Value of Hardware Unboxed benchmarking

[Jupiter]

Linus is by far the worst. He also did a lot of advertising for multiple scam products like that molecular food scanner.

Nevertheless, Hardware Unboxed and Gamers Nexus are also not very good.

What Gamers Nexus said about the RTX 4070 Ti was pure polemic. I don't bother with something like that anymore. I'd much rather get my information from more neutral sources like Der Bauer, DF PCGH or igorsLAB.

 

[Arun]

We shouldn’t confuse instruction scheduling with top level scheduling, they have nothing to do with each other, the changes in Kepler were for the former, not the latter, and NVIDIA did *not* revert those changes in any way.

Also, while GCN to RDNA2 are fully HW-driven for instruction scheduling, RDNA3 is actually a step towards instruction scheduling being SW-driven via the S_DELAY_ALU instruction, but it is still more HW-heavy than NVIDIA as it works at the level of instruction delays rather than cycle delays, which does have the benefit it’s more precise and will waste fewer cycles with certain kinds of false dependencies (I’m happy to explain that in more detail elsewhere if anyone wants me to, but this would be too off-topic here).

Note that in terms of power efficiency, my personal experience is that e.g. register caching (introduced in Maxwell & RDNA3) is a lot more important than HW vs SW instruction scheduling (at least on modern process nodes where wire energy and RAM energy have gone up relatively speaking so the logic power of that scheduling isn’t so horrible in the grand scheme of things).

—

For Kepler/Maxwell instruction level scheduling, see Scott Gray’s analysis for maxas: https://github.com/NervanaSystems/maxas/wiki/Control-Codes

For Volta & all recent NVIDIA GPUs, see Section 2.1 of https://arxiv.org/abs/1804.06826

For RDNA3: Section 5.7 of https://www.amd.com/system/files/Te...uction-set-architecture-feb-2023_0.pdf#page52

—

Instruction scheduling being HW or SW (or a mix of the two) has zero impact on driver overhead (except maybe a tiny impact on compile times but it really shouldn’t be exaggerated).

I’m not an expert on how top level scheduling works for IHVs on these APIs (except on Imagination/PowerVR HW which is completely different) but any video or website that connects the Kepler changes with driver overhead probably has no idea what they are talking about on the subject, and should likely be ignored completely regarding the causes of driver overhead in general.

 

[NightAntilli]

We shouldn’t confuse instruction scheduling with top level scheduling, they have nothing to do with each other, the changes in Kepler were for the former, not the latter, and NVIDIA did *not* revert those changes in any way.

Also, while GCN to RDNA2 are fully HW-driven for instruction scheduling, RDNA3 is actually a step towards instruction scheduling being SW-driven via the S_DELAY_ALU instruction, but it is still more HW-heavy than NVIDIA as it works at the level of instruction delays rather than cycle delays, which does have the benefit it’s more precise and will waste fewer cycles with certain kinds of false dependencies (I’m happy to explain that in more detail elsewhere if anyone wants me to, but this would be too off-topic here).

Note that in terms of power efficiency, my personal experience is that e.g. register caching (introduced in Maxwell & RDNA3) is a lot more important than HW vs SW instruction scheduling (at least on modern process nodes where wire energy and RAM energy have gone up relatively speaking so the logic power of that scheduling isn’t so horrible in the grand scheme of things).

—

For Kepler/Maxwell instruction level scheduling, see Scott Gray’s analysis for maxas: https://github.com/NervanaSystems/maxas/wiki/Control-Codes

For Volta & all recent NVIDIA GPUs, see Section 2.1 of https://arxiv.org/abs/1804.06826

For RDNA3: Section 5.7 of https://www.amd.com/system/files/Te...uction-set-architecture-feb-2023_0.pdf#page52

—

Instruction scheduling being HW or SW (or a mix of the two) has zero impact on driver overhead (except maybe a tiny impact on compile times but it really shouldn’t be exaggerated).

I’m not an expert on how top level scheduling works for IHVs on these APIs (except on Imagination/PowerVR HW which is completely different) but any video or website that connects the Kepler changes with driver overhead probably has no idea what they are talking about on the subject, and should likely be ignored completely regarding the causes of driver overhead in general.

What's the reason for nVidia's 4000 series having a higher overhead than AMD's 7000 series?

 

[DegustatoR]

What's the reason for nVidia's 4000 series having a higher overhead than AMD's 7000 series?

It doesn't. Each case must be reviewed on its own as its app specific and is likely a result of the differences such apps have in rendering on different IHVs.

