News AMD's Radeon 780M integrated graphics hits 3.15GHz in Ryzen 7 8700G overclocking test — 37% faster on average with iGPU and memory overclock

[Admin]

Overclocker SkatterBencher was able to get a 3.15GHz overclock on the Ryzen 7 8700G's 780M integrated graphics.

AMD's Radeon 780M integrated graphics hits 3.15GHz in Ryzen 7 8700G overclocking test — 37% faster on average with iGPU and memory overclock : Read more

 

[Bluoper]

Bit of a goofy ahh chip.
On a more serious note, how do you think the 780m would perform if it had a large amount of 3d v-cache?

 

[jeremyj_83]

Bit of a goofy ahh chip.
On a more serious note, how do you think the 780m would perform if it had a large amount of 3d v-cache?

If it is used like the Infinity Cache on the GPUs it could effectively double the bandwidth from the RAM. That would really help the GPU performance.

"Thanks to the 128MB cache, the framebuffer mostly ends up being cached, which drastically cuts down memory access. AMD says the effective bandwidth of the GDDR6 memory ends up being 119 percent higher than what the raw bandwidth would suggest."
https://www.tomshardware.com/news/the-amd-radeon-rx-6800-xt-and-rx-6800-review

 

[usertests]

Bit of a goofy ahh chip.
On a more serious note, how do you think the 780m would perform if it had a large amount of 3d v-cache?

What jeremy said, except technically 3D V-Cache is inaccessible to iGPU, just like any of AMD's CPU L3 caches. There was that moment where some website reported quadrupled iGPU performance with the 7800X3D but it was just a mistake.

So what we'd want is Infinity Cache to come to mainstream APUs, or even 3D Infinity Cache, or AMD changing how it handles L3 to allow it to be shared between CPU + GPU. No evidence seen that they will do any of that for Strix Point. Not sure about Halo.

Even a measly 16 MiB might be enough to get the ball rolling. That's the amount in Navi 24 (6500 XT and 6400).

 

[jeremyj_83]

What jeremy said, except technically 3D V-Cache is inaccessible to iGPU, just like any of AMD's CPU L3 caches. There was that moment where some website reported quadrupled iGPU performance with the 7800X3D but it was just a mistake.

So what we'd want is Infinity Cache to come to mainstream APUs, or even 3D Infinity Cache, or AMD changing how it handles L3 to allow it to be shared between CPU + GPU. No evidence seen that they will do any of that for Strix Point. Not sure about Halo.

Even a measly 16 MiB might be enough to get the ball rolling. That's the amount in Navi 24 (6500 XT and 6400).

I'd want the 128MB like what the i7-5775c had for eDRAM. However, that would make the chip a lot more expensive.

 

[hotaru251]

so 20% improved performance at the cost of 3x the power consumption & thermals are crazy?

 

[King_V]

so 20% improved performance at the cost of 3x the power consumption & thermals are crazy?

Yeah, that's a mess. I mean, this performance uplift is pretty cool in that it could be accomplished, in the sort of "CAN it be done?" kind of way, but it's not really practical, and absolutely not worth the power-use and thermal management headaches that are the price of that performance gain.

But, of course, the chip wasn't really designed to be pushed like this.

 

[usertests]

I'd want the 128MB like what the i7-5775c had for eDRAM. However, that would make the chip a lot more expensive.

Intel's Adamantine would have brought that back (L4 cache, usable by iGPU) as soon as Meteor Lake. But it has been a no-show so we'll have to wait for Arrow Lake at least. Size could be similar too, I heard 128/512 MiB.

 

[greenreaper]

Note for anyone trying this kind of thing: check that your GFX Curve Optimiser settings are actually set. My experience on the B650E-I is that ASUS's control for GFX CO doesn't work, as shown by a lack of voltage changes - you have to set it in AMD's overclocking menus.

CPU CO works but if setting both you may need to set it on AMD menus anyway to ensure that it doesn't get reset.

 

[Kamen Rider Blade]

This platform wasn't meant to compete with CPU+GPU normal combos.

This APU is optimized for LapTops & tiny NUC's.

That's it, there's only 2x major niches that it works for.

 

[thestryker]

The performance with massively tuned DRAM says everything: the 780m is bandwidth starved and they'd likely be better off switching to a quad channel memory architecture. While this is probably relatively viable for laptops I'm not sure it would be on the desktop side of things.

 

[Notton]

The performance with massively tuned DRAM says everything: the 780m is bandwidth starved and they'd likely be better off switching to a quad channel memory architecture. While this is probably relatively viable for laptops I'm not sure it would be on the desktop side of things.

