Discussion AMD SoC Halo series GPU discussion

[DaaQ]

If a handheld becomes too large for you, then you can simply like.. not buy it.


In the meanwhile, I'm happily rocking one of these:

View attachment 127258

And it's freakin' awesome.

What is that model? A steam deck? MSI, ROG, something else?

 

[igor_kavinski]

What is that model? A steam deck? MSI, ROG, something else?

https://www.lenovo.com/ae/en/legiongo/

 

[Bryo4321]

https://www.lenovo.com/ae/en/legiongo/

Love mine too, played half life and Tony hawk on a flight earlier this week. I like it more than the steam deck honestly. It doubles as a tablet in a pinch and the kickstand rocks. Still has a touchpad unlike the ally, the screen is gorgeous, and double USB4 ports makes it so you don’t need a dock for travel.

 

[marees]

Leaks for the highly anticipated GPD WIN 5 have just surfaced

GPD WIN 5 – Everything We Know So Far - GPD | PC Gaming Handhelds & Mini Laptops

The GPD WIN 5 leaks are here! Discover the AMD-powered handheld with an iGPU so fast, it could make some external GPUs obsolete. Details inside.

gpdstore.net

https://twitter.com/x/status/1948322504516223042

 

[aigomorla]

dude i find it a mission to avoid companies that try to make claims like that.

There is no way a iGPU can be "Faster" then a dedicated GPU.. Impossible.
Your talking about VRAM on the higher tier ones which have the same amount of system RAM on the entire unit like a Intel 770 or AMD 9600 card with 16GB of VRAM.

Also GDDR and DDR are two completely different things.

Its very prediatory, and only applies when your probably looking at the 5050 RTX which we all know is LOLZ.
But the moment u step up to a 5060 / 9060 / B580 even, it will smoke any iGPU u throw at it.

 

[igor_kavinski]

But the moment u step up to a 5060 / 9600 / B580 even, it will smoke any iGPU u throw at it.

Until VRAM limits are hit. So if a game actually benefits from more than 16GB RAM, it will suffer on these dGPUs and run smoother on the 64GB Strix Halo. Ditto for 8GB dGPUs vs. 32GB Strix Halo.

 

[aigomorla]

Until VRAM limits are hit. So if a game actually benefits from more than 16GB RAM, it will suffer on these dGPUs and run smoother on the 64GB Strix Halo. Ditto for 8GB dGPUs vs. 32GB Strix Halo.

Thats another thing tho.

Depending on the resolution, and how poorly the engine is coded, 16GB is enough to run any game you throw at it.
I would think under most conditions, even 12GB of VRAM is enough for almost any game at 1440p.

Its the 8GB's which suffer a lot, because i think 8GB was intended for the 1080p segement.

Also LPDDR5X which the strix has is still desktop memory, vs GDDR6.
And we know what happened when Nvidia tried to pull the 1050 a long time ago on DDR to save money and cost of GDDR.
It turned the card into garbage.
Lets also not get into HBM memory which made the AMD Fury so extremely powerful that 99.999% of Ethereum miners wanted that card.

Now im not saying igpu's are all bad... im just saying, i don't like it when a company claims it can out perform a dedicated GPU.
10/10 cases if we ignore budget on both systems, the iGPU will lose, and by MAGNITUDES. Ie.. again ignoring budget throw a 9070 XT against any of the hand helds and see by how many MAGNITUDES it loses.

Its very predatory because most people will think all hand helds are like nintendo switches or even the Legion which has its own special code OS designed to run much more efficiently then windows.

But if you were to put a occulink on these units and put any of the dedicated cards i listed, it would absolutely WRECK the claims they were stating. And this is on a Occulink, and not a full bandwith 8x pci-e dedicated slot.

This is why i have a laptop with a oculink now, and i use the external gpu only when i need it.
These handhelds are a fun toy, i was even interested in getting that new legion, but im waiting for a OLED version like the steam deck to come out.

But will it beat my laptop with oculink? No way. Wont even be close, as i have my old 3090 in a egpu dock.

