Comment by simonw
9 hours ago
The numbers I was most interested in are tucked away in a chart towards the bottom - the speed comparison of the Mac Studios v.s. a RTX 5090:
Qwen3.8 27B tokens/sec generation speed
Prompt size 8K 64K 128K 256K
RTX 5090 PC 59 51 44 n/a
M5 Ultra 48 39 32 24
M3 Ultra 31 23.5 20 15
A whole bunch more comparison numbers in this section: https://www.macstories.net/stories/m5-ultra-mac-studio-revie...
Those RTX 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.
can confirm.
I dont' know why people spend huge money on these and Spark. The 5090 is running qwen 3.8 at 200+ tps!! That's 1-2 orders of magnitude faster.
Have a 5090, and yes it's very fast. But it's like the worst ADHD team member and requires constant supervision and review from larger models. It's context size on-card is good for super, suuuuuper shallow precision work. The gb10/spark on top of it, that thing can refactor enormous monorepo architecture. The time it takes the 5090 to compact, reiterate and execute a plan is often the same time as the gb10.
1 reply →
How are you deciding which work to send to the 5090 vs a frontier model, or making the two work together nicely?
Correct is much more important than fast for me, but if I could get correct and fast, that would obviously be amazing.
People don't buy Sparks and M5 Ultras to run a 27B model - you buy it to run an MoE model like Qwen Next which this M5 excelled at.
2 replies →
Is a 5090 still cost efficent when it is (currently) unobtainable? Or when obtainable only at current prices (min. $6500 USD)?
1 reply →
Because you can run Qwen 3.8 Flash Next, Laguna S 2.1 and other medium-sized models that simply don't fit on a 5090?
A) the macos value add is enormous if you have any investment in the ecosystem, B) for me at least a GPU is completely useless for anything but being a token generator.
3 replies →
A 5090 has a 1.79TB/s memory bandwidth. Qwen 3.8 27B NVFP4 is 22GB. You cannot generate tokens faster than the weights can traverse the GPU memory, so that makes max generation speed without MTP to be 81T/s. Say MTP is giving you 0.5 acceptance rate (very good), that is 1.5 * 81 is 121T/s. Even with a perfect acceptance rate you would only get 162T/s.
4 replies →
Qwen3.8-Flash-Next is pretty damn worth the extra ram you need.
same reason they spend huge amounts of money on rolexes when seikos work better (the tech crowd isn't immune from vanity).
14 replies →
The issue is that the moment you want to run the more capable models that will no longer fit in a single 5090's memory, performance falls off a cliff.
Both are probably single-token decode performance, which is reasonable to show. Otherwise agree RTX 5090 should shinebetter with NVFP4.
... or with llama.cpp with MTP.
[flagged]
I guess it is possible, but Apple has had very vocal fans for decades. I suspect, rather than astroturfing, it is just people who are in their ecosystem.
Tok/sec is 0 on a 3090 for most of the models that the mac can run
2 replies →
Those are some incredible graphs, that leap in prompt processing going from M3 to M5.
Also: ~30 token/s on GLM 5.3-flash, locally. (That's roughly Opus 4.8-tier. I think).
/meta Here's a CSS filter that stops those nuisance chart animations,
A dense 27B doesn't really make sense for the Mac. A MoE makes way more sense when you have modest bandwidth but lots of memory.
A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures. But when you hit the limit of what you can hold in memory, you reach the limitation of the platform.
Whereas a hybrid architecture with distinct DRAM and VRAM with sparse MoE, you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers and arbitrage the difference in cost for each of those in distinct classes of hardware.
> A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures
Inference time is going to be dominated by the low memory bandwidth on these Macs, so a dense model will suffer most. It’s more of an opportunity for large MoE models with a low number of active experts since you can keep all experts in VRAM but not pay the bandwidth cost until they are used.
> you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers
This is an interesting direction that I expect to see more of. But for most models currently you need basically all experts loaded since they are chosen per token.
Apple seems to be researching longer horizon expert caching, where they keep experts swapped in for longer runs of tokens [1]. Other labs are offloading ngram caches but not sure if they’re pursuing anything like this?
1. https://machinelearning.apple.com/research/introducing-third...
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two heads on.
1.2 T/s is not that modest is it? That's very close to an RTX pro 5000
A basic llama-bench on Qwen 3.8 27B UD-Q4_K_M gives pp512 3920 tok/s / tg128 81 tok/s on a 500W RTX PRO 6000 (should be similar speeds to a 5090, chip is basically the same, just less VRAM). With MTP3 this is 140 tok/s on mtp-bench.
