Did you see somewhere that it would have unified memory? I am extremely curious about that, but haven’t picked up on anything.Strix Halo points to the future of windows laptops. It will likely be expensive and it will be a niche product but it’s the most innovative product out of all the client Zen 5 products.
The few issues I have with it are the CCD is not 3NE and it’s using RDNA 3.5 and not 4. But it’s builds a base and I hope it’s successful because of it is we don’t have deal with crap like VRAM limitations ever on a laptop cause of the unified memory and who doesn’t like big APUs.
AMD APUs have unified memory.Did you see somewhere that it would have unified memory? I am extremely curious about that, but haven’t picked up on anything.
They've had it since Kaveri.Did you see somewhere that it would have unified memory?
It didn’t make it to Ryzen until the MI300A for the datacenter. None of the consumer Ryzen APUs have unified memory.They've had it since Kaveri.
?It didn’t make it to Ryzen until the MI300A for the datacenter.
Yeah they do. Since Kaveri.None of the consumer Ryzen APUs have unified memory.
Yeah a Strix Halo in that form factor would be a buy for me as long as the price isn't egregious.Oh no, I'm waiting to see it in something the size of this, or probably a bit bigger and throw my money on it
Where have you seen that? AMD’s hUMA documents end with Carrizo from what I’ve seen. I even bought and still have a Trinity laptop 9 or so years ago, thinking it had unified memory.AMD APUs have unified memory.
CPU and GPU and NPU may share a unified per-process virtual address space...
That's because HSA died with Carrizo.AMD’s hUMA documents end with Carrizo from what I’ve seen
You got any links stating this? I have Googled my ass off for years and have been looking for it since the first Ryzen launched. The last reference to unified memory I’m finding in consumer APUs is Carrizo.?
Yeah they do. Since Kaveri.
It's somewhere in Linux driver notes.You got any links stating this?
Of course, HSA died in 2016!The last reference to unified memory I’m finding in consumer APUs is Carrizo.
I am still looking, but so far see zero mention of unified memory. To be clear, I am not talking about * shared * memory - as in both the CPU and the GPU separately using the same system RAM. in that scenario, the CPU gives the GPU a chunk of the system RAM that then becomes slow VRAM for the GPU. Once that happens, the CPU units are not allowed to operate on it directly, information is copied back-and-forth as if it was on a separate physical device just as it was a discrete GPU.It's somewhere in Linux driver notes.
Of course, HSA died in 2016!
Unified memory is very not relevant outside of thin slice of GPGPU space.
Here's a conversation from an AMD GPU driver developer:Where have you seen that? AMD’s hUMA documents end with Carrizo from what I’ve seen. I even bought and still have a Trinity laptop 9 or so years ago, thinking it had unified memory.
I’ve been watching for indications of unified memory since the first Ryzens. I am not seeing anything. Searching now, I still don’t see anything. Any links?
consoles, phones?Unified memory is very not relevant outside of thin slice of GPGPU space.
I greatly appreciate that link! From just last month, late June 2024. The first mention that Ryzen APUs can do zero copy that I know of since the launch in 2017. And I have been looking for it since then.Here's a conversation from an AMD GPU driver developer:
AMD's AOMP 19.0-2 Compiler Brings Zero-Copy For CPU-GPU Unified Shared Memory - Phoronix Forums
Phoronix: AMD's AOMP 19.0-2 Compiler Brings Zero-Copy For CPU-GPU Unified Shared Memory AMD compiler engineers have released AOMP 19.0-2 as the newest version of their downstream LLVM/Clang compiler that carries all of their latest work around OpenMP/AOCC GPU device offloading to Radeon and...www.phoronix.com
APU have supported it, but no one runs ROCm on an APU except the big $10000 one. They can all share pointers to system memory across CPU/GPU/NPU. Where it needs more work is sharing VRAM to the CPU. But since Strix etc don't have VRAM that isn't an issue.
That's a very separate ecosystem.consoles
Phone GPU APIs are even more antiquated than what we have on big boy platforms.phones
Yeah that's the selling point.Strix Halo will be very interesting as a machine learning workstation if AMD is smart enough to expose unified memory in the drivers and allow that 40 CU GPU access to all 128 GBs of RAM
That's server levels of memory setups, not feasible.it would have been so much better if they had gone with a 512-bit bus
I suspect that's their plan now. ROCm + Halo large memory space could finally give them a niche to get a few sales.Strix Halo will be very interesting as a machine learning workstation if AMD is smart enough to expose unified memory in the drivers and allow that 40 CU GPU access to all 128 GBs of RAM.
it’s feasible, just much more expensive, However, there is a growing community that would gladly pay the higher price, since LLM text generation is severely memory bandwidth constrained. Llama 3 70B 4-bit will get about 8 tokens per second on an enabled Strix Halo, 15 tok/s would have made it a slam dunk.Yeah that's the selling point.
That's server levels of memory setups, not feasible.
Technically, yes; market-wise, no, at least for now.it’s feasible
Bubble huffers don't have much time left to live.there is a growing community that would gladly pay the higher price, since LLM text generation is severely memory bandwidth constrained
Oh yes, but that's niche.There are people who hate Apple almost as much as Igor does who are buying Macs for that unified memory and the 512-bit and 1024-bit buses.
It’s certainly possible, it will increase costs. just requires more channels. If AMD wants to keep it around $2000 then 256-bit is the right choice.That's server levels of memory setups, not feasible.
I love this term😂Bubble huffers don't have much time left to live.
You're not routing 512b on laptop PCBs.It’s certainly possible
How do I route all that stuff.just requires more channels.
Less.If AMD wants to keep it around $2000
How does Apple do it? Does the on-package memory help?You're not routing 512b on laptop PCBs.
MoP.How does Apple do it?
Yeah lmao.Does the on-package memory help?
I am a straight up bubble huffer. I say that AI functions will be the main reason for sales of computing devices by 2040. I still find it bizarre there’s so many people on various text site forums don’t see that AI is THE future of computing in society.Technically, yes; market-wise, no, at least for now.
Bubble huffers don't have much time left to live.
Oh yes, but that's niche.
-halo exists to displace x106 dies and below out of existence.