Yeah, it would only require a completely new socket...
or reuse sTR5 w/ TRX50 with an even more cut down CPU.

 

[TerryLaze]

Intel's Adamantine would have brought that back (L4 cache, usable by iGPU) as soon as Meteor Lake. But it has been a no-show so we'll have to wait for Arrow Lake at least. Size could be similar too, I heard 128/512 MiB.

Since they use a lot of ram on CPU/GPU max for datacenter I doubt they will cut into that business by giving it to the desktop, it's gonna be a long wait for it to show up on desktop.
https://www.intel.com/content/www/us/en/products/docs/processors/max-series/overview.html

 

[palladin9479]

This platform wasn't meant to compete with CPU+GPU normal combos.

This APU is optimized for LapTops & tiny NUC's.

That's it, there's only 2x major niches that it works for.

*Cough* SFF systems.

Seriously SFF platforms usually don't have the internal space nor the power budget for a separate GPU. Silverstone does make a few "large" SFF designs that can fit something but it may or may not go with the decor. And that's the next thing, usually SFF stuff goes in places where you don't want to draw attention to the SFF device, it should be practically invisible. Glowy RGB or fan noise are huge detractors and spoil the mood. Think console only more general purpose.

Back when I was as single bachelor living in a high rise apartment overseas, my HTPC was practically invisible and it was setup that all I had to do was hit a button on the keyboard for music to start playing. I've tried a bunch of dedicated items, and most work for awhile, but eventually I always end up going back to the a HTPC platform because it can install anything.

Currently I'm using an InWin BP655 which can fit a 200~300W SFF PSU, but they are hard to find nowadays. The OG SFF case is the Mini-Box M350. This case requires a special external PSU called a Pico-PSU which is usually only 150W worth of power, though there is a 192W version.

https://www.mini-box.com/M350-universal-mini-itx-enclosure

12v/12.5A 150W AC-DC Power Adapter

AC-DC 12V, 12.5A/ 150W Switching Power Adapter (110/220V) for higher power applications

www.mini-box.com

APU's are ideal fits for that segment and let you do some really cool stuff without having a blinking, RGB filled monstrosity sitting in the middle of the room.

 

[TerryLaze]

*Cough* SFF systems.

APU's are ideal fits for that segment and let you do some really cool stuff without having a blinking, RGB filled monstrosity sitting in the middle of the room.

But how many people want an SFF, so small that it won't fit a gpu, but then somehow need good GPU performance?!

I mean there was probably a reason that intel stopped making these NUCs where the whole PC was a GPU sized card and it went on a daughterboard that would accept another GPU sized card.
https://www.anandtech.com/show/15720/intel-ghost-canyon-nuc9i9qnx-review

 

[palladin9479]

But how many people want an SFF, so small that it won't fit a gpu, but then somehow need good GPU performance?!

I mean there was probably a reason that intel stopped making these NUCs where the whole PC was a GPU sized card and it went on a daughterboard that would accept another GPU sized card.
https://www.anandtech.com/show/15720/intel-ghost-canyon-nuc9i9qnx-review

Intel stopped making NUC's because they couldn't beat the leader in that segment, Via's Epia platform. Where Intel focused on performance and shiny stuff, Via focused on stability, hardware compatibility and connectivity with generous amounts of GPIO, RS-232, LVDS and being able to operate in extremely harsh environments without needing maintenance. VIA eventually ditched it's x86 Epia's for ARM related products that accomplish the same thing. The M930 is still available though it's really only for older installations that require compatibility with some old x86 software, usually some sort of medical device or factory automation control computer.

VIA EPIA-M930 Mini-ITX Board

Fast-track the development of embedded systems for demanding use cases.

www.viatech.com

As for SFF, there is quite a bit of demand for it, though not amongst the hardcore "MOAR FPS, MOAR LIGHTS, MOAR POWAW MOAR MOAR" crowd. The executive directory doesn't want a loud blinking monstrous computer in his living room, but he does want the ability to do various tasks. He doesn't know or care to build something, so he pays a boutique installer $10,000+ to set the whole thing up for him.

If you are a gamer who only cares about the FPS counter or bragging about system specs, then it's not a product for you.

