I don't understand the name "Strix". It's a name a GPU and motherboard partner of theirs, Asus, uses (used?) for their products. It's impossible for me to read "AMD Strix" and not think of it as some ASU's GPU with an AMD chip in it, or some motherboard for AMD sockets.
Aren't there enough syllables out there to invent a combination which doesn't collide with your own board partners?
I don't think AMD really uses the name "Strix Halo" to market it to a large audience, it's just an internal codename. Just two other recent internal names are "Hawk Point" and "Dragon Range" internally, where Hawk and Dragon are names that MSI and PowerColor use to market GPUs as as well. Heck, PowerColor even exclusively sells AMD cards under the "Red Dragon" name!
AMD's marketing names for especially their mobile chips are just so deliberately confusing that it makes way more sense for press and enthusiasts to keep referring to it by its internal code name than whatever letter/number/AI nonsense AMD's marketing department comes up with.
For me the question is: what does this mean for future of desktop CPUs? High bandwidth unified memory seems very compelling for many applications, but the GPU doesn't have as much juice as a separate unit. Are we going to see more these supposedly laptop APUs finding their way into desktops, and essentially a bifurcation of desktops into APUs and discrete CPU/GPUs? Or will desktop CPUs also migrate to becoming APUs?
iGPUs have been getting ever closer to entry level and even mid-range GPUs.
In addition there's a interest in having a lot of memory for LLM acceleration. I expect both CPUs to get more LLM acceleration capabilities and desktop pc memory bandwidth to increase from its current rather slow dual channel 64bit DDR5-6000 status quo.
We're already hearing the first rumors for Medusa Halo coming in 2026 with 50% more bandwidth than Strix Halo.
The sentence "In addition there's a interest in having a lot of memory for LLM acceleration" was supposed to say "In addition there's a interest in having a lot of memory bandwidth for LLM acceleration" but it's too late to edit it now.
I think the market is very limited for high end iGPUs in practice with the compromises that occur with them.
On Desktop, upgradability is very popular and obviously the returns from the cooling on discrete GPUs are immense. With GPU dies costing so much, due to their size and dependency on TSMC, pushing the faster but hotter is probably a cost effecient compromise.
On Laptops with APUs, you currently ususally give up upgradeable memory - the fastest LPDDR is only soldered on (today), and the fastest solution would be on-die memory for bandwith gains that only really Apple is doing.
Marketing wise, low core count Laptops appear to be hard to sell. Gaming laptops seem to ship with more cores than the desktop you would build - the CPU appears out-specced.
I think this is because CPUs are cheaper, but that means a high-end APU would also need large CPU to compete. Now you've got a relatively unbalanced APU, with expensive hot CPU and relatively hot iGPU crammed in a small space - cooling is now tricky.
This is going to be compared with cheap RTX 4060 laptops - and generally look bad by comparison. I think what's changing now to narrow the gap is Handhelds, and questionable practices from Nvidia.
The Steam Deck kicked big OEMs into requesting AMD for large APUs.
Nvidia seems to have influence on OEM AMD Laptops - Intel CPU and Nvidia GPU for years now seem to ship first, in larger quantities, and get marketing push despite CPU arguably being worse.
Intel despite their issues seem to raising the iGPU bar too - their Desktop GPU investment seems to be paying off, and might be pressuring AMD to react.
Memory bandwidth. Besides LLMs, gaming on an iGPU will always be more expensive for the same performance as dedicated GPUs due to memory bandwidth.
Before someone tells me consoles using iGPUs, keep in mind that consoles use GDDR as its main system memory which has slow access times for the CPU. In a non-console, CPU performance is important. GDDR is also power hungry so they can't be used as the main system RAM in a laptop form.
It is the thermal envelope that defines pretty much everything nowadays. Without active management of it chips would die a heat death very fast. Which also means chips are designed with a certain chip external heat management in mind. The more heat you can get out of a system and away from a chip, the more powerful you can design these things. And game consoles do have active cooling, i.e. they sit between desktop PCs and thin laptops, probably sharing the thermal handling capacity with larger gaming laptops, if anything.
Just look at how a discrete GPU vs. an integrated GPU look like in terms of size, power, cooling, and other constraints like memory type and placement. That’s why both options still exist. If one size did it all, the other option would just die out.
APUs are going to replace low end video cards, because they no longer make economical or technical sense.
Historically those cards had narrow memory bus and about a quarter or less video memory of high end (not even halo) cards from the same generation.
That narrow memory bus puts their max memory bandwidth at a comparable level to desktop DDR5 with 2 DIMMs. At the same time quarter of high end is just 4GB VRAM which is not enough for low details for many games and prevents upscaling/frame gen from working.
From manufacturing standpoint low end GPUs aren't great either - memory controllers, video output and a bunch of other non-compute components don't scale with process node.
At the same time unified memory and bypassing PCIE benefits igpus greatly. You don't have to build an entire card, power delivery, cooler - you just slightly beef up existing ones.
tl;dr; sub-200 dollas GPUs are dead and will be replaced by APUs. I won't be surprised if they will start nibbling at lower mid-range market too in the near future.
My main gaming rig (for admittedly not very intensive games) has been a 7000 series Ryzen APU with a 780M, and my next one will also be an APU. It makes zero economic sense to build a discrete CPU system for casual gaming, even if I believe that APU prices will be artificially inflated to "cozy up" to low-end discrete GPU prices for a while to maximize profits.
Which is why for the games I play, a graphics workstation laptop like Thinkpad P series is much more useful, including GPGPU coding outside gaming, without being an heavyweight circus laptop whose battery lasts half-hour.
>> Are we going to see more these supposedly laptop APUs finding their way into desktops, and essentially a bifurcation of desktops into APUs and discrete CPU/GPUs?
I sure hope so. We could use a new board form factor that omits the GPU slot. Although my case puts the power connector and button over that slot on the back so it's not completely wasted, but the board area is. This has seemed like a good idea for a long time to me.
