News AMD slides claim Strix Halo can beat the RTX 4070 laptop GPU by up to 68% in modern games

[beyondlogic]

Marketing slides show AMD's Radeon 8060S beating the RTX 4070 across 17 games.

AMD slides claim Strix Halo can beat the RTX 4070 laptop GPU by up to 68% in modern games : Read more

amds apus are doing really well they just need to get the software down more on there drivers which is what i am hoping that is seriously what they are doing with 9000 parts. they need to come out swinging and so does intel in the markets of 100-600. least thats my opinion. i know the apus dominate alot of the laptop market.

 

[illrigger]

1. It almost certainly will come to mini-PCs in time. It’s not even out yet.
2. It has quad-channel unified memory, and thus cannot be a drop-in option for AM5 based desktops.
3. It’s equal to a 4070 laptop with a 65w TGP, but that configuration is easily surpassed in gaming by midrange desktop dGPUs.
4. The complexity and large iGPU and extremely fast unified RAM and the fact it can’t drop into AM5 means it would likely be very expensive for a desktop APU, and desktops needing good graphics performance but not willing to use a dGPU and being willing to pay a premium for that is a slim niche, so it’s not AMD’s launch priority. Someone will fill it eventually and people will probably complain it’s a poor value.

On top of that, they have confirmed that Strix Halo is hard wired with only two x4 PCIe slots for NVMe, so it literally can't use a dGPU without it being crippled in the same way as an eGPU. It was designed from the ground up to be a SBC at best. While I am sure it will eventually make its way into an ITX form factor, when it does it won't be overly useful as a desktop replacement or even a small server board. Its best real use case use is in mini PCs where it can get full power and good cooling.

 

[kealii123]

55 Watts probably because they aren't compute bound but limited by DRAM bandwidth. My RTX 4090 stops its fans runing any model that won't fit inside 24GB, because there is so little to do but wait.

I try to simplify things by converting a token to a syllable and making words 2-4 syllables (I'm German, my words are longer ).

A little more than one word per second? Try that on your wife and see what that does to her patience!

As a benchmark it's perhaps not bad, but in practical terms I'd consider it unusable.


Have a closer look at NVlink, not all versions and variants are created equal. For the 3090 it's a little over 100GByte/s, pretty much DRAM speed these days and only two cards max for anything PC.

CUDA code is designed to exploit the Terabyte/s aggregate bandwidth of massive register files. VRAM access is already falling off a cliff, so much so that common subexpression elimination, a typical staple of compiler optimization on CPUs, is essentially reversed, because that would often be slower than recomputing inside registers.

And even with the greatest and latest NVlink switches (7200 Gbyte/s on Hopper), that's not counting latency.

HPC and AI hardware is a little more complex than just putting Lego bricks together. And yeah, I hoped it was much simpler, too. But then I had the opportunity to test and researched a bit deeper.

And now I understand better why prices are the way they are.

Yes, my wife already complains about my verbal communication skills.

HOWEVER, practically, 14b param models are more useful for actual work. I run qwen 2.5 7b on my M1 macbook for copilot autocomplete and it works great. The diminishing returns for larger LLMs isn't something that most people will notice. Its frustrating that we don't even know when the review embargo will be lifted, so we can see what this chip can really do at more realistic sized models, as well as gaming on and off the wall charger.

 

[abufrejoval]

UP TO is doing heavy lifting here, but it is a good showing.

...

As for the high TDP, you have to remember that there are 16 CPU cores in the top part. This is meant to replace a i9-14900HK/Ryzen 9 9955HX and a 4070/5060 at the same time. What is the CPU going to use during gaming tests, surely at least 15-20 Watts?

I don't think gamers are going to find these relevant because the prices will probably be way too high. Even the 8-core, 32 CU model. Maybe it doesn't have to be, idk. But this is the start of something that AMD can improve each generation.

The cost of producing that extra CCD for the 16 core part will be quite different from the premium on the initial sales price. It's part of the magic behind AMD's modular approach, I've heard that even the 1st generation V-cache was only $20 extra production cost: not what they charge on them, especially these days, after improving things.

I bought a Minisforum Mini-ITX board that includes a previous gen Ryzen 9 7945HX with 16 Zen 4 cores for €500, including EU taxes, a few weeks ago. Apart from being the previous generation the main difference is the much smaller IOD with only 2 CUs instead of 40. It's doing 1619 at Cinebench 2024 multi using 100 Watts vs 1936 on my Ryzen 9 7950X3D at 147 Watts, leaving my 5950X behind with 1496 at 150 Watts.

(It's still no good at AI and for gaming you might want to shut down the 2nd CCD: really more of a workstation)

I don't think AMD is selling those CPUs at a loss quite yet, so there is no reason to believe that they couldn't sell these at say the added price of an Intel B580 or €750 total without loosing money... in a year's time.

