> Put two machines on a desk, each about $2,000. One is a tower with an NVIDIA RTX 5090: 32GB of the fastest consumer memory ever shipped, 1,792 GB/s. The other is a mini PC the size of a paperback, an AMD Ryzen AI Max+ 395 "Strix Halo" box with 128GB of soldered memory at roughly 256 GB/s.
Doesn't change the conclusions of the article, but each of those machines is more like $4k+
"integrated graphics processor, using system memory" had its name dragged through the mud for decades.
So we had to rebadge it to "unified memory".
Curious if we'll ever see some old integrated graphics processor "hacked" to manage to handle 128 GB of allocated system RAM and be able to serve diffusion-LLMs at a decent rate on "old" hardware...
I'm writing my own inference engine for Strix Halo and the same model. I already have 30%+ performance plus a more graceful decay over long contexts; that said, their point stands: memory bandwidth is what you really want.
For some reason, this reminds me of my last shared memory system. It was an Athlon XP 1800+ with VIA ProSavage back around 2002. It was just barely able to run CS 1.6.
Why would a RTX 5090 with 32 GB not be able to deal with a 40 GB model? Is there anything preventing me from swapping the weights that do not fit into VRAM in and out of RAM? PCIe 5.0 x16 should max out around 64 GB/s, so slower than the unified memory machine, but at least it should be possible.
For prefill, you are compute bound, and it is trivial to batch multiple input tokens together. When using cpu offload, software like llama.cpp will batch weight uploads with tokens that need those weights and perform work on the GPU. It works very well. With a large batch size and pcie5 you can get prefill speeds close to having all weights on the GPU.
For decode, you are bandwidth bound, and it is difficult to batch multiple output tokens together. There is no benefit to sending your weights to the GPU because even if it internally has insane bandwidth, you are still bottlenecked by system RAM (and adding a pcie5 upload would bottleneck it further). This is the number people usually talk about when they say they are getting a certain tk/s.
> For decode, (...) it is difficult to batch multiple output tokens together.
I think it's the other way around? The GPU has to stream gigabytes of active layer weights to compute the next token, so having a batch of next-token predictions sitting there on the GPU goingh through the layers makes better use of the bandwidth.
At least that's what I observed on a Strix Halo, batching 4 predictions yields like 2-3x the total tps.
MTP can give you small batches, but it is still WAY smaller than the batches you can get with prefill, which is limited only by the number of input tokens you have (but has diminishing returns on performance).
But:
1) It still makes no sense to upload the weights to the GPU with MTP as you are still bottlenecked by the weight upload.
5090 can do all but lossless NVFP4 (OMMA) and NVIDIA does fairly good quants of most anything popular. This isn't quite a 4x reduction from what you see on the label because they're a little conservative with the QKV projections (IMHO) but it's on the order of that. So a dense model at 50-70B parameters is the sweet spot. It's a great card for strong dense models.
In principle you could have bidirectional PCIe x16 pipelining at it would move the roofline a little with fast DDR5, I think llama.cpp has a flag for it.
Or go rent a B200 on vast.ai for 4 bucks an hour or thereabouts, a single heavy Opus session for a couple hours is like a week of any model on vast or RunPods.
NVIDIA publishes something called NGC containers that generally work out of the box. I started running Qwen3.6-NVFP4-MTP locally and then I'll put something heavy on Baseten if I'm lazy or Vast if I want a good deal.
Opus (and maybe now 5.6) are still the strongest for like, the really delicate shit, kernel modules or something, but that's on pace to cross over this year, and the overtraining and misalignment are getting so bad when they phase 4.6 out I'm pulling my plan. I don't pay to get gaslit about Constitutional AI.
I'm going to say this that we're not even close to the limits of what actually needs to be accomplished so at some point, memory will start needing better tiering for inference some day ....
If compute is not the bottleneck, memory is easy-ish to produce (the hard part is mostly on the fab side); what stops a Chinese NVIDIA (huawei) from being 10x cheaper?
Making memory is easy. Packaging that memory within a few millimeters of a piece of silicon using TSVs and maintaining signal integrity on a 1024 bit bus is really, really hard.
LLMs arenât all that compute constrained or even memory constrained. Itâs just that pushing dozens of terabits per second through a piece of silicon is a physics problem.