 

[NightAntilli]

It doesn't. Each case must be reviewed on its own as its app specific and is likely a result of the differences such apps have in rendering on different IHVs.

If there was ever a cherry-picking contest, you would be the global champion.

 

[DegustatoR]

If there was ever a cherry-picking contest, you would be the global champion.

The "cherry-picking contest" is what HUB did back when they've claimed this bs with higher driver overhead. This has been disproven time and time again by their own data. What I've posted is just the latest from that.

 

[NightAntilli]

The "cherry-picking contest" is what HUB did back when they've claimed this bs with higher driver overhead. This has been disproven time and time again by their own data. What I've posted is just the latest from that.

Really? How was this disproven? Let's go back to your own video shall we...

So apparently Nvidia doesn't have "a driver overhead problem due to s/w scheduling" (or w/e b.s. Steve said back when he benched five games and made wide reaching claims).

The video clearly shows the opposite of what you are saying. Let's just put some basics here...

1) Your framerate is primarily capped by either the CPU or the GPU, provided there are no software-induced limits like frame ratecaps etc.
2) The game engine is either more CPU heavy or GPU heavy at any given point in time.
3) The higher your resolution, the more likely that your GPU will become the bottleneck
4) The lower the resolution, the more likely the CPU will be the bottleneck
5) Considering that the CPU does not care about resolution, all else being equal, it is the driver of the GPU that limits how many frames the CPU can at most handle.
6) If the game engine is low enough on CPU usage to not limit the GPU in any way, then the power of the GPU will be visible at all resolutions, including 1080p and 1440p.


Hogwarts Legacy: 7900XTX Faster than 4090 at 1080p and 1440p, but over 25% faster at 4K. Same RT performance at 1080, but slower 7900XT at 1440p and 4K. -> Clearly the 4090 is inducing more CPU resources than the 7900XTX.
Spider-Man Remastered: 7900XTX Faster than 4090 at 1080p, about equal at 1440p but with much better 1% lows than 4090, but over 35% faster at 4K. Generally equal at RT except 4k-> Clearly the 4090 is inducing more CPU resources than the 7900XTX.
Far Cry 6: About equal at 1080p but much better 1% lows for 7900XTX, 7900XTX faster at 1440p, 7900XTX slightly slower at 4k -> Clearly the 4090 is inducing more CPU resources than the 7900XTX, limiting even its speed at 4K.
Hitman 3: 4090 faster at all resolutions -> Low CPU usage game.
Horizon Zero Dawn: 4090 faster at all resolutions -> Low CPU usage game
Rainbow Six Siege (DX11): 4090 faster at all resolutions -> It's quite well-known that nVidia had the lower driver overhead in DX11. This switched with DX12 in most cases.

The rest are all pretty much a variant of the above. The console ports generally use the 4090 very well at 1080p, i.e. no CPU bottleneck at all. Something like CoD MWII the 4090 is clearly more CPU limited than the 7900XTX is, just like Hogwarts Legacy. It's also quite clear when the 4090 is being CPU bottlenecked when the 7800X3D is giving it more performance than the i9.


So tell me...

Is the 7900XTX just as strong as the 4090
or
does the RTX 4090 have driver overhead?

You can't have it both ways.

 

[DegustatoR]

The video clearly shows the opposite of what you are saying.

This video shows exactly what I'm saying. It is you who can't seem to see what it's showing because this goes against your agenda.

 

[Pete]

I miss Tech Report’s super detailed frametime + power draw GPU reviews. The power draw figures would be a neat sanity check (and possible substitute for CPU usage graphs) for outlier results.

 

[trinibwoy]

I miss Tech Report’s super detailed frametime + power draw GPU reviews. The power draw figures would be a neat sanity check (and possible substitute for CPU usage graphs) for outlier results.

I think we just need better instrumentation and metrics reporting through the GPU hardware, driver and OS stack. The GPU usage stats reported by rivatuner etc are really misleading.

 

[DegustatoR]

I think we just need better instrumentation and metrics reporting through the GPU hardware, driver and OS stack. The GPU usage stats reported by rivatuner etc are really misleading.

We have profiling tools to look beyond that and they are available for some time now but are IHV specific and require a deeper knowledge than what most YT reviewers have.

 

[Scott_Arm]

We have profiling tools to look beyond that and they are available for some time now but are IHV specific and require a deeper knowledge than what most YT reviewers have.

intel is planning on bringing more information into their branch of presentmon that would make it kind of like a light profiler to identify bottlenecks. Not sure how deep they could really go.