This can also be a play against nVidia. When mainstream systems use "good enough" integrated GPUs and get rid of that slot, there is no place for nVidia except in high-end systems.
There is no need for a new board form factor, because they have existed for many decades.
Below the mini-ITX format with a GPU slot, there are 3 standard form factors that are big enough for a full-featured personal computer that is more powerful than most laptops: nano-ITX (120 mm x 120 mm, for 5" by 5" cases; half the area of mini-ITX), 3.5" (from the size of the 3.5 inch HDDs, approximately the same area with nano-ITX, but rectangular instead of square) and the 4" x 4" NUC format introduced by Intel.
With a nano-ITX or 3.5" board you can make a computer not bigger than 1 liter that can ensure a low noise even for a 65 W power dissipation for the CPU+iGPU and that can have a generous amount of peripheral ports, to cover all needs.
Keeping the low noise condition, one could increase the maximum power-dissipation to 150 W for the CPU+iGPU in a somewhat bigger case, but certainly still smaller than 2.5 liter.
I expect that we will see such mini-PCs with Strix Halo, the only question is whether their price would be low enough to make them worthwhile.
The fabrication cost for Strix Halo must be lower than for a combo of CPU with discrete GPU, but the initial offerings with it attempt to make the customer pay more for the benefit of having a more compact system, which for many people will not be enough motivation to accept a higher price.
Having a system level cache for low latency transfer of data between CPU and GPU could be very compelling for some applications even if the overall GPU power is lower than a dedicated card. That doesn't seem to be the case here, though?
Strix Halo has unified memory, which is the same general architecture as Apple's M series chips. This means the CPU and GPU share the same memory, so there is no need to copy CPU <-> GPU.
The bifurcation is already happening. The last few years have seen lots of miniPC/NUC like products being released.
One of (many) factors that were holding back this form factor was the gap in iGPU/GPU performance. However with the frankly total collapse of the low end GPU market in the last 3-4 years, there's a much larger opening for iGPUs.
I also think that within the gaming space specifically, a lot of the chatter around the Steam Deck helped reset expectations. Like if everyone else is having fun playing games at 800p low/medium, then you suddenly don't feel so bad playing at maybe 1080p medium on your desktop.
I don't really like these "lightly edited" machine transcripts. There are transcription errors in many paragraphs, just adds that little bit of extra friction when reading.
I really want LPDDR5X (and future better versions) to become standard on desktops, alongside faster and more-numerous memory controllers to increase overall bandwidth. Why hasn't CAMM gotten anywhere?
I also really want an update to typical form factors and interconnects of desktop computers. They've been roughly frozen for decades. Off the top of my head:
- Move to single-voltage power supplies at 36-57 volts.
- Move to bladed power connectors with fewer pins.
- Get rid of the "expansion card" and switch to twinax ribbon interconnects.
- Standardize on a couple sizes of "expansion socket" instead, putting the high heat-flux components on the bottom side of the board.
- Redesign cases to be effectively a single ginormous heatsink with mounting sockets to accept things which produce heat.
- Kill SATA. It's over.
- Use USB-C connectors for both power and data for internal peripherals like disks. Now there's no difference between internal and external peripherals.
There's a rumor that future desktops will use LPDDR6 (with CAMMs presumably) instead of DDR6. Of course CAMMs will be slower so they might "only" run at ~8000 GT/s while soldered LPDDR6 will run at >10000.
Framework asked AMD if they could use CAMM for their new Framework Desktop.
AMD actually humored the request and did some engineering, with simulations. According to Framework, the memory bandwidth on the simulations was less than half of the soldered version.
This completely defied the entire point of the chip - the massive 256 bit bus ideal for AI or other GPU-heavy tasks, which allows this chip to offer the features it does.
This is also why Framework has apologized for non upgradability, but said it can’t be helped, so enjoy fair and reasonable RAM prices. Previously, it had been speculated that CAMM had a performance penalty, but Framework’s engineer on video saying it was that bad was fairly shocking.
The problem was specifically routing the 256-bit LPDDR5X out of the chip into the CAMM2 connector. This is hard to do with such a wide bus, because LPDDR5X wasn't originally designed for it.
I'm curious how much the CUDIMM thing Intel is doing, where the RAM has its own clock, can help in the CAMM context. The Zen 4/5 memory controller doesn't support it but a future one might.
I do not believe they were asking for CAMM as replacement for soldered RAM, but as an upgrade for DIMMs in desktop.
CAMM is touted as being better than DIMMs when it comes to signal integrity and possible speed. Soldered of course beat any socket, in-package beats any soldered RAM, and on-die beat any external component.
That AMD Strix Halo is unable to maintain signal integrity for any socketed RAM is a Strix Halo problem, not a socket problem. They probably backed themselves a bit into a corner with other parts of the design sacrifying tolerances on the memory side, and it's a lot easier to push motherboard design requirements than redoing a chip.
If this wasn't a Strix Halo issue, then they would have been able to run with socketed memory with a lower memory clock. All CPUs, this one included, has variable memory clocks that could be utilized and perform memory training as even the PCB traces to the chip cause significant signal degradation.
For signal integrity issues increasing the link power can often overcome some of the issues caused by longer traces and connectors in the line - while less of an issue for desktop devices, then that goes against the ideal of a low-powered device with limited cooling. Doubly so as it's hard to re-clock in the timescales needed for intermittent use power saving, so will be using that extra power when idle.
I suspect the earlier comment about "Half the performance with CAMM" is likely at iso-power, but that might still be a pretty big dealbreaker.
> - Move to single-voltage power supplies at 36-57 volts.
Why? And why not 12V? Please be specific in your answers.
> - Get rid of the "expansion card" and switch to twinax ribbon interconnects.
If you want that, it's available right now. Look for a product known as "PCI Express Riser Cable". Given that the "row of slots to slot in stiff cards" makes for nicely-standardized cases and card installation procedures that are fairly easy to understand, I'm sceptical that ditching slots and moving to riser cables for everything would be a benefit.