This is a very understandable misconception that I wouldn't blame anyone for having but MoE is actually terrible for inference in most any local LLM / home lab scenario. MoE is popular because it's cheap to train, but because most modern routing needs the previous layer's activations (except at the very beginning) it winds up being just this side of impossible to pipeline / prefetch without all the experts resident. Plus the grouped GEMM kernels have terrible support on any card in most people's house, it's just really unwieldy.
Dense models are very straightforward to share/pipeline because you know all the shapes and geometry up front, that's the inference friendly option.
> I guess they're just welding the memory to the CPU chip, but still curious.
Unified memory is more of an architectural and performance characteristic, and does not imply much about the physical layout of the machine. Most unified memory PCs not from Apple don't have the memory on the same package as the SoC. For stuff like AMD Strix Halo and NVIDIA DGX Spark, it's just standard LPDDR packages soldered on the motherboard in the general vicinity of the SoC, and the only difference from mainstream laptops for the past decade+ is that the memory bus is twice as wide.
Yes. The memory is just located very close to the cpu with wires "welded" directly to it. This allows the memory to be run as fast as possible but it's still a RAM component.
The cache parts of memory are on the CPU itself but they are on the order of MB not GB.
They are usually the same family, LPDDR is used for amd and macs, but the fabs are the same as the most expesive HBM memory, if they have a choice they are going to produce the ones that they can sell for more $$.
The current âbig GPUâ has 96gb of memory, but thatâs not a âconsumer GPUâ apparently, while a $5000 Spark is a âconsumer PCâ I guess. In any case youâre probably better off running a large open weights model on the cloud.
Can't really run it as well, though. My "mini PC" is an M4 Max with 128GB of unified memory and the memory bandwidth is still sorely lacking for inference (although it's far better than any non-unified consumer architecture!).
Yeah this is such a funny thing going around. Try to run or train a small/medium sized model on a MacBook. It doesnât go very good compared to a dedicated gpu
This is likely the right path in the future but it isnât there yet today
Sorry, thought we were talking about tokens. M5 Max is great for bandwidth and Iâm looking forward to seeing what Apple does for AI inference in the M7. The 6000 kills everything else when it comes to TTFT and tokens/s.
I was under the impression that when you're streaming the weights from disk because the full model won't fit in memory, that it is solely reading from the SSD, not writing, so it wouldn't be causing wear on your SSD.
NAND[0] has a fun thing called "read disturbance" where repeated reads from disk will eventually flip 0s to 1s. You have to erase and rewrite the block before the bits flip[1], or you lose the data, but doing so is the same amount of wear as a write.
[0] I heard this being an issue with TLC, I don't know if it also applied to MLC or SLC.
[1] I suspect in practice they use an error correction code and rewrite blocks that read with corrected errors.
It's kinda irresponsible to talk about read disturbance without clarifying that it takes an extremely large number of reads to cause a read disturb error, and it can be corrected by a single rewrite of the data. Read disturb errors are something SSD engineers need to account for, but from an end user perspective it's a smaller problem by multiple orders of magnitude than write endurance, which is already rarely a real problem in practice.
No one can reliably track uses of "it", and definitely not multiple uses, so please take your time to qualify everything explicitly. Currently I have no idea what each "it" is referring to.
> Put two machines on a desk, each about $2,000. One is a tower with an NVIDIA RTX 5090: 32GB of the fastest consumer memory ever shipped, 1,792 GB/s. The other is a mini PC the size of a paperback, an AMD Ryzen AI Max+ 395 "Strix Halo" box with 128GB of soldered memory at roughly 256 GB/s.
Doesn't change the conclusions of the article, but each of those machines is more like $4k+
https://www.microcenter.com/product/711961/amd-ryzen-ai-halo...
"integrated graphics processor, using system memory" had its name dragged through the mud for decades.
So we had to rebadge it to "unified memory".
Curious if we'll ever see some old integrated graphics processor "hacked" to manage to handle 128 GB of allocated system RAM and be able to serve diffusion-LLMs at a decent rate on "old" hardware...
afaik you can do that now on DDR4 platform Mini PC's that can handle 48gb or 64gb DIMM's.
I'm writing my own inference engine for Strix Halo and the same model. I already have 30%+ performance plus a more graceful decay over long contexts; that said, their point stands: memory bandwidth is what you really want.
