Discussion Cinebench 2024 Released

[tamz_msc]

It is better simply on basis of including GPU and shows how much faster those are at this kind of tasks.

In a way, yes. Most reviewers simply do not know how to benchmark workflows that 3D artists use in Blender or CInema 4D. Nobody uses the CPU to render working with these kind of applications.

 

[Nothingness]

According to Techspot/HUB, scaling is 48x for the 7980X in CB 2024. Anandtech shows 52x scaling in CB R23.

Geekbench 6 Ray Tracer scaling is 52-55x (I only looked at two results). So it seems to scale a bit better than CB R24.

 

[Timmah!]

In a way, yes. Most reviewers simply do not know how to benchmark workflows that 3D artists use in Blender or CInema 4D. Nobody uses the CPU to render working with these kind of applications.

I like TechGage, they are like the first reviewer i noticed to test GPUs for pro work, when everyone else did only games.
I guess CPUs have still their niche, because of higher amount available RAM compared to GPUs, but nowadays at least from my experience with Octane Render, stuff that does not fit into VRAM can be stored in system RAM, so its not limiting factor anymore. Though it renders slower then - but seeing how RTX 4090 is like 20x faster than 14900k/7950x, i presume even then its still faster than that.

 

[StefanR5R]

Go on chips and cheese site, they say that there is quite bit more trips to RAM,

Could be the higher RAM bandwidth that helps the CB2024 renderer.

LTT posted Threadripper 7000 results a) with quad-channel and b) dual-channel RAM. I am not sure if it was DDR5-5200 or -6400 in those tests, seemingly the latter.
(source: https://www.youtube.com/watch?v=WE9NIZShfZY&t=355s – "It’s Back and I’m SO Excited! - Threadripper 7000", 5:55 "Dual Channel Memory Results")
Cinebench 2024 multi-core,

7980X quad channel ............. 5,605 points
7980X dual channel .............. 5,146 points (-8 %)
7970X quad channel ............. 3,669 points
7970X dual channel .............. 3,522 points (-4 %)​

A more traditional image rendering benchmark for comparison,
Blender 3.6.5 LTS Barbershop,

7980X quad channel ............. 197 s
7980X dual channel .............. 197 s
7970X quad channel ............. 320 s
7970X dual channel .............. 319 s​

________
Normally I don't watch stuff like that, but the video came up when I looked for something else and the dual channel chapter title seemed interesting.

 

[eek2121]

This edition, even though memory bound, is a much better benchmark among renderers. Regardless of single core or multi core, all of the CPU works in tandem to render a SINGLE tile, unlike the previous versions. It is closer to how MT scaling works in the real world. Kind of like Geekbench 6.

I have to disagree with this statement. Many multithreaded applications (like your web browser) do the opposite approach…like your web browser. (1 process per tab…50 tabs, 50 processes).

As a developer, as long as I can successfully spin a workload safely into a new process, I will do it.

Even for the workloads GB6 is supposed to be measuring, many of them do scale beyond what GB6 will do.

What is really needed is a hybrid score that can represent both.

Note that I am investigating the possibility of building an open source benchmark similar to GB 5/6.

 

[tamz_msc]

I have to disagree with this statement. Many multithreaded applications (like your web browser) do the opposite approach…like your web browser. (1 process per tab…50 tabs, 50 processes).

Tabs that are not active are usually put to sleep on any modern browser.

 

[igor_kavinski]

Tabs that are not active are usually put to sleep on any modern browser.

Except Firefox. Maybe more a badly designed Javascript infested web page issue but it's there. Often imgur.com will spike the CPU utilization up to 100% when no imgur tab is currently open. Seems the JS process is not killed when the imgur tab is closed.

 

[coercitiv]

I have to disagree with this statement. Many multithreaded applications (like your web browser) do the opposite approach…like your web browser. (1 process per tab…50 tabs, 50 processes).

Loading a web page in Chrome in 2015, on a phone:

 

[Timur Born]

FYI: Cinebench R24 causes less load/stress than R23 on my undervolted 13900K and both do *not* trigger the AVX offset.




CB R23 in comparison:

As seen in my screenshots R24 doesn't hit the 253 W power-limit on my CPU while R23 brushes against it and slightly crosses it when CPU temps increase.

And it seems like the UI thread was optimized. You don't have to minimize the window anymore to keep the UI from constantly utilizing a CPU core (aka using two cores when measuring one core). Now the results with window open seem to match the results with window minimized. My last scores screenshot was done with minimized window, here is one with open window:

 

[Abwx]

According to Techspot/HUB, scaling is 48x for the 7980X in CB 2024. Anandtech shows 52x scaling in CB R23.