> - Kill SATA. It's over.
I disagree, but whatever. If you just want to reduce the number of ports on the board, mandate Mini SAS HD ports that are wired into a U.2 controller that can break each port out into four (or more) SATA connectors. This will give folks who want it very fast storage, but also allow the option to attach SATA storage.
> - Use USB-C connectors for both power and data for internal peripherals like disks.
God no. USB-C connectors are fragile as all hell and easy to mishandle. I hate those stupid little almost-a-wafer blades.
> - Standardize on a couple sizes of "expansion socket" instead...
What do you mean? I'm having trouble envisioning how any "expansion socket" would work well with today's highly-variably-sized expansion cards. (I'm thinking especially of graphics accelerator cards of today and the recent past, which come in a very large array of sizes.)
> - Redesign cases to be effectively a single ginormous heatsink with mounting sockets...
See my questions to the previous quote above. I currently don't see how this would work.
> Why? And why not 12V? Please be specific in your answers.
Higher voltages improve transmission efficiency, in particularly for connectors, as long as sufficient insulation is easy to maintain. Datacenters are looking at 48V for a reason.
Nothing comes for free though, and it makes for slightly more work for the various buck converters.
> God no. USB-C connectors are fragile as all hell and easy to mishandle. I hate those stupid little almost-a-wafer blades.
They are numerous orders of magnitude more rugged than any internal connector you've used - most of them are only designed to handle insertion a handful of times (sometime connectors even only work once!), vs. ten thousand times for the USB-C connector. In that sense, a locking USB-C connector would be quite superior.
... on that single metric. It would be ridiculously overcomplicated, driving up part costs when a trivial and stupidly cheap connector can do the job sufficiently. Having to run off 48V to push 240W and have no further power budget at all also increase complexity, cost and add limitations.
USB-C is meant for end-user things where everything has to be crammed into the same, tiny connector, where it does great.
Interesting read, and interesting product. If I understand it right, this seems like it could be at home in a spiritual successor to the Hades Canyon NUCs. I always thought those were neat.
I wish Chips and Cheese would clean up transcripts instead of publishing verbatim. Maybe I'll use the GPU on my Strix Halo to generate readable transcripts of Chips and Cheese interviews.
Although I appreciate the drive for small profile I wonder where the limits are if you put a big tower cooler onto it, seeing as the broad design direction is for laptops or consoles I doubt there's too much left on the table. I think that highlights a big challenge - is there a sizeable enough market for it, or can you pull in customers from other segments to buy a NUC instead. You'd need a certain amount of mass manufacturing with a highly integrated design to make it worthwhile.
> can you pull in customers from other segments to buy a NUC instead
I've never understood the hype for NUCs for non-office settings. You can make SFF builds that are tiny and still fit giant GPUs like the RTX 3090 /4090, say less for something like a 4080 Super. And then you can upgrade the GPU and (woe is you) CPU later on. Although a high-end X3D will easily last you 2-3 GPU generations.
You could fit a NUC in a pair of cargo shorts, FWIW. Or many bicycle under-seat bags, which was nice for biking to school without needing any backpack. They were in a sort of… qualitatively smaller size class than laptops.
Yeah would it have killed them to read over it just once? Can they not find a single school kid to do it for lunch money or something? Hell I'll do it for free, I've read this article twice now, and read everything they put out the moment it hits my inbox.
I think having a (small desktop) system with Strix Halo plus a GPU to accelerate prompt processing could be a good combo, avoiding the weakness of the Mac Ultra. The Strix Halo has 16 PCIe lanes.
Note that none of the PCIe interfaces on Strix Halo are larger than x4. The intent is to allow multiple NVMe drives and a Wi-Fi card. We also used PCIe for 5Gbit Ethernet.
Feedback: That 4x slot looks like it's closed on the end. Can we get an open-ended slot there instead so we can choose to install cards with longer interfaces? That's often a useful fallback.
Max RAM for Strix Halo is 128GB. It's not a competitor to the Mac Ultra which goes up to 512GB.
You shouldn't need another GPU to do prompt processing for Strix Halo since the biggest model it can realistically run is a 70B model. Prompt processing isn't going to help much because it has a good enough GPU but its memory bandwidth is only 256GB/s (~210 GB/s effective).
Despite the hype, the 512GB Mac is not really a good buy for LLMs. The ability to run a giant model on it is a novelty that will wear off quickly... it's just too slow to run them at that size, and in practice it has the same sweet spot of 30-70B that you'd have with a much cheaper machine with a GPU, without the advantage of being able to run smaller models at full-GPU-accelerated speed.
2 to 3 tokens per second was actually probably fine for most things last year.
Now, with reasoning and deep searching, research models, you’re gonna generate 1000 or more tokens just as it’s talking to itself to figure out what to do for you.
So everyone’s focused on how big a model you can fit inside your ram, the inference speed is now more important than it was.
The thinking models really hurt. I was happy with anything that ran at least as fast as I could read, then "thinking" became a thing and now I need it to run ten times faster.
I guess code is tough too. If I'm talking to a model I'll read everything it says, so 10-20 tok/s is well and good, but that's molasses slow if it's outputting code and I'm scanning it to see if it looks right.
counterpoint: thinking models are good since they give similar quality at smaller RAM sizes. if a 16b thinking model is as good as a 60b one shot model, you can use more compute without as much RAM bottleneck
The $2000 strix halo with 128 GB might not compete with the $9000 Mac Studio with 512 GB but is a competitor to the $4000 Mac Studio with 96 GB. The slow memory bandwidth is a bummer, though.
but is a competitor to the $4000 Mac Studio with 96 GB. The slow memory bandwidth is a bummer, though.
Not really. The M4 Max has 2x the GPU power, 2.13x the bandwidth, faster CPU.
