News AMD's Strix Halo being tested with 128GB RAM — shipping records reveal more about extreme 120W APU

[Admin]

AMD's upcoming Strix Halo APU, never confirmed but seen in shipping records and other leaks, has been spotted again in trade data: now with 128GB of RAM and 120W TDP.

AMD's Strix Halo being tested with 128GB RAM — shipping records reveal more about extreme 120W APU : Read more

 

[Deleted member 2731765]

The mobile processor is already threatening to be the beefiest APUs AMD has ever produced, mobile or otherwise, and has now been spotted paired with 128GB of onboard RAM.

Which memory are you referring to here ? Those are NOT on-package memory modules.

128 GB memory is definitely not on-board or on-package like Intel's Lunar Lake, but we are most likely looking at soldered designs.

Customs data tracking site NewBid Data revealed a new shipment of Maple Rev.B PCBs shipped on May 27, though tech news outlets have only spotted the manifest today. The "Maple" circuit board assembly has been previously associated with Strix Halo test boards

That's the "Maple DAP" reference evaluation and testing platform. I wouldn't refer it as any PCB or board. Strix HALO will also support the FP11 socket platform, with TDPs ranging from 55W and up to 130W.

 

[Deleted member 2731765]

A very confusingly written article.

The processing and graphics will also share a 32MB L4 "Infinity" cache.

To be clear, that's called MALL cache for the iGPU. There will be an additional 32 MB of MALL cache onboard the IOD that will be eliminating bandwidth bottlenecks for the iGPU.

Strix Halo APU will add a 32MB MALL cache to the SoC. Its function is similar to the current Infinity Cache, which can reduce the use of memory bandwidth and improve GPU performance.

Strix Halo will also reportedly come with a NPU capable of 70 TOPS,

Sorry, it is max up to 60 TOPS. Here is proof from AMD's own slide.

The extreme TDP and RAM are likely for speculative testing to push the APU to its extreme limits and see how it performs when supercharged with huge power draw and huge tracts of RAM. However, it is still possible that these are the top-end specs AMD will ship Strix Halo with.

They are not pushing anything to the limit. These are indeed high-end Strix Halo parts like you mentioned, because HALO was supposed to be the flagship APU lineup from AMD, targeting the ultra enthusiast segment.

Strix Halo will fit within much slimmer machines than a comparable CPU/mobile GPU combo. Many speculate that Strix Halo will be seen inside a handheld gaming PC.

What ? Slimmer machines and handhelds with Strix Halo ? Seems unlikely, and out of equation since these are high-end SKUs. For handheld devices, AMD has a different APU lineup in the pipeline, like e.g "Kraken Point", having a 4 Zen 5 and 4 Zen 5C core configuration.

AMD also had some lower-end parts in the 'Strix Halo' lineup like some rumors pointed out before, but those were later debunked.

 

[TechyIT223]

+1 for the above input.

I want an Asus G14 with this processor. The current chassis while being thin only has a 1.5KG weight, has no problem pushing up to 80w+100W watts to the 8845HS CPU and 4070M GPU.

HALO would make it the perfect high-end laptop.

Having 2 CCDs and 16-cores will certainly help professionals and any CPU-centric workloads or a combination of CPU/GPU work. I just hope it's not the only configuration.

 

[thestryker]

Since this is a test platform that's probably why we're seeing 128GB capacity though with the 256-bit bus it's not impossible that could be a shipping configuration. I don't know enough about OEM memory IC pricing and board pricing to say for sure but it's possible 8x 32-bit packages with lower capacity and ball count end up being a net win.

 

[usertests]

Sorry, it is max up to 60 TOPS. Here is proof from AMD's own slide.

There have been conflicting leaks for the NPU TOPS in Strix Halo, including the 70 one. At this point, who cares? We can be pretty sure that it will be a little higher than Strix Point, like Hawk Point compared to Phoenix, which is simply from higher (default) clock speeds.

 

[Deleted member 2731765]

The TOPS value shouldn't matter much for gamers though anyway. The 70 TOPS value has not been confirmed by AMD yet.

In any case, I just hope these APUs are powerful enough to provide decent discrete GPU level gaming performance, especially the top-end part sporting 40 CUs, without any throttling as well.

There is speculation that the flagship APU might also offer performance equivalent to that of an RTX 4060/70 GPU , though I wouldn't place my bet on this claim yet.

