*Official Benchmark Thread* - Post it here or it didn't happen :D

[Mr. Fox]

3 hours ago, electrosoft said:

Yeah, the 5080 really has the headroom too fully appreciate overclocking at what tops out at an extreme but "sane"ish type level vs the 5090 requiring a nuclear reactor.

 

Extrapolating, that Steel Nomad is like scoring ~ 21k+ on a 5090...

 

That Timespy score is bonkers both GPU, CPU and combined...

 

I see you right behind Jwick on the bot in the 2 slot a few times, lots of top 5/10 finishes and #1 in TSE!

 

Excellent work!

Thank you. Yeah, it does have a lot of headroom. The scores are lower but it is actually more fun due to the headroom.

 

Here are the settings applied with mvolt+ Extreme. I still need to test with the 5090. If they can make the function survive NVIDIA driver updates this eliminates the need for hardware mods like EVC2 and shunt resistors.

 

 

[electrosoft]

M5 Ultra vs a 9950X3D + 5090 kitted out PC:

 

CPU:

 

Bonkers, next level CPU performance. Absolutely trashes Intel and AMD at 150w with their 36 core model both single and multi.

 

Single core:  ~44% faster

Multi core:  ~80% faster

 

All this while consuming 25% less power in a small, compact box that makes zero noise 99% of the time and even under load is a fraction of the noise a typical PC generates.

 

You're going to need a Threadripper 9970x (32 core)  to compete, but not beat, at this point on multi. To flat out beat Apple, you will need a 9980x. On both, they absolutely get destroyed for single core performance.

 

On Single, Intel and AMD are going to have to bring some serious magic in their next gen CPUs or Apple will just continue to dominate and I don't think they're going to even get close to catch Apple in single threaded performance.

 

Looking forward to the 52 core Intel Nova "Open all salvos!" Lake beast vs the M5 Ultra. Heck, even the 28 core might get close in multi.

 

Blender Render = M5 Ultra is ~72% faster

 

GPU:

 

Time to come back to reality Apple vs Nvidia.....

 

Optimized for Nvidia, 5090 pulls ahead showing Nvidia's strength in terms of raw power, vertical integration and software stacks. 

 

LLMs: If memory isn't an issue, 5090 is a beast but obviously 32GB can't stack up against 96GB or 256GB. Toss in an RTX 6000 (or two or three) and it's game over but then again you can score a fully kitted M5 Ultra with 96GB of memory and 1TB storage and max CPU/GPU config for $6799.99 for less than half the price of an RTX 6000. You can score a 256GB model for $10799 which is 2/3 the cost of a single RTX 6000 at this point.

 

In the end, Apple's massively upgraded Neural Processors are no match for Nvidia's raw power but their unified memory architecture is where they have always shined. Expect the 512GB model (not available yet) to cost ~$17k as the 256GB model is currently $10,799.,99 and upgrading from 96GB to 256GB is a $4k up charge. So you end up with a 512GB M5 Ultra for what is quickly becoming the cost of a single RTX 6000.

 

Video editing, M5 Ultra:

 

DaVinci = tied

Abobe Pro = M5 Ultra finished in less than half the time

 

Gaming:

 

Time for a brutal reality check. Even Metal optimized, 5090 just trashes the M5 Ultra and is over twice as fast. M5 Ultra looks to be right between 5070ti-5080 in gaming performance but looking at some of the results that are actually taking advantage of the 80 core GPU, there is some optimizations to be had but....yeah, the 5090 just drops the hammer on Apple....just like Nvidia dropped the hammer on AMD with the 5090.

 

 

My final verdict? M5 CPU architecture is next level. Absolutely next level. GPU has really come along way even vs M4 and its strength lies in its unified memory bandwidth especially in these AI driven days.

 

For my personal use, target LLMs for particular use cases the 5090's 32GB is working out just fine and this is where it absolutely shines. If you need more than 32GB, performance scarily drops off as expected and you need to start looking at the 5080 based 4500 card with 48GB or even 5500 (barf) up to a 6000 to match the 96GB config from Apple.

 

256GB config for $10,799 really makes a compelling config for many small AI shops, startups or individuals because it costs less than a single 6000 by almost $5k and you get over 2.5x the memory. I suspect the 512GB will also be a very compelling buy for those on a budget but needing 512GB of memory.

 

 

 

[ssj92]

Guys, I remember someone here posting that the new 6xx drivers are not good. 

 

Can someone catch me up on why?

 

I'm on 596.xx on both desktop & laptops. 

 

I finally got SLI working on 596.xx drivers using the same old patch for 446.xx drivers. SO now I am aiming for 616.xx drivers. 

[Mr. Fox]

1 hour ago, ssj92 said:

Guys, I remember someone here posting that the new 6xx drivers are not good. 

 

Can someone catch me up on why?

 

I'm on 596.xx on both desktop & laptops. 

 

I finally got SLI working on 596.xx drivers using the same old patch for 446.xx drivers. SO now I am aiming for 616.xx drivers. 

I haven't heard anything about newer drivers other than mvolt+ not working with some driver versions.