Starfield CPU Benchmarks & Bottlenecks: Intel vs. AMD Comparison
video description
Date: 2023-09-05
Related videos
Comments and reviews: 20
aggies11
Hmm, the Ratio , while a nice visualization might be a bit misleading. Especially when we want to use descriptive terms like balanced . A 0.5 ratio looks nice and balanced, but semantically it's quite the opposite. A balanced ratio is closer to 1.0, which is not intuitively how one might think about it, especially with it visualized in the bar chart.
For me personally (and I'd guess for the average viewer also) when we use terms like balanced we'd want a roughly equal amount of CPU time and GPU time per frame, that sounds balanced and fair. But in this metric it's the opposite, we want the GPU work to be as close to 100% of the entire time spent on the frame, and the CPU work to be as small/short as possible. That intuitively just doesn't read as balanced .
There is also a redundancy in the ratio (both sides must add up to 1) which while nice from a math standpoint, but be extra information that can confuse the point. It's early days yet so I wouldn't be able to give a concrete suggestion about an improvement, only that maybe just a raw GPU efficiency percentage value instead of the ratio eg. 97% vs 0.97:0.03 . It is a tricky one though as that might be too close to Gpu utilization and lead to confusion. There is definitely an inherrent challenge with this data. We want the GPU to be spending as little time waiting for the rest of the system. I'd be tempted to want to use a metric like GPU Bottleneck percentage and then that 0.97:0.03 ratio would become 3% . Which indicates how much the rest of the system is holding back the GPU. But again semantics, GPU bottleneck ususally means the GPU is the limit, and so the terminology would be reversed which is confusing.
I very much look forward to how this metric involves. It's undeniably useful, but how to present it is definitely a new challenge.
reply
Hmm, the Ratio , while a nice visualization might be a bit misleading. Especially when we want to use descriptive terms like balanced . A 0.5 ratio looks nice and balanced, but semantically it's quite the opposite. A balanced ratio is closer to 1.0, which is not intuitively how one might think about it, especially with it visualized in the bar chart.
For me personally (and I'd guess for the average viewer also) when we use terms like balanced we'd want a roughly equal amount of CPU time and GPU time per frame, that sounds balanced and fair. But in this metric it's the opposite, we want the GPU work to be as close to 100% of the entire time spent on the frame, and the CPU work to be as small/short as possible. That intuitively just doesn't read as balanced .
There is also a redundancy in the ratio (both sides must add up to 1) which while nice from a math standpoint, but be extra information that can confuse the point. It's early days yet so I wouldn't be able to give a concrete suggestion about an improvement, only that maybe just a raw GPU efficiency percentage value instead of the ratio eg. 97% vs 0.97:0.03 . It is a tricky one though as that might be too close to Gpu utilization and lead to confusion. There is definitely an inherrent challenge with this data. We want the GPU to be spending as little time waiting for the rest of the system. I'd be tempted to want to use a metric like GPU Bottleneck percentage and then that 0.97:0.03 ratio would become 3% . Which indicates how much the rest of the system is holding back the GPU. But again semantics, GPU bottleneck ususally means the GPU is the limit, and so the terminology would be reversed which is confusing.
I very much look forward to how this metric involves. It's undeniably useful, but how to present it is definitely a new challenge.
reply
LaymensLament
Well, the reason people try to avoid cpu bottlenecks (and talk about that) is because historically cpus were fast enough to max out your gpu, for quite a while even with mid tier models. You could sometimes keep the same CPU for multiple generations of GPU upgrades and the GPU also always had been the more expensive part. This is illustrated by the fact that the metric is called gpu busy and not cpu busy (so we want the cpu to be good enough to fully load a gpu, while expecting the gpu to still be the primary bottleneck). Technically a balanced system would be GPU utilization 1 and then we have a CPU busy chart that is also a 1 (ps.: this would mean something like how often are cpu instructions piling up because the gpu is too slow to process them and also the whole thing is only a fictional idea of perfect load because it will always depend on the context and workload/game).