 

[jtuck9]

The TOPS value shouldn't matter much for gamers though anyway. The 70 TOPS value has not been confirmed by AMD yet.

In any case, I just hope these APUs are powerful enough to provide decent discrete GPU level gaming performance, especially the top-end part sporting 40 CUs, without any throttling as well.

There is speculation that the flagship APU might also offer performance equivalent to that of an RTX 4060/70 GPU , though I wouldn't place my bet on this claim yet.

I'm wondering (hoping) that they'll leverage these NPU's to compete with the NVIDIA software suite, upscaling (DisplayPort 2.1 UHBR10 output) etc, allowing them to punch above their weight. Especially considering the direction of the upcoming Windows 11 (12?!) update.

 

[usertests]

In any case, I just hope these APUs are powerful enough to provide decent discrete GPU level gaming performance, especially the top-end part sporting 40 CUs, without any throttling as well.

There is speculation that the flagship APU might also offer performance equivalent to that of an RTX 4060/70 GPU , though I wouldn't place my bet on this claim yet.

I hope the 256-bit memory bus becomes a trend. It seems like a successor to Strix Halo is planned, and in general it would be nice to see AM6 go quad-channel to give desktop APUs the kick they need.

 

[jtuck9]

I hope the 256-bit memory bus becomes a trend. It seems like a successor to Strix Halo is planned, and in general it would be nice to see AM6 go quad-channel to give desktop APUs the kick they need.

It will be interesting to see how the "AI Craze May Have Nerfed AMD’s & Intel’s Upcoming Chips" plays out...

I'd buy a system with one of these chips in regardless, but as I'm looking at cloud gaming and it looks like upscaling using Geforce Now is tied solely to Nvidia hardware, if that's still a thing?! It's something I'll have to factor in going forward.

 

[usertests]

It will be interesting to see how the "AI Craze May Have Nerfed AMD’s & Intel’s Upcoming Chips" plays out...

The blame can be shifted onto Microsoft. Cool claim by uzzi about the cancelled cache, but AMD hasn't shown any obvious interest in unifying (reunifying?) L3/L4 cache for CPU and iGPU. Maybe in a patent somewhere? AMD has always been stingy with L3 cache on mainstream APUs. Remember how Renoir had a whopping 4 MiB per CCX?

There are many directions AMD could go with their mainstream APUs, but we should have come to expect disappointment by now. Strix Halo breaks APUs out of the mainstream and into enthusiast territory with the larger size, 256-bit memory, and Infinity Cache. It can't be put onto AM5 but we could cope with it being soldered to a motherboard.

 

[Lucky_SLS]

This could be the testing platform for the next gen consoles.

 

[Deleted member 2731765]

Well a chiphell member reckons that 128 GB could work with LPDDR5X: so 16 GB memory chips and eight in total make 128 GB.

What's up with AMD releasing PRO variants as well ? 🤔

Ryzen AI 9 HX PRO

Ryzen AI 7 PRO.

 

[KnightShadey]

Which memory are you referring to here ? Those are NOT on-package memory modules.

128 GB memory is definitely not on-board or on-package like Intel's Lunar Lake, but we are most likely looking at soldered designs.

Well on-board usually does refer to being "on the board" ie PCB, so soldered. The terms have been co-opted, but if it is soldered, that would be the classic term for "on-board / onboard". Perhaps they need terms like boat engine "inboard / outboard" too. 😉

People (including reviewers)rarely used the term on-package or on-die as they should because they usually generalize for a generalized audience, It doesn't help that "onboard" is also just a widely used term for 'included'.

In this case I would assume it means soldered, which as a buyer & user I HATE the trend towards due to needing to decide and stick to those decisions at budget time and not when needed and no uograde path for when memory is a dime on a dollar a year or two later. For completely new platforms like LNL an STXH it's even worse because we won't know the sweet spot for a week or two after launch, so either risk making a mistake, or risk missing out on first allocation.

This very 'package' of Strix Halo with 128GB is precisely what I would like to get for our field folks as the alternative for the 7945HX + 4080/4090 (which are disappearing from marketplace as we spend the last of our 2023+ budget) We got 3 Alienwares [2 4090 , 1 4080 [will be mix/matching shipped RAM with more to 64GB]) just this past week as I saw stock vanishing (did get nearly $1K off each though 🤗) I even wanted to get one with the RX7900 to compare but they were all gone through our Dell rep. Should tide us over to Strix Halo.