This is with llama.cpp. You can of course use vLLM/SGLang well on these cards and they're even faster. On vLLM w/ NVIDIA/Qwen3.8-27B-NVFP4 baseline has a prefill of about 13,000 tok/s. The baseline tok/s is 72 tok/s, but at mtp7, it's 157 tok/s, and w/ dflash7 that goes up to 215 tok/s. On mtp-bench, DFlash2 gets a hair under 300 tok/s w/ the code_python prompt.
That's a dense model. Of course it will do worse.
Now try running that Qwen 3.8 Next model on the 5090 and tell me what TPS you get (hint: it's near 0 since it doesnt fit the 32GB VRAM on 5090 vs the 256 in OPs M5).
Surprisingly, the Reddit crowd are reporting 50–60 tokens/s (for the 32 GiB 5090 + 128 GiB RAM)—on par with the M5 Ultra benchmarks, despite both the PCIe bottleneck and much smaller DDR5 bandwidth,
https://old.reddit.com/r/LocalLLaMA/comments/1wl06np/qwen38f...
(Note it's a sparse MoE with only 6B active).
I have a 3 year old gaming system. RTX 4080 w/128GB of DDR5. It runs Qwen 38 Flash around 44-40 t/s with 128K context. It is on a specialized build that caches MoE experts and uses an optimized 3bit quant that basically is within a few points of the full 8 bit quant. In general, in casual benchmarking with Alibaba's endpoint I could not tell much of a difference. Overall this model is very good on long horizon agentic work. The main pain point for it is that its input processing speed is slow. Regardless, it gets meaningful work done.
I paid $500 for the RAM in Nov 2023 :)
2 replies →
Good to know thanks.
That's with CPU offload to a DDR5 6000 RAM though which is around $3-4k at least.
2 replies →
I really appreciate seeing these dense model numbers. For a large unified memory system though I expect that MoE numbers are what people are more interested in.
These numbers could and should get much better. As an example I can run Qwen3.8-27B-MXFP4 (W4A8) on 2x AMD R9700 that gets 260+ tokens/sec to start and slows down to ~110 tokens/sec over 128k context and can do the max 256k. These are for batch size 1 and throughput goes higher with batching. This is due to speculative decoding, efficient all-reduce inter-gpu compression, and custom GEMM kernels for the specific hardware. Note each R9700 only has 644 GB/s memory bandwidth.
On my m5 max 27b model does 75tps on 256k ctx and starts at 80 on the 8k ctx when you add https://huggingface.co/collections/z-lab/dflash-2 to it. So yeah base might be 30tps (I used iq4) but mtp or dflash help a lot and should be used when checking what is useful and what is not for running models as it is not fare to judge without them.
Thank you for this. I wish Apple focused their silicon design on improving the TTFT metrics but coming from an M3 Pro, it still looks laggard compared to Nvidia's TensorCores in the 5090.
Maybe Apple is an acquisition away from changing that balance.
The rumor on Apple's processor roadmap is that they're skipping other M6 variations (all previous generations had Pro and Max, a few had Ultra) in order to focus on the M7 generation for AI reasons. What exactly the M7 improvements are who knows.
https://www.macrumors.com/2026/06/25/2027-macs-m7-chips/
I think that comes down to TSMC. Nvidia apparently booked out the whole A18 or 16 node. Apple is on 2nm right now and M7 will jump right to A14. According to my quick AI research anyway.
2 replies →
> What exactly the M7 improvements are who knows.
> Apple's planned M7 Ultra chip is being designed to support up to 1.5 TB of unified memory and to push AI performance toward the class of Nvidia's Blackwell accelerators
https://www.tomshardware.com/tech-industry/semiconductors/ap...
M6 got another prompt processing boost. Likely no M6 Ultra though because Apple is reportedly going all in on AI performance in M7 generation.
I assume those are non-batched. I think the M series GPU can do 4X to 8X depending on model quant, which means if you can batch queries you'll get almost 4X to 8X performance.
The selling point of the M5 Ultra Mac Studio is that you can run much larger models that the 5090 can't without swapping. NVidia aggressively segments the market on VRAM for this reason. That's why a 5090 has an MSRP of ~$2k (but good luck getting one for less than $4k) while a 6000 Pro, which is basically a 5090 with 96GB of RAM has now soared beyond $15k where 3-6 months ago it was more like $10-11k. A 6000 Pro has the same memory bandwidth but slightly more CUDA units (IIRC ~24k vs ~21k).
This advantage won't be apparent with a 27B model. The 256GB MS can probably run the newer Flash models locally, something you can't do on a 5090.
I don't think we'll get a successor to the 5090 until late 2028, maybe even 2029. I'm basing this on the launch date of the 5000 series and that we haven't got a midcycle refresh yet. Rumor has it the chips are ready but the 3GB RAM modules are 3-4x the price of the 2GB modules used on the current cards.