 

[igor_kavinski]

But if you were to put a occulink on these units and put any of the dedicated cards i listed, it would absolutely WRECK the claims they were stating. And this is on a Occulink, and not a full bandwith 8x pci-e dedicated slot.

Interesting claim and I hope someone tests it out!

 

[aigomorla]

This is what i meant about the legion with steam OS which almost got me buying it.
It just needs a OLED screen.

https://www.tomshardware.com/video-games/handheld-gaming/lenovo-legion-go-s-steamos-review

Just give me a OLED one now...

 

[poke01]

Thats another thing tho.

Depending on the resolution, and how poorly the engine is coded, 16GB is enough to run any game you throw at it.
I would think under most conditions, even 12GB of VRAM is enough for almost any game at 1440p.

Its the 8GB's which suffer a lot, because i think 8GB was intended for the 1080p segement.

Also LPDDR5X which the strix has is still desktop memory, vs GDDR6.
And we know what happened when Nvidia tried to pull the 1050 a long time ago on DDR to save money and cost of GDDR.
It turned the card into garbage.
Lets also not get into HBM memory which made the AMD Fury so extremely powerful that 99.999% of Ethereum miners wanted that card.

Now im not saying igpu's are all bad... im just saying, i don't like it when a company claims it can out perform a dedicated GPU.
10/10 cases if we ignore budget on both systems, the iGPU will lose, and by MAGNITUDES. Ie.. again ignoring budget throw a 9070 XT against any of the hand helds and see by how many MAGNITUDES it loses.

Its very predatory because most people will think all hand helds are like nintendo switches or even the Legion which has its own special code OS designed to run much more efficiently then windows.

But if you were to put a occulink on these units and put any of the dedicated cards i listed, it would absolutely WRECK the claims they were stating. And this is on a Occulink, and not a full bandwith 8x pci-e dedicated slot.

This is why i have a laptop with a oculink now, and i use the external gpu only when i need it.
These handhelds are a fun toy, i was even interested in getting that new legion, but im waiting for a OLED version like the steam deck to come out.

But will it beat my laptop with oculink? No way. Wont even be close, as i have my old 3090 in a egpu dock.

Also in most situations the iGPU shares power either CPU unless it’s on a mini-pc where it can go all out.

The most limiting factor for iGPUs is GPU clock speed, memory bandwidth and clocks. dGPUs can also go above 200 watts.

But the extra shared RAM is handy for some applications, not gaming though

 

[ToTTenTranz]

dude i find it a mission to avoid companies that try to make claims like that.

There is no way a iGPU can be "Faster" then a dedicated GPU.. Impossible.

AMD Ryzen AI Max+ 395 Analysis - Strix Halo to rival Apple M4 Pro/Max with 16 Zen 5 cores and iGPU on par with RTX 4070 Laptop

The new Ryzen AI Max+ 395 is AMD's latest high-end mobile processor. With up to 16 Zen 5 cores, a powerful Radeon GPU, fast NPU and up to 128 GB RAM, the Ryzen AI Max+ is supposed to be the ideal companion for gaming, content creation, and AI development.

www.notebookcheck.net

STX Halo uses an iGPU with 32MB MALL, 40 CUs RDNA3.5, RAM is 256bit LPDDR5X and the IOD is built on N3E. Power management on the STX Halo is actually spectacular, as the thing can work decently even at sub-10W.

Those laptop 4060 and 4070 discrete GPUs have a 128bit bus GDDR6 and they're all limited to 8GB VRAM meaning it's a stutterfest in many of the most recent games.

 

[aigomorla]

Those laptop 4060 and 4070

are you really trying to pull benchmarks from laptop gpu's in comparision to a full bloated desktop gpu?

When people say dGPU 99.9% of the time they mean the real GPU and not a laptop varient.

Did you entirely miss me mentioning Oculink which allows said dGPU to be attached to a Laptop via 4x PCI-E bus, which is en entirely different thing then a laptop GPU.