They are going for fruitish margins, which is fine if you can afford it. But others may want to hold on to their money for a bit, not accounting for crazy tarifs, obviously.

Integrating discrete transistors into chips and discrete chips into chiplets costs a lot of engineering, but pays off in reduced production cost: Strix Halo is no different.

It will win over systems made of discrete components because it's cheaper to produce at scale. Whether that integration offers value as significant as the markup they are trying to charge you is up to you to judge. But boy are they trying hard to make you think that way.

 

[KyaraM]

Depends on what power the Strix used, plust lets not forget we should compare CPU with iGPU to CPU+GPU when it comes to performance/watt.

For me the BS started way before with Nvidia using the same names for their desktop and laptop parts, by hiding the details in the fine print that is a blatant attempt at misleading their customers. Simply put many computer buyers don't read the fine print or don't have the foundation to really understand the details.

That was stated in the article. At SAME power consumption. Which also means that they could have tested at higher TGP, but didn't,

UP TO is doing heavy lifting here, but it is a good showing.

I haven't seen any indication of LPCAMM models, just soldered. I hope we see LPCAMM in the future.

As for the high TDP, you have to remember that there are 16 CPU cores in the top part. This is meant to replace a i9-14900HK/Ryzen 9 9955HX and a 4070/5060 at the same time. What is the CPU going to use during gaming tests, surely at least 15-20 Watts?

I don't think gamers are going to find these relevant because the prices will probably be way too high. Even the 8-core, 32 CU model. Maybe it doesn't have to be, idk. But this is the start of something that AMD can improve each generation.

It's of course not 1:1, but my 7640HS in my gaming laptop can eat up to 54W under full load. Usually around 30W - 35W during gaming, though. No clue how that compares to this chip, though. But that would still leave around 90W for the GPU, at which point the 4070 Mobile approaches its optimum.

1. I think the point of the test, despite using a 65w mobile 4070, is to show how much more efficient this chip is by testing both in a thermally constrained environment, so you might actually get more than 30 min of battery life if you're gaming on it.

2. Comparing an RTX 4090 running a 70b param model is very disingenuous , since the card can't really run something that large. HOWEVER, if I'm reading the results correctly, the AMD iGPU is hitting around 6 tokens per second on that same 70b model, which is actually very impressive.

Does anyone know when the review embargo drops for the new z13? Or maybe that HP laptop version with the same chip?

Some laptops already have a low-power eco mode, though. I can’t tell you how much mine uses in that mode since I don't game on battery (and I have ask why one would even do somethingstupid like that...), but regular stuff like videos etc lasts a couple hours.

 

[usertests]

On top of that, they have confirmed that Strix Halo is hard wired with only two x4 PCIe slots for NVMe, so it literally can't use a dGPU without it being crippled in the same way as an eGPU. It was designed from the ground up to be a SBC at best. While I am sure it will eventually make its way into an ITX form factor, when it does it won't be overly useful as a desktop replacement or even a small server board. Its best real use case use is in mini PCs where it can get full power and good cooling.

If you give me a Max 385/390/395 desktop, I will use that level of graphics performance until it breaks. There's no need to care about a dGPU if it can handle 1080p/1440p as well as it should. RDNA3.5 will probably end up supported by FSR4, so there's your ticket to 4K and higher FPS.

Why not for a server? Not enough storage/PCIe? Depends on the use case, but I can see how that could be a problem.

It's going to have its best showing in gaming laptops (and there's even a tablet), where it is clearly a better idea to use a mega APU than a mobile dGPU. Then AI workstations (mini size) where it can benefit from having a lot of "VRAM". Maybe the models with 64 GB won't be popular since you'll want the full 128 GB for 70B LLMs.

Strix Halo's problem will be the pricing. It is starting out as an expensive, niche, AI-oriented product, with few models on the market. Using it for desktop gaming will be a waste of money when you can just pick up a 7600 XT or better and weaker CPU on your own.

The cost of producing that extra CCD for the 16 core part will be quite different from the premium on the initial sales price. It's part of the magic behind AMD's modular approach, I've heard that even the 1st generation V-cache was only $20 extra production cost: not what they charge on them, especially these days, after improving things.

Recently MLID (and I think some old leak) suggested that the Strix Halo CCDs are not identical to Zen 5 CCDs used in Granite Ridge and Turin. Probably because they are using Infinity Links or something to connect them. If that's true, maybe all of these product lines (including mainstream mobile) will use the same Zen 6 CCD and connection methods. Strix Halo could be considered the beta test for it.

Don't get me wrong, AMD is still making a killing on high yield chiplets.