Those numbers are wrong, scaling should be measured at same frequency, using a 5GHz+ ST test to compute scaling with a MT score at 4GHz is forcibly terribly inaccurate, real number is at 0.7-0.75x, wich is still not impressive but surely better than GB5/6.

 

[Nothingness]

Even for the workloads GB6 is supposed to be measuring, many of them do scale beyond what GB6 will do.

What is really needed is a hybrid score that can represent both.

What is needed is that people stop looking at an aggregated score as if it could represent the real performance of every workload.

Note that I am investigating the possibility of building an open source benchmark similar to GB 5/6.

That'd be very interesting but quite a demanding task. And a frustrating one as you'll have people criticizing your work because it doesn't match what they believe and their biases. Or just because they're convinced they know so much better than you.

 

[Timur Born]

Could be the higher RAM bandwidth that helps the CB2024 renderer.

My results were done on tight 5600 MT and I was under the impression that the scores are high enough to not be memory or Ring bottlenecked. This might also depend on the number of cores/threads being served by the memory, so those quad channel CPUs might benefit more from memory bandwidth because of their high core-count.

 

[igor_kavinski]

so those quad channel CPUs might benefit more from memory bandwidth because of their high core-count.

Yeah. Those beasty cores want moar data to munch on. Makes sense that more memory channels feed them better.

 

[Timur Born]

This is the result of a Youtuber running 8600 MT and 50x Ring on a 14900K at 57/44x:


Compared to my 2334 pts. at 5600 MT and 45x Ring on a 13900K at 55/43x this may indicate a very small memory and/or Ring based improvement, but most of it still seems CPU frequency based.

2422 / 2334 = 1.038
57/55 = 1.036 (P cores carry a bit more than half the load)
44/43 = 1.023 (E cores are inefficient for floating-point load)

 

[igor_kavinski]

And to think that single core CPUs used to be a quarter of that.

 

[Iron Woode]

I upgraded my CPU to an R7 5700x:

 

[mikegg]

I have to disagree with this statement. Many multithreaded applications (like your web browser) do the opposite approach…like your web browser. (1 process per tab…50 tabs, 50 processes).

Not a good example. The use case for browser tabs is that once you load the page, it usually does not do any more computing work. If you have 50 tabs, 49 of them are using 0 CPU. Occasionally, you might have a tab with an active compute need but it's usually very low.

People need more RAM and single-threaded speeds for browsers for the exact reasons mentioned above. Multithread performance is almost never a problem for browsers.

Heck, a very fast ST dual-core CPU with a lot of RAM is probably better than a CPU with 64 slow cores for web browsing.

 

[Timur Born]

CB24's GPU test turns out to be quite useful for quickly finding a basic GPU undervoltage frequency. It crashes or produces bad pixels rather quickly. It seems to be all CUDA load at maybe 60% power usage, but it qualifies as 100% GPU load and adds a useful tool to the box.

 

[moinmoin]

Not a good example. The use case for browser tabs is that once you load the page, it usually does not do any more computing work.

Yeah, I agree adblocking is great and highly recommended. But when not doing so you absolutely have tabs wasting compute just for cycling through useless ads.

 

[mikegg]

Yeah, I agree adblocking is great and highly recommended. But when not doing so you absolutely have tabs wasting compute just for cycling through useless ads.

The point is that the old GB5 does not measure how fast web browsing with multiple tabs is, like @eek2121 claims here.

I have to disagree with this statement. Many multithreaded applications (like your web browser) do the opposite approach…like your web browser. (1 process per tab…50 tabs, 50 processes).

What GB6 actually needs is a separate MT test that is more like GB5 MT. Both MT use cases are valid. But the GB6 MT is much more common for consumers - whom GB targets.

 

[FlameTail]

What GB6 actually needs is a separate MT test that is more like GB5 MT. Both MT use cases are valid. But the GB6 MT is much more common for consumers - whom GB targets.

EXACTLY.

I was thinking this very same thing the other day.

Geekbench needs two types of MT tests.

 

[FlameTail]

Nowadays the only thing I use GB6 MT is in comparing phone CPUs.

For laptop/desktop chips, CB2024 MT is my go to benchmark now. And it's not flawed like previous Cinebench versions which did not support ARM libraries.

 

[qmech]

Yeah, I agree adblocking is great and highly recommended. But when not doing so you absolutely have tabs wasting compute just for cycling through useless ads.

It's not necessarily cycling through the ads, it's that the ads contain video or audio (even if muted) which suppresses the browser's background tab deprioritization.

 

[moinmoin]

It's not necessarily cycling through the ads, it's that the ads contain video or audio (even if muted) which suppresses the browser's background tab deprioritization.

Indeed, an many sites cycling through ads over time ensures that such imo exploitative ads can randomly appear and disappear at any time.

 

[Det0x]

*double post*