reply
Well, the reason people try to avoid cpu bottlenecks (and talk about that) is because historically cpus were fast enough to max out your gpu, for quite a while even with mid tier models. You could sometimes keep the same CPU for multiple generations of GPU upgrades and the GPU also always had been the more expensive part. This is illustrated by the fact that the metric is called gpu busy and not cpu busy (so we want the cpu to be good enough to fully load a gpu, while expecting the gpu to still be the primary bottleneck). Technically a balanced system would be GPU utilization 1 and then we have a CPU busy chart that is also a 1 (ps.: this would mean something like how often are cpu instructions piling up because the gpu is too slow to process them and also the whole thing is only a fictional idea of perfect load because it will always depend on the context and workload/game).
reply
aggies11
RE: CPU Bottleneck testing - since you have the data, what might be interesting to see is from the 13100F all the way up to the 13900K, how close do the absolute GPU Busy time values in ms themselves stay or change. If the GPU busy time is pretty much the same for for all the cpus, then technically one might actually be able to extrapolate the max performance of a GPU (for the given test settings) without actually needing the fastest CPU.
My suspicion would be that not only does a faster CPU give you a better GPU Busy ratio (by decreasing the amount of time it takes to complete CPU work) but it also might lower the GPU Busy ms also (perhaps the GPU can be more efficiently fed ?). But I wonder which is be relatively negligible, or enough to make a difference? It's always interesting to have more insight into exactly how a CPU can bottleneck/impact GPU performance.
reply
RE: CPU Bottleneck testing - since you have the data, what might be interesting to see is from the 13100F all the way up to the 13900K, how close do the absolute GPU Busy time values in ms themselves stay or change. If the GPU busy time is pretty much the same for for all the cpus, then technically one might actually be able to extrapolate the max performance of a GPU (for the given test settings) without actually needing the fastest CPU.
My suspicion would be that not only does a faster CPU give you a better GPU Busy ratio (by decreasing the amount of time it takes to complete CPU work) but it also might lower the GPU Busy ms also (perhaps the GPU can be more efficiently fed ?). But I wonder which is be relatively negligible, or enough to make a difference? It's always interesting to have more insight into exactly how a CPU can bottleneck/impact GPU performance.
reply
LeDooni
Your testing in general is very helpful, but your memory testing (the 5600 MT/s example) pretty clearly shows that lower latency has a very significant impact on performance. Why you don t at least test a decent 7600C36 kit on the 13th Gen Intel rigs is beyond me with how cheap those have become and same goes for Z790 2-DIMM boards that can reliably run such kits. Going further into manual tuning would still improve performance by quite a lot, but non-tuned 6000-7000 MT/s just isn t where we re at with Intel platforms anymore. Test the CPUs at their performance sweet spot, there s no need to have the same memory kit on each CPU if the architectures handle RAM differently. It would be an easy change that would represent the actual performance potential of those platforms. Thanks for the work!
reply
Your testing in general is very helpful, but your memory testing (the 5600 MT/s example) pretty clearly shows that lower latency has a very significant impact on performance. Why you don t at least test a decent 7600C36 kit on the 13th Gen Intel rigs is beyond me with how cheap those have become and same goes for Z790 2-DIMM boards that can reliably run such kits. Going further into manual tuning would still improve performance by quite a lot, but non-tuned 6000-7000 MT/s just isn t where we re at with Intel platforms anymore. Test the CPUs at their performance sweet spot, there s no need to have the same memory kit on each CPU if the architectures handle RAM differently. It would be an easy change that would represent the actual performance potential of those platforms. Thanks for the work!
reply
Cybernetic
Lots of work, which is cool. The problem is how crazy it seems, by the end of the video, that you spent this much effort and never once commented on how the Intel architecture smashes AMD architecture in a brand new massive game, regardless of how poorly optimized the game is. That's a pretty big story, I would think. The much-lauded X3D CPUs are being trounced by a 360 13700k. And yet in your entire conclusion section you dance around ever stating that Intel CPUs are the clear winner. It's actually super strange once I noticed it by that time in the video. Clear winner, not a single mention of current-generation AMD being beaten or Intel being the best on the market under a pretty demanding gaming load. Bizarrely absent of mention.
reply
Lots of work, which is cool. The problem is how crazy it seems, by the end of the video, that you spent this much effort and never once commented on how the Intel architecture smashes AMD architecture in a brand new massive game, regardless of how poorly optimized the game is. That's a pretty big story, I would think. The much-lauded X3D CPUs are being trounced by a 360 13700k. And yet in your entire conclusion section you dance around ever stating that Intel CPUs are the clear winner. It's actually super strange once I noticed it by that time in the video. Clear winner, not a single mention of current-generation AMD being beaten or Intel being the best on the market under a pretty demanding gaming load. Bizarrely absent of mention.
reply
PineappleOnPizza
I highly doubt that Bethesda will provide a big performance fix in a few days. Optimization should be and (before AAA garbage began) always was a huge part of the total development time, but I guess with all of these artificial frame rate boosting methods most publishers simply stopped caring about that and push the devs to be done much earlier. 60ish frames on a 3070 with a good CPU? Lmao. The game looks average at best and shouldn't run like it does and with all the loading screens it has it should not be the size of the game's universe that is at fault.