 

[KnightShadey]

That's the "Maple DAP" reference evaluation and testing platform. I wouldn't refer it as any PCB or board. Strix HALO will also support the FP11 socket platform, with TDPs ranging from 55W and up to 130W.

That there is a DAB likely does fit the description of 'All of the Above" in that it should be a small test PCB with an F11 socketed STXHalo , with soldered RAM & resources around it ( "on-board" ) and with a small DAB IC also soldered beside it with an I/O port (likely USB but could be PCIe).

So in that case it would match the initial description in the article above, even if NOT typical/indicative of the final shipping product... I hope, as again I HATE soldered RAM.

(Using ARM example here because AMD's isn't as easy to cut & paste and more cluttered).

 

[Lucky_SLS]

Pro variants for all those corporate laptop business XD

 

[thestryker]

So in that case it would match the initial description in the article above, even if NOT typical/indicative of the final shipping product... I hope, as again I HATE soldered RAM.

Strix Halo will not have any configuration using SODIMMs due to the 256-bit bus and these being aimed primarily at consumer segments not business. The best you could hope for is a manufacturer doing dual LPCAMM2, but this would also currently be limited to 128GB so I'm not sure any would bother.

 

[KnightShadey]

There have been conflicting leaks for the NPU TOPS in Strix Halo, including the 70 one. At this point, who cares? We can be pretty sure that it will be a little higher than Strix Point, like Hawk Point compared to Phoenix, which is simply from higher (default) clock speeds.

Yeah that confusion went all the way to the AI 9 370 HX LMNOP launch at Computex, with ASUS even showing of long prepared PR material with the 45 TOPs number on it, so I'll believe final numbers when they arrive. Does that mean a late production boost once they knew the intel number of 48, but they couldn't get partner ASUS' marketing to change quick enough, or forgot the initial number was there? 🤷🏻‍♂️

Well some of us care, but we are truly in the minority, and we already know where it could be useful both for office tools, and engineering tools in the field. Definitely we are in the minority, and not just following the hype.
Most people should be fine with the 50 of Strix because they want to do something not necessarily do it in 7.9s vs 9.7 seconds. I'm just hoping we can freely assign it and nit get stuck in an M$ copilot model, so we can use it for low-level tasks (security, office) at the same time as the iGPU crunches something bigger, and if sharing resource that's where the 128GB would be handy of course.

The TOPS value shouldn't matter much for gamers though anyway. The 70 TOPS value has not been confirmed by AMD yet.

Yeah, but we're likely at the same point as when Strix was 40-45, then 45-50 TOPs in the leaks last December. So at this point I'd just say the only certain thing is that it will be more than the 50 TOPs of plain Strix.

https://videocardz.com/newz/amd-nex...-offer-3x-performance-boost-for-generative-ai

 

[KnightShadey]

Strix Halo will not have any configuration using SODIMMs due to the 256-bit bus and these being aimed primarily at consumer segments not business. The best you could hope for is a manufacturer doing dual LPCAMM2, but this would also currently limited to 128GB so I'm not sure any would bother.

I understand, and I am resigned to that fate.... but again a boy can hope. 😉

And I don't want it for more than 128GB... at the moment, just more for the price diff of OEM minimum shipped vs Maximum shipped, and also the confusion everyone has of thinking their customers are Apple customers and don't mind paying $1-2K extra for a RAM bump that also requires a forced SSD bump. When buying a bunch, I'd rather get the base configuration until we know what they need, but not be stuck with an 8/16GB e-reader. 😉

 

[KnightShadey]

What's up with AMD releasing PRO variants as well ? 🤔

Ryzen AI 9 HX PRO

Ryzen AI 7 PRO.

First, the naming convention makes no sense and is reverting to old habits, so it's an AI 9 365HX but also an AI 7 Pro when a corp setup? Hopefully just a typo/quick nickname for manufest.

Why not the original one an AI 7 365HX (other than AMD typically making all HX an R9)?
Are they now using AI 9 instead of Ryzen 9xxx to describe the family, because I thought that was the 3xx's job (3- generation, 6/7-performance, 0/5 desk/mobile) ? 🤯 🤬

Adding the HX in between for the 9 Pro seems like a typo, or a dumb marketing scheme turning the TDP into naming convention. 🤷🏻‍♂️ 🤦🏼

As for the Pro plan itself, it's like intel's vPro ... uhh... pro-gram, were it comes with security or VM added bits, and also carries a better support plan (24/36 mths vs typical 12).