Especially if you over scale it, with a dGPU /w 16GB of ram, like a Intel B770 / 9060 XT or even a 5070 12GB even.

 

[poke01]

When taking mobile, dGPU means a laptop GPU that is separate from the CPU.

 

[ToTTenTranz]

are you really trying to pull benchmarks from laptop gpu's in comparision to a full bloated desktop gpu?

It's pretty obvious that me and everyone else are talking about laptop dGPUs when making comparisons with Strix Halo, because the latter is a laptop chip.

When people say dGPU 99.9% of the time they mean the real GPU and not a laptop varient.

Laptop GPUs are real GPUs.

Did you entirely miss me mentioning Oculink which allows said dGPU to be attached to a Laptop via 4x PCI-E bus, which is en entirely different thing then a laptop GPU.

I didn't miss, but I don't get how a niche (oculink users) within a niche (eGPU users) should matter when comparing the Strix Halo iGPU to discrete laptopt GPUs.

 

[Rigg]

Leaks for the highly anticipated GPD WIN 5 have just surfaced

GPD WIN 5 – Everything We Know So Far - GPD | PC Gaming Handhelds & Mini Laptops

The GPD WIN 5 leaks are here! Discover the AMD-powered handheld with an iGPU so fast, it could make some external GPUs obsolete. Details inside.

gpdstore.net

https://twitter.com/x/status/1948322504516223042

 

[aigomorla]

It's pretty obvious that me and everyone else are talking about laptop dGPUs when making comparisons with Strix Halo, because the latter is a laptop chip


Laptop GPUs are real GPUs.


I didn't miss, but I don't get how a niche (oculink users) within a niche (eGPU users) should matter when comparing the Strix Halo iGPU to discrete laptopt GPUs.

No.. when most people hear dGPU they think this, im pretty sure:


Before putting something like that on a laptop or even a mini was impossible physically... because ur obviously trying to fit a square in a triangle. But we started getting things like Occulink on the newer minis... and now dGPU became this:


This is all done though a single cable and port which most newer laptops and higher end mini's have.



And they don't need to even get that big... infact they also come tiny and all in one package like this:


So no.. when someone says dGPU... this is the last thing i think of unless you want to troll the audience into thinking there is no version of what i listed above:


Because as i said comparing that thing above to a eGPU of today is extemely shaddy as hell, and its not physically impossible to put a Occulink in these hand helds for when they are are plugged in for even greater gaming performance, since they can fit on ultra portable laptops.


Lastly not to sound condensending to people who buy "gaming" laptops, but i honestly think they are the biggest pieces of junk you can buy.
Why? because they can't do well in anything.

Because of the gpu they draw stupid amount of power, require massive cooling for the gpu, and a large battery making them obnoxiously large and heavy. This is why when anyone ever asks me for advice i always push the ultra portable route with a oculink now, so when they are not playing games, they have the battery life + it being thin and light, and when they are docked at home or at whatever location they are, and want to play, they can plug themselves in and enjoy a real dGPU.


Once again, i am not completely saying these hand helds are bad, i am saying its bad marketing and predatory wording.
I am also very close in in getting the new legion go 2 which is reported to come with a OLED:

Lenovo Legion Go 2 First Look: Is the Upgrade Worth the Hype?

Picture this. You’ve finally gotten your hands on Lenovo’s new shiny handheld gaming PC, the Legion Go 2 , and within minutes, you’re ready to show it off to the world (or at least to your co-op squad on Discord).

www.yardbarker.com

Why? because i can't really lay back in my couch or in the car with autopilot and play games with a laptop + eGPU. So yes they do have its nitche, but it will never replace a real dGPU / Gaming setup, so don't even cross that line.

But i bet you in 10/10 games my laptop with my oculink RTX 3090 will absolutely slaughter the living hell out of it in any game you throw at it. To make matters even more depressing, it also has Intel processor on top, which should speak for itself.