 

[TerryLaze]

Intel stopped making NUC's because they couldn't beat the leader in that segment, Via's Epia platform. Where Intel focused on performance and shiny stuff, Via focused on stability, hardware compatibility and connectivity with generous amounts of GPIO, RS-232, LVDS and being able to operate in extremely harsh environments without needing maintenance. VIA eventually ditched it's x86 Epia's for ARM related products that accomplish the same thing. The M930 is still available though it's really only for older installations that require compatibility with some old x86 software, usually some sort of medical device or factory automation control computer.

Was intel ever advertising the NUCs as industrial boards?
All I know is that they were extremely client/desktop oriented.

 

[usertests]

Since they use a lot of ram on CPU/GPU max for datacenter I doubt they will cut into that business by giving it to the desktop, it's gonna be a long wait for it to show up on desktop.
https://www.intel.com/content/www/us/en/products/docs/processors/max-series/overview.html

64 GB of High Bandwidth Memory is a different ballpark from 128-512 MiB of Adamantine L4 cache. Adamantine could be Intel's response to AMD's 3D V-Cache (64 MiB on top of 32 MiB standard), except it could be more for the benefit of iGPUs than CPU cores.

I think we'd want to see much larger L4 caches come to consumer products eventually, so large that they could even replace external memory for some use cases, but that was one of the rumors, half a gigabyte maximum to start out. Now that Meteor Lake has launched we are not at the starting line yet.

Intel's Patent Details Meteor Lake's 'Adamantine' L4 Cache

But how many people want an SFF, so small that it won't fit a gpu, but then somehow need good GPU performance?!

Make it cheap enough and I'll want that. I think we've seen barebones 680M/780M boxes around $300.

Strix Halo will be the true test of what people are willing to pay for small form factor, big iGPU performance.

 

[palladin9479]

Was intel ever advertising the NUCs as industrial boards?
All I know is that they were extremely client/desktop oriented.

Yes Intel heavily pushed NUC's for industrial, medical and otherwise small computing platforms. They also did consumer advertising as a way to generate hype. The problem was Intel had little experience designing for the industrial space. Where they pushed performance and "new" features, industrial application designers wanted stability, compatibility and lots of old school IO ports. How often have you heard of RS-485, or CAN bus support? Those are critical requirements for industrial and automotive automation. The Via platforms could operate reliably 24/7/365in 100F+ and -30F environments. Many designs were fanless, so no worries about the board fan jamming.

I remember when Intel started offering NUC's because one of my co-workers came in talking about how awesome he though they were. I then pulled up the designs for my specialized home router / server that was based on Epia and a SFF box. He was dumb struck because he though Intel had invented the form factor when it was Via who came up with Nano-ITX, Pico-ITX and Mini-ITX.

Anyhow just to illustrate what I'm talking about with regards to HTPC stuff. My receiver is an Yamaha RX-v3900 that I purchased many years ago for something like $2000 along with a set of speakers that cost another $1200 or so. The original TV was an LG X-Canvas for $1600 but that eventually died many years later and got replaced by a Samsung 4K SmartTV. So as you can see, price-per-performance is not the metric used for HTPC as a hundred USD here or there is nothing in the grand scheme of things. Instead things like aesthetics and power use (correlated with noise) are more important. The device needs to be discrete while matching the rooms décor and do it's job without being loud or obnoxious.

 

[TerryLaze]

64 GB of High Bandwidth Memory is a different ballpark from 128-512 MiB of Adamantine L4 cache. Adamantine could be Intel's response to AMD's 3D V-Cache (64 MiB on top of 32 MiB standard), except it could be more for the benefit of iGPUs than CPU cores.

I think we'd want to see much larger L4 caches come to consumer products eventually, so large that they could even replace external memory for some use cases, but that was one of the rumors, half a gigabyte maximum to start out. Now that Meteor Lake has launched we are not at the starting line yet.

Intel's Patent Details Meteor Lake's 'Adamantine' L4 Cache

Yeah, if they ever do this this cache will not be usable by the end-user, they call it L4 cache because that's the technical definition of it, but as they describe it it will just be a protected ram drive to make the system boot up faster.
That's all the intel quote talks about.

If it were anything to counter x3d it would be on the compute dies or on top of them, much closer to the cores.
The patent shows it on the SoC tile so that it's immediately available even if only the two LP/soc cores boot up.

 

[palladin9479]

GPU cache memory is not the same as CPU cache memory, nor should they be the same. Try to think of the GPU as a completely separate processor that just happens to share access to the same memory controller as the CPU. Cache is designed to reduce the number of access's to that memory controller and stands between the processor and the memory controller. What AMD would have to do is equip it's iGPU's with it's own 32-128MB of internal graphics memory cache, which might be economically possible in the next few generations of their modular die design.