Think future generations of AMD could get quite interesting. Theyâre no doubt seeing people whining about mem throughput specifically
For some reason, this reminds me of my last shared memory system. It was an Athlon XP 1800+ with VIA ProSavage back around 2002. It was just barely able to run CS 1.6.
Why would a RTX 5090 with 32 GB not be able to deal with a 40 GB model? Is there anything preventing me from swapping the weights that do not fit into VRAM in and out of RAM? PCIe 5.0 x16 should max out around 64 GB/s, so slower than the unified memory machine, but at least it should be possible.
There are two phases to LLMs:
1) prefill
2) decode
For prefill, you are compute bound, and it is trivial to batch multiple input tokens together. When using cpu offload, software like llama.cpp will batch weight uploads with tokens that need those weights and perform work on the GPU. It works very well. With a large batch size and pcie5 you can get prefill speeds close to having all weights on the GPU.
For decode, you are bandwidth bound, and it is difficult to batch multiple output tokens together. There is no benefit to sending your weights to the GPU because even if it internally has insane bandwidth, you are still bottlenecked by system RAM (and adding a pcie5 upload would bottleneck it further). This is the number people usually talk about when they say they are getting a certain tk/s.
> For decode, (...) it is difficult to batch multiple output tokens together.
I think it's the other way around? The GPU has to stream gigabytes of active layer weights to compute the next token, so having a batch of next-token predictions sitting there on the GPU goingh through the layers makes better use of the bandwidth.
At least that's what I observed on a Strix Halo, batching 4 predictions yields like 2-3x the total tps.
MTP can give you small batches, but it is still WAY smaller than the batches you can get with prefill, which is limited only by the number of input tokens you have (but has diminishing returns on performance).
But:
1) It still makes no sense to upload the weights to the GPU with MTP as you are still bottlenecked by the weight upload.
2) I'm not sure MTP helps much with MoE models.
It's slower than the 4:1 ratio would imply, but it does indeed work.
Things get really slow if the model doesn't for in vram + ram and you have to go from disk to ram to vram.
5090 can do all but lossless NVFP4 (OMMA) and NVIDIA does fairly good quants of most anything popular. This isn't quite a 4x reduction from what you see on the label because they're a little conservative with the QKV projections (IMHO) but it's on the order of that. So a dense model at 50-70B parameters is the sweet spot. It's a great card for strong dense models.
In principle you could have bidirectional PCIe x16 pipelining at it would move the roofline a little with fast DDR5, I think llama.cpp has a flag for it.
Or go rent a B200 on vast.ai for 4 bucks an hour or thereabouts, a single heavy Opus session for a couple hours is like a week of any model on vast or RunPods.
NVIDIA publishes something called NGC containers that generally work out of the box. I started running Qwen3.6-NVFP4-MTP locally and then I'll put something heavy on Baseten if I'm lazy or Vast if I want a good deal.
Opus (and maybe now 5.6) are still the strongest for like, the really delicate shit, kernel modules or something, but that's on pace to cross over this year, and the overtraining and misalignment are getting so bad when they phase 4.6 out I'm pulling my plan. I don't pay to get gaslit about Constitutional AI.
It's time to have an exit strategy.
I'm going to say this that we're not even close to the limits of what actually needs to be accomplished so at some point, memory will start needing better tiering for inference some day ....
If compute is not the bottleneck, memory is easy-ish to produce (the hard part is mostly on the fab side); what stops a Chinese NVIDIA (huawei) from being 10x cheaper?
Making memory is easy. Packaging that memory within a few millimeters of a piece of silicon using TSVs and maintaining signal integrity on a 1024 bit bus is really, really hard.
LLMs arenât all that compute constrained or even memory constrained. Itâs just that pushing dozens of terabits per second through a piece of silicon is a physics problem.
I think it's mostly the ramp-up time, but ChangXin Memory Technologies (CXMT) is basically aspiring to do just this.
"Can't" is not really correct.
Nowadays, specially with MoE models you can run parts of the model on GPU and still get some speed up.
This is a very understandable misconception that I wouldn't blame anyone for having but MoE is actually terrible for inference in most any local LLM / home lab scenario. MoE is popular because it's cheap to train, but because most modern routing needs the previous layer's activations (except at the very beginning) it winds up being just this side of impossible to pipeline / prefetch without all the experts resident. Plus the grouped GEMM kernels have terrible support on any card in most people's house, it's just really unwieldy.