 

[DavidC1]

Those laptop 4060 and 4070 discrete GPUs have a 128bit bus GDDR6 and they're all limited to 8GB VRAM meaning it's a stutterfest in many of the most recent games.

Nitpicking but important. The RTX 4060/70 runs at 16GT for the memory meaning it's same bandwidth as Strix Halo at 256GB/s despite the 128-bit width.

Lastly not to sound condensending to people who buy "gaming" laptops, but i honestly think they are the biggest pieces of junk you can buy.
Why? because they can't do well in anything.

Because of the gpu they draw stupid amount of power, require massive cooling for the gpu, and a large battery making them obnoxiously large and heavy. This is why when anyone ever asks me for advice i always push the ultra portable route with a oculink now, so when they are not playing games, they have the battery life + it being thin and light, and when they are docked at home or at whatever location they are, and want to play, they can plug themselves in and enjoy a real dGPU.

I agree with gaming laptops. Even 6 year old computers have enough CPU to run Strix Halo level GPU fine. That kind of performance could be had for $249 on ARC B580, or Radeon 9060XT which is still in the $300 range.

It's both a waste of money and e-waste as you have no upgrade options with them, unlike a desktop GPU or even the oculink suggestion. Guess it's acceptable if all you do is borrow, or are in the 5 percentile.

 

[ToTTenTranz]

No.. when most people hear dGPU they think this, im pretty sure:
(...)
So no.. when someone says dGPU... this is the last thing i think of unless you want to troll the audience into thinking there is no version of what i listed above:

No one's trolling anyone.. We're talking about Strix Halo, a chip that according to AMD is made for laptops:



So it's pretty obvious to the majority of people that if we're comparing Strix Halo to dGPUs, we're comparing to laptop dGPUs and not desktop dGPUs that consume more power than 6x Strix Halo APUs combined.

And they don't need to even get that big... infact they also come tiny and all in one package like this:

I know you can get small-ish eGPUs. I use an eGPU box myself with a 3060 to pair with my Legion Go.
But I'm also aware it's a niche market, and even moreso the people with Oculink solutions.

Nitpicking but important. The RTX 4060/70 runs at 16GT for the memory meaning it's same bandwidth as Strix Halo at 256GB/s despite the 128-bit width.

Big diference is the 4060/70 laptop have 256GB/s for the GPU alone, whereas Strix Halo's iGPU needs to share the same 256GB/s with the CPU.
Not to mention the big power budget / power efficiency difference between a ~60W STX Halo and a ~25W Intel/AMD CPU + >90W RTX4060/70.

 

[fastandfurious6]

@aigomorla your sentiment is not invalid but it's outdated

it's true that before Halo, igpus were always anemic and couldn't compare with actual desktop gpu

however that's not the case anymore

basically, Halo igpu (8060S) = 2080super desktop = 5060m laptop

10k timespy = 120fps 1080p every game, 60fps 1400p every game



Halo igpu is the king of midrange and there's some (rare) $1000 halo (9955hx) laptops now

then you have:

5080 laptop = 5070 desktop !!!
5070ti laptop > 3080 desktop
etc

and silent fans + 4k 60fps is a thing on mobile now, 5080 laptop can do that

(just 3dmark scores on nbc comparing table but gaming fps are similar. throw in 9955hx3D cache and you have fluid gaming 4k 60fps silent no stutters. what else you want 🤣)

(also oculink etc really not worth it anymore since new rtx5000 gen, next gen it will even be negative also rdna5 mobile etc)

 

[marees]

How To Run OpenAI’s GPT-OSS 20B and 120B Models on AMD Ryzen™ AI Processors​

OpenAI has just released two new AI models – gpt‑oss‑20b and gpt‑oss-120b – which are the first open‑weight models from the firm since GPT‑2.

PC spec requirements for both gpt‑oss‑20b – the more restrained model packing 21 billion parameters – and gpt‑oss-120b, which offers 117 billion parameters. The latter is designed for data center use, but it will run on a high-end PC, whereas gpt‑oss‑20b is the model designed specifically for consumer devices.