Here's the list of features for Ryzen Pro 7xxx series (So does it now become the Ryzen PRO AI xxx series or Ryzen AI Xxxx Pro series ?);
https://www.amd.com/content/dam/amd...n/7000/ryzen-pro-7000-security-whitepaper.pdf


PS, sorry for all the posts in this thread, was travelling the past two days, just catching up, and definitely focused on StxHalo for both work & personal, and combining them all would be a mess. 🥴

 

[usertests]

Well a chiphell member reckons that 128 GB could work with LPDDR5X: so 16 GB memory chips and eight in total make 128 GB.

24 GB LPDDR5X packages are already on the market, so I see nothing stopping a 192 GB configuration, and maybe 256 GB using 32 GB packages by the time a Zen 6 Strix Halo successor comes out.

In this case I would assume it means soldered, which as a buyer & user I HATE the trend towards due to needing to decide and stick to those decisions at budget time and not when needed and no uograde path for when memory is a dime on a dollar a year or two later.

The trend sucks. You have business concerns, but for me, I'll wait until the laptop is discounted 75%. If it comes with 32 GB soldered/on-package (e.g. Lunar Lake), I'm probably not going to mind if the price was right (dirt cheap, "refurbished" but brand new, etc...).

Strix Halo will not have any configuration using SODIMMs due to the 256-bit bus and these being aimed primarily at consumer segments not business.

I would not count out models with SO-DIMM or CAMM. You can find laptops with soldered *and* an empty SO-DIMM slot, and somehow that works. Unless you have a good argument for why this can't happen, I think we could still see Strix Halo with user-upgradeable memory. However, I have seen some roadmaps that explicitly mention LPDDR5X support and not DDR5 for Strix Halo, so that is concerning.

CAMM is so early on the market, I don't think there has even been one example of two CAMMs being stacked yet. It was proposed by JEDEC recently but not implemented by anyone as far as I know. Strix Halo is likely a 2025 product though, so it could exist by the time it launches.

So at this point I'd just say the only certain thing is that it will be more than the 50 TOPs of plain Strix.

Nothing is certain but I will be surprised if it doesn't land at 60 or 70.

Given that clocks are the deciding factor here, how about we overclock it ourselves? Is anyone trying to overclock NPUs yet?

 

[KnightShadey]

Nothing is certain but I will be surprised if it doesn't land at 60 or 70.

I guess 69 is too obvious, Eh? 🤪

Given that clocks are the deciding factor here, how about we overclock it ourselves? Is anyone trying to overclock NPUs yet?

I don't know if anyone has specifically before, but I'm an old fart nowadays and not as interested in OC'ing as I was. But it would be interesting just to see effects and what improves comparing OC to better CPU to more memory, etc to scope out sweet spots.

As for strix's NPU itself has it's own clock domain, and if it dynamically goes up/down depending on workload (or marketing to go from 45-50 a number that should've been concrete when relaying to ASUS) then I'm sure that even if the tools aren't provided up-front then someone will figure out how to do it once they have one in their hands long enough and can figure out.

Just looking at THG's article Skatterbencher OC'ing a 780M https://www.tomshardware.com/pc-com...ter-on-average-with-igpu-and-memory-overclock would've been the perfect time to investigate impact on last generation's more co-dependent design without extrapolating from more generic benchmarks. 🤔

Edit: Looks like the link page has some OC'd iGPU Ai results https://skatterbencher.com/skatterbencher-ai-benchmark-leaderboard/ , so I'm pretty sure he'll add strix as soon as he can get his hands on one and figure it out.

Edit2: just looking through the list, saw his process below the scores, gonna investigate that once back at the office in a coupla weeks... anything to delay getting back to work once back at work. 😉

Also, not sure even with the 780M results he was able to discern NPU/iGPU and CPU loads apart, but will read up later (still winding down from watching SC final).

 

[thestryker]

I would not count out models with SO-DIMM or CAMM. You can find laptops with soldered *and* an empty SO-DIMM slot, and somehow that works. Unless you have a good argument for why this can't happen, I think we could still see Strix Halo with user-upgradeable memory.

It works extremely poorly as it messes with memory channels. There's a reason this isn't done anymore as you basically had the soldered as one channel then SODIMM as the other. So you can end up with mismatched capacity which causes performance problems any time the dual channel capacity is exceeded. In the case of Strix Halo there are 4x 64-bit memory channels which means no reasonable way to have both soldered and expandable.