$2000 M4 Pro Mini is more of a direct comparison. The Mini only has 64GB max ram but realistically, a 32B model is the biggest model you want to run with less than 300 GB/s bandwidth.
People keep saying "to compete with Apple" which of course, is nonsense. Apple isn't even second or third place in laptop marketshare last I checked.
So why build powerful laptops? Simple: people want powerful laptops. Remoting to a desktop isn't really a slam dunk experience, so having sufficient local firepower to do real work is a selling point. I do work on both a desktop and a laptop and it's nice being able to bring a portable workstation wherever I might need it, or just around the house.
This is a really good point. It's not easy to use both a laptop and a desktop at the same time. There's challenges around locality, latency, limited throughput, unavailability that software can't easily deal with, so you need to be aware and smart about it, and you'll need to compromise on things.
I'd work from my workstation at all times if I could. Tramp is alright, but not too fast and fundamentally can't make things transparent.
128GB is actually a step down. The previous generation (of sorts) Strix Point had maximum memory capacity of 256GB.
The mini-PC market (which basically all uses laptop chips) seems pretty robust (especially in Asia/China). They've basically torn out the bottom of the traditional small form factor market.
I've never seen a good technical comparison showing what's new between "Unified Memory" vs traditional APUs/iGPUs memory subsystems laptops have had for over a decade, only comparisons to dGPU setups which are rarer in laptops. The biggest differences comparing Apple Silicon or Strix Halo to their predecessors seems to be more about the overall performance scale, particularly of the iGPU, than the way memory is shared. Articles and blogposts most commonly reference:
- The CPU/GPU memory are shared (does not have to be dedicated to be used by either).
- You don't need to copy data in memory to move it between the CPU/GPU.
- It still uses separate caches for the CPU & GPU but the two are able to talk to each other directly on the same die instead of an external bus.
But these have long been true of traditional APUs/iGPUs, not new changes. I did even see some claims Apple put the memory on die too and that's what makes it unified but checking that it seems to still actually be "on package", which isn't unique either, and it wouldn't explain any differences in access patterns anyways. I've been particularly confused as to why Strix Halo would now qualify as having Unified Memory when it doesn't seem anything is different than before, save the performance.
If anyone has a deeper understanding of what's new in the Unified Memory approach it'd be appreciated!
There is no such thing as simple CUDA C API, that is the mistake most folks do when talking about CUDA.
It won over OpenCL, because it is a polyglot ecosystem, with first tier support for C, C++, Fortran, and Python (JIT DSL), plus several languages that have toolchains targeting PTX, the IDE integration, graphical debugger, compute and graphical rendering libraries.
All of the above AMD and Intel could have provided for OpenCL, but never did when it mattered, not even after SPIR was introduced.
Now they finally have GPGPU support for Fortran, C++, Python JIT DSLs, but a bit too late to the party, because contrary to NVidia it isn't like those tools are available regardless of the card.
The early versions had been only C. Then they added a lot of stuff.
You don't need all the fancy stuff, but OpenCL (and even more so Vulkan) are too complicated when all you want to do is some gpu number crunching.
Being able to write a kernel with something that looks like C. Having pointers on gpu and cpu and being able to call these kernels somewhat conveniently (like CUDA C) would be a great starting point.
Seems like Apple's M2 is a sweet spot for AI performance at 800 GB/s of memory bandwidth which can be added under $1,500 refurbished for 65 gigs of RAM.
I don't understand the name "Strix". It's a name a GPU and motherboard partner of theirs, Asus, uses (used?) for their products. It's impossible for me to read "AMD Strix" and not think of it as some ASU's GPU with an AMD chip in it, or some motherboard for AMD sockets.
Aren't there enough syllables out there to invent a combination which doesn't collide with your own board partners?
I understand it's internal codename, but also can't read it without thinking Asus. Especially considering I have Asus Strix 4090 in my rig.
I don't think AMD really uses the name "Strix Halo" to market it to a large audience, it's just an internal codename. Just two other recent internal names are "Hawk Point" and "Dragon Range" internally, where Hawk and Dragon are names that MSI and PowerColor use to market GPUs as as well. Heck, PowerColor even exclusively sells AMD cards under the "Red Dragon" name!
AMD's marketing names for especially their mobile chips are just so deliberately confusing that it makes way more sense for press and enthusiasts to keep referring to it by its internal code name than whatever letter/number/AI nonsense AMD's marketing department comes up with.
Strix https://en.wikipedia.org/wiki/Strix_(mythology) Halo is the code name.
AMD Ryzen AI MAX 300 is the product name. This continuing to use the code name.
Well, it's a public enough code name that it surprises me that they just used Asus's name.
Confused me too.
Did AMD not know?
Or did AMD know and not care?
AMD makes exclusive deals with ASUS regularly. I guess they're just good friends.
oh great, they are not using a confusing name anymore... wait, now they are using a stupid name!
Spy X Family!
AMD is captured.
For me the question is: what does this mean for future of desktop CPUs? High bandwidth unified memory seems very compelling for many applications, but the GPU doesn't have as much juice as a separate unit. Are we going to see more these supposedly laptop APUs finding their way into desktops, and essentially a bifurcation of desktops into APUs and discrete CPU/GPUs? Or will desktop CPUs also migrate to becoming APUs?
iGPUs have been getting ever closer to entry level and even mid-range GPUs.
In addition there's a interest in having a lot of memory for LLM acceleration. I expect both CPUs to get more LLM acceleration capabilities and desktop pc memory bandwidth to increase from its current rather slow dual channel 64bit DDR5-6000 status quo.
We're already hearing the first rumors for Medusa Halo coming in 2026 with 50% more bandwidth than Strix Halo.
The sentence "In addition there's a interest in having a lot of memory for LLM acceleration" was supposed to say "In addition there's a interest in having a lot of memory bandwidth for LLM acceleration" but it's too late to edit it now.
This has been the case for decades now.
GPU:s have existed about 30 years. Embedded ones for 20 years or so? Why are the embedded GPU:s always so stunted?