Dense models are very straightforward to share/pipeline because you know all the shapes and geometry up front, that's the inference friendly option.
MoE sells a lot of HBMe3.
Do unified memory CPUs suffer from the same memory shortages as normal memory?
I guess they're just welding the memory to the CPU chip, but still curious.
> I guess they're just welding the memory to the CPU chip, but still curious.
Unified memory is more of an architectural and performance characteristic, and does not imply much about the physical layout of the machine. Most unified memory PCs not from Apple don't have the memory on the same package as the SoC. For stuff like AMD Strix Halo and NVIDIA DGX Spark, it's just standard LPDDR packages soldered on the motherboard in the general vicinity of the SoC, and the only difference from mainstream laptops for the past decade+ is that the memory bus is twice as wide.
Yes. The memory is just located very close to the cpu with wires "welded" directly to it. This allows the memory to be run as fast as possible but it's still a RAM component.
The cache parts of memory are on the CPU itself but they are on the order of MB not GB.
They are usually the same family, LPDDR is used for amd and macs, but the fabs are the same as the most expesive HBM memory, if they have a choice they are going to produce the ones that they can sell for more $$.
The current âbig GPUâ has 96gb of memory, but thatâs not a âconsumer GPUâ apparently, while a $5000 Spark is a âconsumer PCâ I guess. In any case youâre probably better off running a large open weights model on the cloud.
Uhh the 5090 alone is double the cost of their quoted PC prices.
Can't really run it as well, though. My "mini PC" is an M4 Max with 128GB of unified memory and the memory bandwidth is still sorely lacking for inference (although it's far better than any non-unified consumer architecture!).
Yeah this is such a funny thing going around. Try to run or train a small/medium sized model on a MacBook. It doesnât go very good compared to a dedicated gpu
This is likely the right path in the future but it isnât there yet today
To be fair it's "only" half the throughput of a 4090 and a third of an RTX 6000. Significant but not an order of magnitude.
Those are the ratios for memory bandwidth, but the GPUs have a much higher ratio for compute, and that affects prefill rate / TTFT, right?
An old ada Rtx 6000 maybe. A Blackwell RTX Pro 6000 is an order of magnitude faster and has 96gb.
That's not what I'm seeing. It is much faster but not an order of magnitude. Not trying to be pedantic, only setting expectations.
"The Blackwell RTX PRO 6000 provides up to 1,792 GB/s of memory bandwidth, while the 40-core Apple M5 Max tops out at 614 GB/s"
Sorry, thought we were talking about tokens. M5 Max is great for bandwidth and Iâm looking forward to seeing what Apple does for AI inference in the M7. The 6000 kills everything else when it comes to TTFT and tokens/s.
For sure. Clearly Nvidia mops the floor with the competition. I'm looking forward to M6/M7 and to see if Apple wants a bigger piece of the pie.
Let's also ensure the SSD doesn't age prematurely.
I was under the impression that when you're streaming the weights from disk because the full model won't fit in memory, that it is solely reading from the SSD, not writing, so it wouldn't be causing wear on your SSD.
You'd need your OS to support, and be configured to use, a disk mounting option that disables file access timestamps, otherwise reads ARE writes.
...and disable swap.
NAND[0] has a fun thing called "read disturbance" where repeated reads from disk will eventually flip 0s to 1s. You have to erase and rewrite the block before the bits flip[1], or you lose the data, but doing so is the same amount of wear as a write.
[0] I heard this being an issue with TLC, I don't know if it also applied to MLC or SLC.
[1] I suspect in practice they use an error correction code and rewrite blocks that read with corrected errors.
It's kinda irresponsible to talk about read disturbance without clarifying that it takes an extremely large number of reads to cause a read disturb error, and it can be corrected by a single rewrite of the data. Read disturb errors are something SSD engineers need to account for, but from an end user perspective it's a smaller problem by multiple orders of magnitude than write endurance, which is already rarely a real problem in practice.
It is and it doesn't. You only get into disk writes if the system starts paging out to disk.
No one can reliably track uses of "it", and definitely not multiple uses, so please take your time to qualify everything explicitly. Currently I have no idea what each "it" is referring to.