AMD's New Strix Halo "Zen 5" Mobile Chips to Feature 40 iGPU CUs

[Darmok N Jalad]

Strix Halo is... not enough "Halo" in CPU performance.

AMD's Strix Halo Zen 5 APU tested in Geekbench AI benchmark — Ryzen AI Max 390 sample falls behind Ryzen 7 7840HS

The Strix Halo sample does not shine in this particular benchmark.

www.tomshardware.com

...in AI performance, running on the CPU only. I dunno, but if I had a SOC with an NPE AND a massive IGP, the last place I'd elect to run an AI workload is the CPU. How CPU-only AI performance translates to other applications is certainly up for debate, but it seems outside the main purpose of such a product.

 

[TumbleGeorge]

...in AI performance, running on the CPU only. I dunno, but if I had a SOC with an NPE AND a massive IGP, the last place I'd elect to run an AI workload is the CPU. How CPU-only AI performance translates to other applications is certainly up for debate, but it seems outside the main purpose of such a product.

If you have different tests with Strix Halo right now, please share it.

 

[dyonoctis]

This will be a proper 45+W chip that will make workstation laptop great again (pardon the pun). One issue that plagues current laptop design is the insistence on using a dedicated GPU on most machines even if they wouldn't need one. You can't get an AMD Phoenix Point laptop without a discrete GPU included, just like you can't get an Intel HX laptop without one. To makes matters worse hybrid graphics was interesting when it was invented but windows quickly made it such a bad idea, I'm yet to have a laptop where the system works correctly

With such a big graphics component maybe, just maybe, we'll finally get good laptop designs with the best CPUs on the market and with no discrete GPU attached to fulfill sales quotas.

Do you mean dragon range? There are a few OLED laptops without GPU using H chips from AMD.

Dragon range and Intel HX chips are essentially desktop chips using the weakest iGPU, that might be enough for some users, but that might be a tad unbalanced for others. Both AMD and Intel designed those chips with the assumption that they would be paired with a dGPU, the "weaker" laptop CPU have better iGPUs. Gotta hope that Intel tries to make a chip with a big iGPU as well.

Laptop pc are already overbolated when it comes to the number of SKUs and confusing as fuck for people who aren't into tech, IMO, the last thing that they need to add is a premium product lineup that would be worse in some tasks over the thin and light laptops with a bigger iGPU

 

[R0H1T]

The NPU goes in the IO die, which, as I said before, is totally new.

I still think it should be monolithic, let's see how this plays out when Halo's released. The I/O die could be a major(?) negative since this is a relatively low volume, albeit high margin part.

 

[igormp]

I still think it should be monolithic, let's see how this plays out when Halo's released. The I/O die could be a major(?) negative since this is a relatively low volume, albeit high margin part.

Low volume, high part is exactly what makes it more reasonable to be a chiplet design.
Instead of designing and entirely new, huge chip (which would warrant yield issues), you can just worry about the IO Die, the only new part, and reuse the existing CCDs that are already cheap since they're used in the desktop Ryzens and Epycs.

 

[Marcus L]

Low volume, high part is exactly what makes it more reasonable to be a chiplet design.
Instead of designing and entirely new, huge chip (which would warrant yield issues), you can just worry about the IO Die, the only new part, and reuse the existing CCDs that are already cheap since they're used in the desktop Ryzens and Epycs.

If AMD could run 8000MTs on a chiplet based IMC then surely 9000 series would have gotten a better IMC, only on monolithic designs can they increase the memory speed by so much compared to chiplet CPU's

 

[igormp]

If AMD could run 8000MTs on a chiplet based IMC

It's LPDDR5, which has higher latencies. It's also a totally new IOD not used in any product before it.

then surely 9000 series would have gotten a better IMC

They could, but instead they just reused the same IOD from Zen 4. 0 cost for then on that front because they don't really give much of a crap to invest that much on Ryzen desktop.

only on monolithic designs can they increase the memory speed by so much compared to chiplet CPU's

Not really, it's just a matter of doing a proper IO Die that can handle such speeds. Strix Halo's IOD is likely going to be made on a better node process than Zen4/5.
This at least could give us some hope that Zen 6's IOD might be improved as well.

 

[trsttte]

Is it chiplet design? If so, they have always had high idle power consumption so far.

I believe it's a chiplet modular design but not like desktop chiplets with high power consumption but tiles like the 7900xt with a much more efficient interconnect. Time will tell.

Do you mean dragon range? There are a few OLED laptops without GPU using H chips from AMD.

No, I mean Phoenix which is monolithic with a bigger GPU like other mobile chips - it's a traditional mobile chip with a high power limit. Sure there are some designs here and there but they are very rare.

 

[Darmok N Jalad]

If you have different tests with Strix Halo right now, please share it.

Yeah, there aren’t any, but it’s such an odd leak, to run a bench using the worst possible part of the SOC to do it. It’s not even the usual geekbench trip. My guess is that SH will not be the best at anything, but it’s probably going to make for a great all-in-one solution in the right applications.

 

[igormp]

Yeah, there aren’t any, but it’s such an odd leak, to run a bench using the worst possible part of the SOC to do it. It’s not even the usual geekbench trip. My guess is that SH will not be the best at anything, but it’s probably going to make for a great all-in-one solution in the right applications.

It'll have unified memory with a reasonable pool size (up to 96 or 128GB, I guess) and okay-ish bandwidth, this would make it a nice machine to run 70b+ LLMs (as long as you can get your stuff working with ROCm).