I think the market is very limited for high end iGPUs in practice with the compromises that occur with them.
On Desktop, upgradability is very popular and obviously the returns from the cooling on discrete GPUs are immense. With GPU dies costing so much, due to their size and dependency on TSMC, pushing the faster but hotter is probably a cost effecient compromise.
On Laptops with APUs, you currently ususally give up upgradeable memory - the fastest LPDDR is only soldered on (today), and the fastest solution would be on-die memory for bandwith gains that only really Apple is doing.
Marketing wise, low core count Laptops appear to be hard to sell. Gaming laptops seem to ship with more cores than the desktop you would build - the CPU appears out-specced. I think this is because CPUs are cheaper, but that means a high-end APU would also need large CPU to compete. Now you've got a relatively unbalanced APU, with expensive hot CPU and relatively hot iGPU crammed in a small space - cooling is now tricky.
This is going to be compared with cheap RTX 4060 laptops - and generally look bad by comparison. I think what's changing now to narrow the gap is Handhelds, and questionable practices from Nvidia.
The Steam Deck kicked big OEMs into requesting AMD for large APUs.
Nvidia seems to have influence on OEM AMD Laptops - Intel CPU and Nvidia GPU for years now seem to ship first, in larger quantities, and get marketing push despite CPU arguably being worse.
Intel despite their issues seem to raising the iGPU bar too - their Desktop GPU investment seems to be paying off, and might be pressuring AMD to react.
Before someone tells me consoles using iGPUs, keep in mind that consoles use GDDR as its main system memory which has slow access times for the CPU. In a non-console, CPU performance is important. GDDR is also power hungry so they can't be used as the main system RAM in a laptop form.
> Memory bandwidth.
It is the thermal envelope that defines pretty much everything nowadays. Without active management of it chips would die a heat death very fast. Which also means chips are designed with a certain chip external heat management in mind. The more heat you can get out of a system and away from a chip, the more powerful you can design these things. And game consoles do have active cooling, i.e. they sit between desktop PCs and thin laptops, probably sharing the thermal handling capacity with larger gaming laptops, if anything.
>Why are the embedded GPU:s always so stunted?
Because gpu want a lot of silicon. 5080 is 300 mm^2. Meanwhile ryzen 9xxx is 50 mm^2.
Meanwhile CPU wants to use that wafer space for themselves. And even if you use 100% of the wafer space for GPU you will have a small gpu and no cpu
Just look at how a discrete GPU vs. an integrated GPU look like in terms of size, power, cooling, and other constraints like memory type and placement. That’s why both options still exist. If one size did it all, the other option would just die out.
I would love a unified memory architecture, even for external GPUs.
Pay for memory once, and avoid all the copying around between CPU/GPU/NPU for mixed algorithms, and have the workload define the memory distribution.
APUs are going to replace low end video cards, because they no longer make economical or technical sense.
Historically those cards had narrow memory bus and about a quarter or less video memory of high end (not even halo) cards from the same generation.
That narrow memory bus puts their max memory bandwidth at a comparable level to desktop DDR5 with 2 DIMMs. At the same time quarter of high end is just 4GB VRAM which is not enough for low details for many games and prevents upscaling/frame gen from working.
From manufacturing standpoint low end GPUs aren't great either - memory controllers, video output and a bunch of other non-compute components don't scale with process node.
At the same time unified memory and bypassing PCIE benefits igpus greatly. You don't have to build an entire card, power delivery, cooler - you just slightly beef up existing ones.
tl;dr; sub-200 dollas GPUs are dead and will be replaced by APUs. I won't be surprised if they will start nibbling at lower mid-range market too in the near future.
My main gaming rig (for admittedly not very intensive games) has been a 7000 series Ryzen APU with a 780M, and my next one will also be an APU. It makes zero economic sense to build a discrete CPU system for casual gaming, even if I believe that APU prices will be artificially inflated to "cozy up" to low-end discrete GPU prices for a while to maximize profits.
Which is why for the games I play, a graphics workstation laptop like Thinkpad P series is much more useful, including GPGPU coding outside gaming, without being an heavyweight circus laptop whose battery lasts half-hour.
>> Are we going to see more these supposedly laptop APUs finding their way into desktops, and essentially a bifurcation of desktops into APUs and discrete CPU/GPUs?
I sure hope so. We could use a new board form factor that omits the GPU slot. Although my case puts the power connector and button over that slot on the back so it's not completely wasted, but the board area is. This has seemed like a good idea for a long time to me.
This can also be a play against nVidia. When mainstream systems use "good enough" integrated GPUs and get rid of that slot, there is no place for nVidia except in high-end systems.
There is no need for a new board form factor, because they have existed for many decades.
Below the mini-ITX format with a GPU slot, there are 3 standard form factors that are big enough for a full-featured personal computer that is more powerful than most laptops: nano-ITX (120 mm x 120 mm, for 5" by 5" cases; half the area of mini-ITX), 3.5" (from the size of the 3.5 inch HDDs, approximately the same area with nano-ITX, but rectangular instead of square) and the 4" x 4" NUC format introduced by Intel.
With a nano-ITX or 3.5" board you can make a computer not bigger than 1 liter that can ensure a low noise even for a 65 W power dissipation for the CPU+iGPU and that can have a generous amount of peripheral ports, to cover all needs.
Keeping the low noise condition, one could increase the maximum power-dissipation to 150 W for the CPU+iGPU in a somewhat bigger case, but certainly still smaller than 2.5 liter.
I expect that we will see such mini-PCs with Strix Halo, the only question is whether their price would be low enough to make them worthwhile.
The fabrication cost for Strix Halo must be lower than for a combo of CPU with discrete GPU, but the initial offerings with it attempt to make the customer pay more for the benefit of having a more compact system, which for many people will not be enough motivation to accept a higher price.