Apple should see a Mac Studio major update in 2028. That might even force NVidia's hand. But it's really impossible to say what the state of the market will be 2-3 years from now. It may have completely crashed. I suspect not however.
The interesting thing will be when the bandwidth demands start forcing HBM memory onto these home/enthusiast solutions.
But what about builds that combine 8 of the 5090 with infiniband between boxes? Wouldn't that be comparable to the mac in terms of price and potentially beat it by a lot in terms of performance for the large MoE? I understand the space/heat/noise considerations, but price wise it may still not make as much sense as people think. (Agreed that it is hard to get the NVIDIA hardware and the 6000 pro are priced less competitively).
> But what about builds that combine 8 of the 5090 with infiniband between boxes?
Why Infiniband ("IB")? If it's for RDMA, that is possible with certain Ethernet cards/chipsets as well. Certainly Mellanox, but Broadcom:
* https://techdocs.broadcom.com/us/en/storage-and-ethernet-con...
and Intel as well:
* https://www.intel.com/content/www/us/en/support/articles/000...
Link level flow control or priority flow control needs to be supported on the switch ports as well.
No, $40K is not comparable to $10K.
1 reply →
>8 of the 5090
Where are you buying 8 5090s for under $10k? With CPU, RAM, and (checks comment) infiniband hardware???
You're probably looking at a lot closer to $60k when all is said and done, and that's before you hire an electrician to run a sub panel for your homelab...
Others have chimed in on cost and size, I'll chime in on power. The 8x 5090s will require a dedicated datacentre grade power source. The Mac Studio runs on a plain jane wall socket.
While that sounds super awesome, How many people are actually going to build and maintain that vs a box you can grab at the mall that fits in a lunchbox?
Sounds like nice utility bill in the making.
I can't speak to Infiniband pricing for something like that. It seems like the cheap option is 56/100Gbps with used Enterprise equipment. You'd need 8 HCAs, DAC cabling and a switch but even then you're into thousands of dollars. If you want 200Gbps+ it gets into the tens of thousands (AFAICT).
Each PC is probably going to cost ~$6k and you're talking about 8000W of electricity draw. That's going to consume multiple 20A circuits even at 240V. And the electricity ain't free either. A Mac Studio seems to draw ~500W max.
Oh and the Mac Studio has an upgrade route to run 1T+ models too by chaining them together with TB5 chaining. OSX supports RDMA this way. That's comparable bandwidth to the 100Gbps Infiniband option.
So you're talking about $50-60k of hardware and more power draw and more heat for something that will I'm sure beat the MS M5U option but at huge cost. Also, at that kind of price point, I'm likely to get a workstation PC and put 2 (or possibly 3) 6000 Pros in it.
openai make a npu google make npu (tpu no mater) amd buy tellas
every company make his own npu (without xai)
probaby in 2028 we will have more concurent firm on market place
Not forgetting of course that an RTX5090 is what 600W+ ? And the Mac is probably half that at most ?
Certainly not forgetting wattage. A 5090 is 575W. The M5 Ultra Studio is 480W.
nvidia-smi -pl 450 for like a 4% reduction in throughput. I tend to set it around 350W because it's a comfortable temperature blowing on my legs under the desk without warming my office in the summer.
I put together this system two years ago, so it's a little out of date, but it only cost $3000 for the same performance and capability as an Ultra. I don't think I would spend $7000 to save 100W, though.
> nvidia-smi -pl 450 for like a 4% reduction in throughput.
Yeah people don't pay enough attention to those settings IMO. The first thing I do when I set up a new machine (or upgrade my OS) is to restore all my powersaving configs.
For example I've got all but one of my virtual desktops that put the CPU in powersave mode: I don't need max Ghz when browsing the Web, not even on demand. But when I switch to the virtual desktop where my development environment is, then I want power on demand.
Now I don't do it to save the planet: I do it because I love a quieter computing experience (coupled with Be Quiet! PSU and Noctua fans, this makes for a very quiet computer). That it consumes less electricity is a nice side-benefit.
sure. so is 2x power worth 10x perf? I think it is in most cases.
When you are doing matrix math, compute is compute. Apple cant be more efficient due to physics. The only reason Macs are more efficient in general is that they have tightly bundled hw and sw for specific tasks.
The next Ultra, supposedly on deck in 2028:
> Apple's planned M7 Ultra chip is being designed to support up to 1.5 TB of unified memory and to push AI performance toward the class of Nvidia's Blackwell accelerators, according to a new Bloomberg report published by Mark Gurman...
Apple plans to release a base M6 chip this fall for entry-level Macs... a base M7 in the first half of 2027, M7 Pro and M7 Max at the end of 2027, and the M7 Ultra in 2028.
https://www.tomshardware.com/tech-industry/semiconductors/ap...