However, I have seen some roadmaps that explicitly mention LPDDR5X support and not DDR5 for Strix Halo, so that is concerning.

The memory controller would likely support DDR5, but SODIMMs cap out at 5600 (I'm not certain whether or not 6400 with CKD will happen) so this would be a very unlikely configuration. If it does turn out to be only LPDDR5/X then that would rule out SODIMMs but not LPCAMM2.

CAMM is so early on the market, I don't think there has even been one example of two CAMMs being stacked yet. It was proposed by JEDEC recently but not implemented by anyone as far as I know. Strix Halo is likely a 2025 product though, so it could exist by the time it launches.

CAMM2 stacked is two 64-bit channels so this wouldn't be the configuration used rather they would use two memory sockets next to the APU.

 

[eX_Arkangel]

I hope the 256-bit memory bus becomes a trend. It seems like a successor to Strix Halo is planned, and in general it would be nice to see AM6 go quad-channel to give desktop APUs the kick they need.

not only desktop "APU's", but desktop in general, memory BW is the main reason/justification AMD has been using for stagnating the core count for 4 gens (assuming Z5 also tops out at 16) just like intel did for a decade, when the solution is as easy as increasing the channels, but they have refused so far to do so.
As for the APU's AMD has been using mobile discarded dies for desktop, rebranding them with the newest naming scheme currently in use and sold as if they were the latest gen, (2000G being vanilla Z1 not Z1+, 3000G same Z1+ instead of Z2, 4000G launched a year later than Z2 and vendor locked initially to OEM channel) the only outlier in naming were 5000G series, but still used VEGA, and they later outright refused to launch Rembrandt for AM4, something every APU enthusiast was eagerly expecting.

The blame can be shifted onto Microsoft. Cool claim by uzzi about the cancelled cache, but AMD hasn't shown any obvious interest in unifying (reunifying?) L3/L4 cache for CPU and iGPU. Maybe in a patent somewhere? AMD has always been stingy with L3 cache on mainstream APUs. Remember how Renoir had a whopping 4 MiB per CCX?

There are many directions AMD could go with their mainstream APUs, but we should have come to expect disappointment by now. Strix Halo breaks APUs out of the mainstream and into enthusiast territory with the larger size, 256-bit memory, and Infinity Cache. It can't be put onto AM5 but we could cope with it being soldered to a motherboard.

Strix Halo is what everyone initially thought AMD would do for APU's since the release of Zen 2 and RDNA 1, but.... as you said they have avoided releasing it for quite a while, and im a 100% sure it wasn't about technical challenges but a corporate decision to not compete against their RADEON low end tier dGPU and mobile GPU's. Now, when Intel threatens them with MTL and LNL with BM graphics and way more matured drivers its when they suddenly decide to go all out with an APU, i just hope they decide to make an AM5 G version, and stop giving us the ultramobile discarded leftovers dies for the desktop users. (which is the reason for the "stingy L3" in desktop APU's), not renoir but their entire lineup since Raven Ridge has always sported only half L3 of current "desktop" gen variant, add to that severely crippled io capabilities like PCIe lanes and version plus usb/TB/M.2 ports.
Just look at current 8000G series, and compare the entire lineup against regular Z4 desktop in cache-io-PCie ports-etc etc, if you dare go lower than 8600G they are closer to mendocino sku's than to Z4 desktop (a sarcastic exaggeration just in case...)

 

[usertests]

As for the APU's AMD has been using mobile discarded dies for desktop, rebranding them with the newest naming scheme currently in use and sold as if they were the latest gen, (2000G being vanilla Z1 not Z1+, 3000G same Z1+ instead of Z2, 4000G launched a year later than Z2 and vendor locked initially to OEM channel) the only outlier in naming were 5000G series, but still used VEGA, and they later outright refused to launch Rembrandt for AM4, something every APU enthusiast was eagerly expecting.

I can forgive the desktop APU neglect and naming. But if the socket is nerfed by design, there's no fixing the situation for years. And for all we know, AM6 will be the same dual-channel affair, but with a faster DDR6 standard to "justify" it.

I think there might have been technical reasons for not launching Rembrandt desktop APUs on AM4. The big one being that it uses DDR5 memory. When the going gets tough, get it soldered to a mobo. Probably not done cheaply by ASRock but maybe Minisforum and other Chinese companies. That's what I hope to see with Strix Halo, although nobody should kid themselves about the price/perf beating traditional CPU+GPU combos.