Having a system level cache for low latency transfer of data between CPU and GPU could be very compelling for some applications even if the overall GPU power is lower than a dedicated card. That doesn't seem to be the case here, though?
Strix Halo has unified memory, which is the same general architecture as Apple's M series chips. This means the CPU and GPU share the same memory, so there is no need to copy CPU <-> GPU.
That new 'desktop' from framework appears to be just that with the AMD Ryzen Al Max 385
We have both Max 385 and Max+ 395 versions.
Any word on putting that in a mobile device? So far there’s only a hp business laptop and a gaming tablet… none of which appeal to the “MacBook crowd”
The bifurcation is already happening. The last few years have seen lots of miniPC/NUC like products being released.
One of (many) factors that were holding back this form factor was the gap in iGPU/GPU performance. However with the frankly total collapse of the low end GPU market in the last 3-4 years, there's a much larger opening for iGPUs.
I also think that within the gaming space specifically, a lot of the chatter around the Steam Deck helped reset expectations. Like if everyone else is having fun playing games at 800p low/medium, then you suddenly don't feel so bad playing at maybe 1080p medium on your desktop.
Framework made a tiny desktop form factor version with this chip in it, so we'll if it gets much traction (at least among enthusiasts).
When, if ever, will this be released as a bare processor/memory/motherboard combination that I can buy and throw in my own case?
Does anyone know?
Framework is selling the board as a stand alone: https://frame.work/fr/en/products/framework-desktop-mainboar...
Too bad there isn't a full PCIe (might not be enough bandwidth left) :(.
I was looking to see if they sold only the motherboard just last night and failed. Thanks!!
$1,299 (64GB) and $1,599 USD (128GB) for the motherboards. Yikes, but I get why.
Here's one motherboard: https://frame.work/products/framework-desktop-mainboard-amd-...
I wouldn't be surprised if Minisforum also offers a motherboard.
Framework Desktop
https://frame.work/products/desktop-diy-amd-aimax300
I don't really like these "lightly edited" machine transcripts. There are transcription errors in many paragraphs, just adds that little bit of extra friction when reading.
I really want LPDDR5X (and future better versions) to become standard on desktops, alongside faster and more-numerous memory controllers to increase overall bandwidth. Why hasn't CAMM gotten anywhere?
I also really want an update to typical form factors and interconnects of desktop computers. They've been roughly frozen for decades. Off the top of my head:
- Move to single-voltage power supplies at 36-57 volts.
- Move to bladed power connectors with fewer pins.
- Get rid of the "expansion card" and switch to twinax ribbon interconnects.
- Standardize on a couple sizes of "expansion socket" instead, putting the high heat-flux components on the bottom side of the board.
- Redesign cases to be effectively a single ginormous heatsink with mounting sockets to accept things which produce heat.
- Kill SATA. It's over.
- Use USB-C connectors for both power and data for internal peripherals like disks. Now there's no difference between internal and external peripherals.
There's a rumor that future desktops will use LPDDR6 (with CAMMs presumably) instead of DDR6. Of course CAMMs will be slower so they might "only" run at ~8000 GT/s while soldered LPDDR6 will run at >10000.
LPDDR6 won't go that low, even on CAMM2. The interface is designed for up to 14.4Gbps, with initial modules aiming for 10.6Gbps.
> Why hasn't CAMM gotten anywhere?
Framework asked AMD if they could use CAMM for their new Framework Desktop.
AMD actually humored the request and did some engineering, with simulations. According to Framework, the memory bandwidth on the simulations was less than half of the soldered version.
This completely defied the entire point of the chip - the massive 256 bit bus ideal for AI or other GPU-heavy tasks, which allows this chip to offer the features it does.
This is also why Framework has apologized for non upgradability, but said it can’t be helped, so enjoy fair and reasonable RAM prices. Previously, it had been speculated that CAMM had a performance penalty, but Framework’s engineer on video saying it was that bad was fairly shocking.
The problem was specifically routing the 256-bit LPDDR5X out of the chip into the CAMM2 connector. This is hard to do with such a wide bus, because LPDDR5X wasn't originally designed for it.
LPDDR6X is designed for it, and an use CAMM2.
I'm curious how much the CUDIMM thing Intel is doing, where the RAM has its own clock, can help in the CAMM context. The Zen 4/5 memory controller doesn't support it but a future one might.
I do not believe they were asking for CAMM as replacement for soldered RAM, but as an upgrade for DIMMs in desktop.
CAMM is touted as being better than DIMMs when it comes to signal integrity and possible speed. Soldered of course beat any socket, in-package beats any soldered RAM, and on-die beat any external component.
That AMD Strix Halo is unable to maintain signal integrity for any socketed RAM is a Strix Halo problem, not a socket problem. They probably backed themselves a bit into a corner with other parts of the design sacrifying tolerances on the memory side, and it's a lot easier to push motherboard design requirements than redoing a chip.
If this wasn't a Strix Halo issue, then they would have been able to run with socketed memory with a lower memory clock. All CPUs, this one included, has variable memory clocks that could be utilized and perform memory training as even the PCB traces to the chip cause significant signal degradation.
For signal integrity issues increasing the link power can often overcome some of the issues caused by longer traces and connectors in the line - while less of an issue for desktop devices, then that goes against the ideal of a low-powered device with limited cooling. Doubly so as it's hard to re-clock in the timescales needed for intermittent use power saving, so will be using that extra power when idle.
I suspect the earlier comment about "Half the performance with CAMM" is likely at iso-power, but that might still be a pretty big dealbreaker.
> - Move to single-voltage power supplies at 36-57 volts.
Why? And why not 12V? Please be specific in your answers.
> - Get rid of the "expansion card" and switch to twinax ribbon interconnects.
If you want that, it's available right now. Look for a product known as "PCI Express Riser Cable". Given that the "row of slots to slot in stiff cards" makes for nicely-standardized cases and card installation procedures that are fairly easy to understand, I'm sceptical that ditching slots and moving to riser cables for everything would be a benefit.
