Woohoo!
Would love to see render times using our Flame benchmark test.
The old Flame benchmark is marginal at this point. It predates a lot of new stuff, like all the ML models.
There have been less than 5 entries in recent times.
wow. seems to beat the threadripper 9975WX by 30% on multicore. and 20% lower than blackwell 6000
The topline numbers are very impressive. But should still be taken with some context.
Where M5 wins is with total universal memory size (if you shell out for 256 or 512GB). It also wins in direct memory speed from CPU to GPU (M5 1.2TB/s to all memory, no PCIe bottleneck).
If your app has proper Metal / ML integration (Flame, Nuke, Resolve, etc. do) they can take full advantage of this raw performance and impress.
Those are the sweet spot use cases that matter to us. And if you stay in this narrow swim lane most of the time, the M5 is hard to beat.
That said, the Blackwell 6000 has a memory bandwidth lead of 1.79TB/s on card, albeit requiring texture transfer via a much slower PCIe5 bus. And the Z8 main memory tops out at 410GB/s. The Blackwell wins on long compute tasks where local memory is more important than interchange.
So it is very workload-dependent. The ecosystem of ML tools still favors Cuda by a lot, whether that is ComfyUI or other tools. And that isn’t changing anytime soon.
You may be happy with native ML tools from Flame, and then you can take full advantage. At IBC two vendors have shown OFX plugins that allow you to ComfyUI models inside Flame without leaving the app. But those require Cuda to run locally. Or you need to run in the cloud and pay for tokens and be allowed to route material through cloud servers. Considerable drawbacks.
Two caveats on the Geekbench number: You cannot mix GB6 and GB7 numbers. GB7’s test does some face tracking, upscale, blur, and fluid simulation, among others. But it’s a burst, not a 40min sustained compute, which has its own challenges. There may be caveats in terms of power/heat management between the M5 and a Z8 to which we have no test data coverage. It could go either way.
Summary: The M5 is amazing. And for many Flame artists, a great platform that is “affordable”. But our workflows are also in flux, as we’ve seen at IBC, and the M5 continues to have some blind spots. At this price point, you need to make sure those blind spots don’t haunt you in the next 3-5 years before you get your next system. The question to ask: “Are you confident that Flame’s ML tools are everything you will need in terms of ML in the next 3+ years, or are there gaps? Is cloud compute allowed and cost appropriate?”
Personally, what I would choose: buy a used M3 Ultra with 256GB, and buy a Z4 with a Blackwell 4500 card. 6000 if you have some spare cash. Those are cost/performance sweet spots at the moment. That covers all your bases much more reliably. And you have the flexibility of an upgradable platform while also having the cheap power of the Mac. Then upgrade one or the other in the midpoint of the lifespan of the other.
Both platforms Mac and Z8 have seen about a 2.25x performance increase since 2021. Neither one is left in the dust; they just have different strengths.
PS: You can substitute Dell for Z8, they have similar systems. I’m just more familiar with the Z8.
Fair and reasonable points.
The price of a single nvidia 6000 is around 19,000 to 26.000 euros (greek price) and 5.500 to 6.000 euros for 4500 blackwell.
A well equipped m5 mac ultra can be had for at least 13.500 euro,add a well equipped external storage plus 5.000 to 6.000 for software. I think the total sum is roughly the price of a single GPU.
These days hardware is expensive and for some economies 20.000 for a single component is hard to justify. I am not saying that macs are the best or you can’t just build a workstation with of the shelf components. I think apple is a more sensible choice and a middle ground till the dust settles with prices.
If you need an exhaustive TL;DR comparison of all the specs in a decision between NVidia GPU, AMD ML box, NVidia ML Box, or Mac M5 Ultria, here’s every number and argument together.
And everywhere he says 5090, just replace with RTX Pro 4500.
The answer - it all depends on your use case, and what you can tolerate more.
this is an interesting watch too https://www.youtube.com/watch?v=OXZ14OErwOw