 

[troyan]

AMDunboxed claims that the 7800XT is the "obvious choice" now. Maybe someone can tell them that sub 60FPS in 1080p is not a great pc gaming experience: https://www.kitguru.net/components/graphic-cards/dominic-moass/amd-rx-7800-xt-review/26/

 

[Kaotik]

AMDunboxed claims that the 7800XT is the "obvious choice" now. Maybe someone can tell them that sub 60FPS in 1080p is not a great pc gaming experience: https://www.kitguru.net/components/graphic-cards/dominic-moass/amd-rx-7800-xt-review/26/

Right, so according to you most of the cards out there, including big portion of RTX 40 line up, don't offer great pc gaming experience. Right on. I wonder why they still keep that old options screen in the games btw, since no-one can't use anything but maxed out settings.

 

[Seanspeed]

AMDunboxed

Good lord. Not that this wouldn't be childish in any effort, but this is not even a remotely clever attempt at making up some demeaning nickname. Super forced and laughable, even ignoring that it's wholly undeserved.

As for the rest of your comment and your 'supporting evidence' being a cherry-picked link to a benchmark only for a very specific game with specific settings, I cannot understand how anybody is supposed to take this seriously.

Things aren't great for GPU's in general right now, but they had an entirely compelling supporting argument for their claim:



For a new, modern gen GPU, it is far and away the best option offered so far. It just is, no matter how much it pains you to admit it.

 

[troyan]

Hm, most cards are either cheaper than $500 or not relevant any more as a new buy.

But lets focus on older games:
Cyberpunk, nearly three years old and is getting a DLC in a few week - sub 60 FPS in 1080p.
Dying Light 2, nearly two years old, under 60 FPS in 1080p.
Marvel and Guardians fan? 8x FPS in 1080p (4070 is nearly 50% faster...).
Still have to finish Shadow the Tomb Raider? 4070 is 36% faster.

Big Spider-Man fan with every comic since Amazing Fantasy 15? 4070 is 30%+ faster in Miles Morales.
Harry Potter fan? In Legacies under 60 FPS in 1080p.

Have a 4K display for home office and wont buy a new one? DLSS performance is a viable option.

How exactly is this product for $499 the "obvious choice"? And no, reducing image quality is not an argument.

 

[Seanspeed]

Hm, most cards are either cheaper than $500 or not relevant any more as a new buy.

But lets focus on older games:
Cyberpunk, nearly three years old and is getting a DLC in a few week - sub 60 FPS in 1080p.
Dying Light 2, nearly two years old, under 60 FPS in 1080p.
Marvel and Guardians fan? 8x FPS in 1080p (4070 is nearly 50% faster...).
Still have to finish Shadow the Tomb Raider? 4070 is 36% faster.

Big Spider-Man fan with every comic since Amazing Fantasy 15? 4070 is 30%+ faster in Miles Morales.
Harry Potter fan? In Legacies under 60 FPS in 1080p.

Have a 4K display for home office and wont buy a new one? DLSS performance is a viable option.

How exactly is this product for $499 the "obvious choice"? And no, reducing image quality is not an argument.

Surely we cant be expected to engage with clearly bad faith nonsense like this, right?

The level of cherry-picking you're doing is genuinely.....I cant even put it nicely enough to say, honestly.

 

[troyan]

They benched Hogwarts Legacies without Raytracing. Can you explain, why somebody wouldnt play this game with 74FPS on a 4070 but has no problem with Starfield running with 76FPS on a 7800XT?

 

[Subtlesnake]

How exactly is this product for $499 the "obvious choice"? And no, reducing image quality is not an argument.

Right, gamers aren't allowed to use non-Ultra settings or they get cast into hell.

 

[Deleted member 2197]

Good lord. Not that this wouldn't be childish in any effort, but this is not even a remotely clever attempt at making up some demeaning nickname. Super forced and laughable, even ignoring that it's wholly undeserved.

As for the rest of your comment and your 'supporting evidence' being a cherry-picked link to a benchmark only for a very specific game with specific settings, I cannot understand how anybody is supposed to take this seriously.

Things aren't great for GPU's in general right now, but they had an entirely compelling supporting argument for their claim:

View attachment 9545

For a new, modern gen GPU, it is far and away the best option offered so far. It just is, no matter how much it pains you to admit it.

Is it a better option than the RX 6800 XT? Usually the fidelity settings highly affect the overall price/performance result and will vary from review to review.

AMD Radeon RX 7800 XT im Test: Mit 16 GiB zum Sieg gegen die Geforce RTX 4070?

Preis-Leistung (Rasterizing & Raytracing) & PCGH-Fazit

www.pcgameshardware.de