> - Kill SATA. It's over.
I disagree, but whatever. If you just want to reduce the number of ports on the board, mandate Mini SAS HD ports that are wired into a U.2 controller that can break each port out into four (or more) SATA connectors. This will give folks who want it very fast storage, but also allow the option to attach SATA storage.
> - Use USB-C connectors for both power and data for internal peripherals like disks.
God no. USB-C connectors are fragile as all hell and easy to mishandle. I hate those stupid little almost-a-wafer blades.
> - Standardize on a couple sizes of "expansion socket" instead...
What do you mean? I'm having trouble envisioning how any "expansion socket" would work well with today's highly-variably-sized expansion cards. (I'm thinking especially of graphics accelerator cards of today and the recent past, which come in a very large array of sizes.)
> - Redesign cases to be effectively a single ginormous heatsink with mounting sockets...
See my questions to the previous quote above. I currently don't see how this would work.
> Why? And why not 12V? Please be specific in your answers.
Higher voltages improve transmission efficiency, in particularly for connectors, as long as sufficient insulation is easy to maintain. Datacenters are looking at 48V for a reason.
Nothing comes for free though, and it makes for slightly more work for the various buck converters.
> God no. USB-C connectors are fragile as all hell and easy to mishandle. I hate those stupid little almost-a-wafer blades.
They are numerous orders of magnitude more rugged than any internal connector you've used - most of them are only designed to handle insertion a handful of times (sometime connectors even only work once!), vs. ten thousand times for the USB-C connector. In that sense, a locking USB-C connector would be quite superior.
... on that single metric. It would be ridiculously overcomplicated, driving up part costs when a trivial and stupidly cheap connector can do the job sufficiently. Having to run off 48V to push 240W and have no further power budget at all also increase complexity, cost and add limitations.
USB-C is meant for end-user things where everything has to be crammed into the same, tiny connector, where it does great.
Interesting read, and interesting product. If I understand it right, this seems like it could be at home in a spiritual successor to the Hades Canyon NUCs. I always thought those were neat.
I wish Chips and Cheese would clean up transcripts instead of publishing verbatim. Maybe I'll use the GPU on my Strix Halo to generate readable transcripts of Chips and Cheese interviews.
The Framework desktop seems like a next step.
Although I appreciate the drive for small profile I wonder where the limits are if you put a big tower cooler onto it, seeing as the broad design direction is for laptops or consoles I doubt there's too much left on the table. I think that highlights a big challenge - is there a sizeable enough market for it, or can you pull in customers from other segments to buy a NUC instead. You'd need a certain amount of mass manufacturing with a highly integrated design to make it worthwhile.
> can you pull in customers from other segments to buy a NUC instead
I've never understood the hype for NUCs for non-office settings. You can make SFF builds that are tiny and still fit giant GPUs like the RTX 3090 /4090, say less for something like a 4080 Super. And then you can upgrade the GPU and (woe is you) CPU later on. Although a high-end X3D will easily last you 2-3 GPU generations.
The size of NUCs is much smaller than any SFF builds with RTX 3090. Some people just like smallness.
Closer to a phone than a laptop, in size!
You can’t mount an SFF build unobtrusively behind a monitor or under a desk, it’s much larger and heavier than a NUC.
Agreed, although this is the smallest I've seen:
https://github.com/phkahler/mellori_ITX/blob/master/images/m...
It's currently got a 5700G - Zen 3 in it and 64GB RAM. I'd like the next one to hang on the back of a monitor or TV via the standard mounting holes.
You could fit a NUC in a pair of cargo shorts, FWIW. Or many bicycle under-seat bags, which was nice for biking to school without needing any backpack. They were in a sort of… qualitatively smaller size class than laptops.
i feel like high-end mini-ITX builds only became viable a few years ago with the introduction of 700W+ SFX PSUs.
Yeah would it have killed them to read over it just once? Can they not find a single school kid to do it for lunch money or something? Hell I'll do it for free, I've read this article twice now, and read everything they put out the moment it hits my inbox.
I think having a (small desktop) system with Strix Halo plus a GPU to accelerate prompt processing could be a good combo, avoiding the weakness of the Mac Ultra. The Strix Halo has 16 PCIe lanes.
Note that none of the PCIe interfaces on Strix Halo are larger than x4. The intent is to allow multiple NVMe drives and a Wi-Fi card. We also used PCIe for 5Gbit Ethernet.
Love what you're doing, I'm in batch 4!
Feedback: That 4x slot looks like it's closed on the end. Can we get an open-ended slot there instead so we can choose to install cards with longer interfaces? That's often a useful fallback.
Max RAM for Strix Halo is 128GB. It's not a competitor to the Mac Ultra which goes up to 512GB.
You shouldn't need another GPU to do prompt processing for Strix Halo since the biggest model it can realistically run is a 70B model. Prompt processing isn't going to help much because it has a good enough GPU but its memory bandwidth is only 256GB/s (~210 GB/s effective).
Despite the hype, the 512GB Mac is not really a good buy for LLMs. The ability to run a giant model on it is a novelty that will wear off quickly... it's just too slow to run them at that size, and in practice it has the same sweet spot of 30-70B that you'd have with a much cheaper machine with a GPU, without the advantage of being able to run smaller models at full-GPU-accelerated speed.
It runs DeepSeek R1 q4 MoE well enough.
It does have an edge on being able to run large MoE models.
There’s so much flux in LLM requirements.
2 to 3 tokens per second was actually probably fine for most things last year.
Now, with reasoning and deep searching, research models, you’re gonna generate 1000 or more tokens just as it’s talking to itself to figure out what to do for you.
So everyone’s focused on how big a model you can fit inside your ram, the inference speed is now more important than it was.
Absolutely.
The thinking models really hurt. I was happy with anything that ran at least as fast as I could read, then "thinking" became a thing and now I need it to run ten times faster.
