow because they didn't clock well. From memory they topped out around 4 GHz whereas Haswell refresh was pushing much higher than that.

 

I kept my 5775C system. Some day I'll get around to benching it with modern games. Don't hold your breath.

[Obioban]

It's just a less efficient allocation of resources.

 

Got a bunch chrome tabs open? That VRAM is useless to help you.

 

Running a game that doesn't need much ram (most games)? That extra RAM is sitting idle.

 

Having GPU and CPU share memory lets you allocate the resource as efficiently as possible-- and because you're not paying for two different ram banks, for the same money you should be able to get more in the common pool.

[podkall]

41 minutes ago, da na said:

doesn't steal 512mb-2GB of your RAM, and if that one small VRAM chip fills up the GPU also has a range of system memory which it can allocate.

to them it's more complicated than simply just allocating a portion of RAM for the iGPU instead, and then changing the allocation/reserve for iGPU

[YoungBlade]

24 minutes ago, porina said:

I think they were priced appropriately, comparable to the equivalent Skylake -K offerings. I wasn't up to speed when Haswell was current but I wouldn't be surprised if they were similarly priced too.

They were priced fine as CPUs, but as APUs, they were way too expensive. The then top-of-the-line AMD APU, the A10-7850K, had an MSRP of $190. While the Core i7 5775C was $380 and the Core i5 5675C was $276. So both were just too expensive to compete in the APU market.

 

Had Intel released a Core i3 with the same Iris Pro graphics at $100-150, AMD would have been totally out of a market. This hypothetical i3 CPU would have become the go-to choice for APU builds.

 

Instead, with the way they were priced, they weren't compelling as APUs.

[Agall]

21 minutes ago, podkall said:

to them it's more complicated than simply just allocating a portion of RAM for the iGPU instead, and then changing the allocation/reserve for iGPU

Precisely. With how well the handhelds work with only 16GB of total RAM, I don't think its much of a concern with today's architecture, especially with how fast system RAM is.

[podkall]

14 minutes ago, Agall said:

Precisely. With how well the handhelds work with only 16GB of total RAM, I don't think its much of a concern with today's architecture, especially with how fast system RAM is.

and how lazy some people are, look at brand new videogames and their optimization

[Mark Kaine]

yea gpu on mobo... i don't know why this never caught on... its genius...

 

i have an old intel board with Intel cpu and ATi (or AMD?) graphics *on the board*.

its completely hassle free and runs every dos game at 60 fps no problem! 

 

i honestly don't see why this couldn't be done for windows games too... (no i don't buy the not enough power excuse, they didn't even try!)

 

i blame "voodoo", yep!  

[Alex Atkin UK]

1 hour ago, da na said:

I work on a good deal of old laptops, and I've noticed that a few particular AMD northbridge chipsets (which contained the iGPU in that era) have a single RAM chip set aside strictly for graphics use.

Its actually quite simple - the northbridge was merged into the CPU to improve memory efficiency and power consumption.  On AMD this is basically what the SoC is, although lots of things that were on the southbridge are also in the CPU package now too.

 

The chipsets we have today are technically the southbridge only and they often can be omitted entirely on small form factor PCs as they only provide additional IO on top of what the CPU already has.

[porina]

8 minutes ago, YoungBlade said:

They were priced fine as CPUs, but as APUs, they were way too expensive. The then top-of-the-line AMD APU, the A10-7850K, had an MSRP of $190. While the Core i7 5775C was $380 and the Core i5 5675C was $276. So both were just too expensive to compete in the APU market.

They were CPUs first. That the iGPU performed well or not wasn't a priority. Again, I feel this was a release only to say they released it. They never intended it to ship in volume for the few months it was current before Skylake replaced it. That A10 probably would have performed worse than any i3 of the time in most CPU tasks, maybe more Pentium/Celeron level.

 

I paid £220 for the i5 5675C in Feb. 2016 (after Skylake launched). I got an i3 6100 in August 2016 for £101. Unfortunately my i5 6600k order didn't separate out pricing so I don't know how much that was. My 2nd i7 6700k was £289 in Feb 2016. My first one was from the vendor that didn't break out the pricing by line in e-mail. Bottom line: I don't think Intel could have sold it at i3 pricing without losing significant money. The eDRAM and packaging likely would have dominated.

 

The 5775C I have was bought used so can't be used for price comparison. I paid £200 for it used in November 2018.

[Agall]

14 minutes ago, podkall said:

and how lazy some people are, look at brand new videogames and their optimization

I think things have changed, I'm actually in the process of using Newegg's trade in program to get $1500 for my RTX 4090 that I got on launch and downgrading to like a 4070(S). Cash out while I can as well since I don't find myself gaming as much nor needing that much processing power. I don't play anything intense nor competitive now a days.

[da na]

1 hour ago, Mark Kaine said:

yea gpu on mobo... i don't know why this never caught on... its genius...

 

i have an old intel board with Intel cpu and ATi (or AMD?) graphics *on the board*.

its completely hassle free and runs every dos game at 60 fps no problem! 

 

i honestly don't see why this couldn't be done for windows games too... (no i don't buy the not enough power excuse, they didn't even try!)

 

i blame "voodoo", yep!  

Very common in ye olden days. Higher end multimedia laptops would have a GeForce 7600, GeForce 8300, etc on board instead of Intel graphics. I've got several this way, absolutely great. Even some business laptops did it, Latitude E6400 has an Nvidia NVS 160M.

[8tg]

8 hours ago, da na said:

Very common in ye olden days. Higher end multimedia laptops would have a GeForce 7600, GeForce 8300, etc on board instead of Intel graphics. I've got several this way, absolutely great. Even some business laptops did it, Latitude E6400 has an Nvidia NVS 160M.

I like the boards that had full on gpus integrated into them

pic related, ECS k7s7ag, a socket 462 board with an sis 746 chipset and a full on 64mb sis xabre 200 GPU onboard 

To kinda equate this to modern spec it would be like if there was a full fledged gt 1030 on your motherboard as an igpu. Is it amazing? No, but it does a lot more than conventional onboard graphics 

 

there’s a couple boards like this, there’s also one with a 3dfx voodoo on it I forget the name of 

[Salted Spinach]

What do CPU manufacturers deliver? Just the processor.

 

They are not the vendor for memory chips.

 

They do not own the rights to include VRAM technology in the CPU package.

 

They do not want an additional component in the package that produces heat and gets in the way of the CPU's primary function.

 

If VRAM to be added on the PCB instead of the CPU package, this is the board partner's domain and exposes them to increased costs when integrated graphics is not an advantage for the vast majority of boards being sold.

 

Additional component means more design complexity, more rejects in the production line. Higher costs and lower margin. This regardless of where the VRAM is placed.

 

Unless the manufacturer passes on all the costs to the customer. Who will be resistant because APUs are traditionally a class of product for the budget-conscious.

 

That said, having VRAM on the board could become a thing for very small prebuilts and ultrabooks (and only these type of products) if handheld gaming booms and AMD/Intel/Nvidia push APU graphics to the point where VRAM will add an amount of performance that cannot be ignored.

 

One more thing that people might not realize or are too quick dismiss, is that the main driver for gaming APUs is the handheld scene. And ARM will not be happy if it encroaches on their mobile gaming territory.