These models can be downloaded from Hugging Face (here's gpt‑oss‑20b and here’s gpt‑oss-120b) under the Apache 2.0 license, or for the merely curious, there's an online demo you can check out (no download necessary).


you can run gpt-oss-20b on any laptop or PC that has 16GB of system memory (or 16GB of video RAM, or a combo of both). However, it's very much a case of the more, the merrier – or faster, rather. The model might chug along with that bare minimum of 16GB, and ideally, you'll want a bit more on tap.


It's the same overall deal with the beefier gpt-oss-120b model, except as you might guess, you need a lot more memory. Officially, this means 80GB

AMD's recommendation in this case, CPU-wise, is for its top-of-the-range Ryzen AI Max+ 395 processor coupled with 128GB of system RAM (and 96GB of that allocated as Variable Graphics Memory). speeds of up to 30 tokens per second

Want to run OpenAI's new AI models on your laptop or phone? Here's what you'll need and how to do it

Powerful on-device AI can be yours – if your PC can handle it

www.techradar.com

Loading…

www.amd.com

 

[poke01]

How To Run OpenAI’s GPT-OSS 20B and 120B Models on AMD Ryzen™ AI Processors​

OpenAI has just released two new AI models – gpt‑oss‑20b and gpt‑oss-120b – which are the first open‑weight models from the firm since GPT‑2.

PC spec requirements for both gpt‑oss‑20b – the more restrained model packing 21 billion parameters – and gpt‑oss-120b, which offers 117 billion parameters. The latter is designed for data center use, but it will run on a high-end PC, whereas gpt‑oss‑20b is the model designed specifically for consumer devices.

These models can be downloaded from Hugging Face (here's gpt‑oss‑20b and here’s gpt‑oss-120b) under the Apache 2.0 license, or for the merely curious, there's an online demo you can check out (no download necessary).


you can run gpt-oss-20b on any laptop or PC that has 16GB of system memory (or 16GB of video RAM, or a combo of both). However, it's very much a case of the more, the merrier – or faster, rather. The model might chug along with that bare minimum of 16GB, and ideally, you'll want a bit more on tap.


It's the same overall deal with the beefier gpt-oss-120b model, except as you might guess, you need a lot more memory. Officially, this means 80GB

AMD's recommendation in this case, CPU-wise, is for its top-of-the-range Ryzen AI Max+ 395 processor coupled with 128GB of system RAM (and 96GB of that allocated as Variable Graphics Memory). speeds of up to 30 tokens per second

Want to run OpenAI's new AI models on your laptop or phone? Here's what you'll need and how to do it

Powerful on-device AI can be yours – if your PC can handle it

www.techradar.com

Loading…

www.amd.com

View attachment 128395

The fact that AMD is leading Windows local AI for large LLMs is good.

Nvidia missed the opportunity.

 

[adroc_thurston]

Nvidia missed the opportunity.

It's a meme.
No one actual uses local LLMs since they lack the capability to do like, anything useful.

ML is firmly a big iron market (for better or for worse).

 

[Darkmont]

It's a meme.
No one actual uses local LLMs since they lack the capability to do like, anything useful.

ML is firmly a big iron market (for better or for worse).

The membw gruel that client gets served is laughable for a workload that does almost nothing but chug membw.

 

[adroc_thurston]

The membw gruel that client gets served is laughable for a workload that does almost nothing but chug membw.

LPDDR6-cels slurping on their hopium barrels as we speak.

 

[MS_AT]

Seems they are indeed working on the frameworks

I’m actually re-running some tests now atm btw - the latest Vulkan (amdvlk 2025.Q2) and ROCm (TheRock 7.0 nightly, rocWMMA HEAD) have shown some pretty decent perf improvements just in the past month or so. I do wonder if any media reviewers will get this testing right, as there are easily 2-4X differences based on backend, compile/runtime flags, and with kernel/driver configs. (Of course, all my testing is for Linux, I have no idea how things are on the Windows side).

https://community.frame.work/t/desktop-reviews/73250/4