I guess code is tough too. If I'm talking to a model I'll read everything it says, so 10-20 tok/s is well and good, but that's molasses slow if it's outputting code and I'm scanning it to see if it looks right.
counterpoint: thinking models are good since they give similar quality at smaller RAM sizes. if a 16b thinking model is as good as a 60b one shot model, you can use more compute without as much RAM bottleneck
The $2000 strix halo with 128 GB might not compete with the $9000 Mac Studio with 512 GB but is a competitor to the $4000 Mac Studio with 96 GB. The slow memory bandwidth is a bummer, though.
$2000 M4 Pro Mini is more of a direct comparison. The Mini only has 64GB max ram but realistically, a 32B model is the biggest model you want to run with less than 300 GB/s bandwidth.
You will be limited to a much smaller context size with half the RAM even if you're using a smaller model.
Of course it's a competitor. Only a fraction of M3 Ultra sold will have 512GB RAM
> Max RAM for Strix Halo is 128GB. It's not a competitor to the Mac Ultra which goes up to 512GB.
What a... strange statement. How did you get to that conclusion?
Why do you think it's strange?
Running something like Qwq 32b q4 with a ~50k context will use up those 128GB with the large KV cache.
But why did they choose to build this as a mobile cpu though? I don’t need 128GB of unified RAM on my laptop. It’s the desktop where things happen.
People keep saying "to compete with Apple" which of course, is nonsense. Apple isn't even second or third place in laptop marketshare last I checked.
So why build powerful laptops? Simple: people want powerful laptops. Remoting to a desktop isn't really a slam dunk experience, so having sufficient local firepower to do real work is a selling point. I do work on both a desktop and a laptop and it's nice being able to bring a portable workstation wherever I might need it, or just around the house.
This is a really good point. It's not easy to use both a laptop and a desktop at the same time. There's challenges around locality, latency, limited throughput, unavailability that software can't easily deal with, so you need to be aware and smart about it, and you'll need to compromise on things.
I'd work from my workstation at all times if I could. Tramp is alright, but not too fast and fundamentally can't make things transparent.
Compete with Apple is my guess. There is a decent market for super high end laptops.
Framework (I believe) made one of these into a purchasable desktop.
128GB is actually a step down. The previous generation (of sorts) Strix Point had maximum memory capacity of 256GB.
The mini-PC market (which basically all uses laptop chips) seems pretty robust (especially in Asia/China). They've basically torn out the bottom of the traditional small form factor market.
Because desktops are a much smaller market and AMD caught the Apple Silicon FOMO.
I wonder if you could actually put these into a socket or issues would occur.
While I think the market is small, and they don't release a lot of these, AMD has sold desktop / socketed APUs in the past.
They tend to come out much slower than the laptop chips, or the "CPU-only" desktop chips.
This is one of the more recent examples: https://www.amd.com/en/products/processors/desktops/ryzen/80...
Because people are accustom to unified memory in laptops and also complain about the low amounts of ram and inability to upgrade.
This solves those problems but apparently uncovers a new one.
> Because people are accustom to unified memory in laptops
Surely the vast majority of laptops sold in the last five years don't have unified memory yet.
I've never seen a good technical comparison showing what's new between "Unified Memory" vs traditional APUs/iGPUs memory subsystems laptops have had for over a decade, only comparisons to dGPU setups which are rarer in laptops. The biggest differences comparing Apple Silicon or Strix Halo to their predecessors seems to be more about the overall performance scale, particularly of the iGPU, than the way memory is shared. Articles and blogposts most commonly reference:
- The CPU/GPU memory are shared (does not have to be dedicated to be used by either).
- You don't need to copy data in memory to move it between the CPU/GPU.
- It still uses separate caches for the CPU & GPU but the two are able to talk to each other directly on the same die instead of an external bus.
But these have long been true of traditional APUs/iGPUs, not new changes. I did even see some claims Apple put the memory on die too and that's what makes it unified but checking that it seems to still actually be "on package", which isn't unique either, and it wouldn't explain any differences in access patterns anyways. I've been particularly confused as to why Strix Halo would now qualify as having Unified Memory when it doesn't seem anything is different than before, save the performance.
If anyone has a deeper understanding of what's new in the Unified Memory approach it'd be appreciated!
Yes, around 90% of laptops sold in the last ten years have unified memory.
Cool I guess for a mini PC but Im one of those desktop PC tower nerds :)
As long as they cant even provide something similiar to a simple CUDA C API on consumer hardware i dont buy their stuff.
There is no such thing as simple CUDA C API, that is the mistake most folks do when talking about CUDA.
It won over OpenCL, because it is a polyglot ecosystem, with first tier support for C, C++, Fortran, and Python (JIT DSL), plus several languages that have toolchains targeting PTX, the IDE integration, graphical debugger, compute and graphical rendering libraries.
All of the above AMD and Intel could have provided for OpenCL, but never did when it mattered, not even after SPIR was introduced.
Now they finally have GPGPU support for Fortran, C++, Python JIT DSLs, but a bit too late to the party, because contrary to NVidia it isn't like those tools are available regardless of the card.
The early versions had been only C. Then they added a lot of stuff.
You don't need all the fancy stuff, but OpenCL (and even more so Vulkan) are too complicated when all you want to do is some gpu number crunching.
Being able to write a kernel with something that looks like C. Having pointers on gpu and cpu and being able to call these kernels somewhat conveniently (like CUDA C) would be a great starting point.
Early meaning until CUDA 3.0 in 2010, we are now on CUDA 12.8, 15 years later.
Seems like Apple's M2 is a sweet spot for AI performance at 800 GB/s of memory bandwidth which can be added under $1,500 refurbished for 65 gigs of RAM.
Where for $1500?
Not on Apple Refurbs. That would cost you about $2200.
And the M2 Max has a memory bandwidth of 400GB/s.
I’m guessing a reference to M2 Ultra? Not sure about that price though…
M2 Ultra refurb was over $4,000, last I checked.