Gaming is in a shameful state and most of the AAA games should be avoided. Most aren't worth it (with some exceptions) and are a mess on release.
reply
I highly doubt that Bethesda will provide a big performance fix in a few days. Optimization should be and (before AAA garbage began) always was a huge part of the total development time, but I guess with all of these artificial frame rate boosting methods most publishers simply stopped caring about that and push the devs to be done much earlier. 60ish frames on a 3070 with a good CPU? Lmao. The game looks average at best and shouldn't run like it does and with all the loading screens it has it should not be the size of the game's universe that is at fault.
Gaming is in a shameful state and most of the AAA games should be avoided. Most aren't worth it (with some exceptions) and are a mess on release.
reply
TAP7a
If you were offering short term contracter position or wanted consulting, I'd be keen on having some talks about this and developing some more densely informative standard methodologies for looking at these new data sources. Already in just your hypothesising we've got use cases for regressions, distribution tests like K-S testing, obviously testing for differences in means, calculating kurtosis for frame times/rates would be a good figure to indicate how tailed the distribution is or more precisely how frequent the extremes are and how consistent the pacing is... would be good fun
reply
If you were offering short term contracter position or wanted consulting, I'd be keen on having some talks about this and developing some more densely informative standard methodologies for looking at these new data sources. Already in just your hypothesising we've got use cases for regressions, distribution tests like K-S testing, obviously testing for differences in means, calculating kurtosis for frame times/rates would be a good figure to indicate how tailed the distribution is or more precisely how frequent the extremes are and how consistent the pacing is... would be good fun
reply
Alexander
For a more robust test of the theory that the CPU is actually memory constraint, could you not pair a 13th gen Intel CPU with different kits of DDR4 and DDR5 RAM?
This would allow you to keep the CPU constant while manipulating memory, thus isolating the 'RAM effect'. I think this would be informative, as you did observe a substantial drop-off in FPS when DDR5 clock speeds went below 6000Mhz. This could be even more pronounced when using DDR4 kits. If I am not mistaken, 13th gen Intel CPUs can be combined with DDR4 and DDR5 compatible motherboards.
reply
For a more robust test of the theory that the CPU is actually memory constraint, could you not pair a 13th gen Intel CPU with different kits of DDR4 and DDR5 RAM?
This would allow you to keep the CPU constant while manipulating memory, thus isolating the 'RAM effect'. I think this would be informative, as you did observe a substantial drop-off in FPS when DDR5 clock speeds went below 6000Mhz. This could be even more pronounced when using DDR4 kits. If I am not mistaken, 13th gen Intel CPUs can be combined with DDR4 and DDR5 compatible motherboards.
reply
Zolt n
So imagine that you have a 10400f (because come on, a 6C 12T cpu is more than enough for comfortable gaming) and a 6700 XT (which is a mid-range GPU that should still hold up decently 1080p medium settings) and you get 40-45 FPS on 1080p low... That tells you the current state of game unoptimizations and heavy relying on the upscaling... This reminds me the PS4/PS5 scene where the 60FPS patches coming out from the community, so I expect lots of mods from the community here to optimize the game instead of the developers.
reply
So imagine that you have a 10400f (because come on, a 6C 12T cpu is more than enough for comfortable gaming) and a 6700 XT (which is a mid-range GPU that should still hold up decently 1080p medium settings) and you get 40-45 FPS on 1080p low... That tells you the current state of game unoptimizations and heavy relying on the upscaling... This reminds me the PS4/PS5 scene where the 60FPS patches coming out from the community, so I expect lots of mods from the community here to optimize the game instead of the developers.
reply
Thor
You wanted to highlight the differences in CPU in these charts - got it.
But for future GPU-Busy analyses I would much more be interested in Benchmarks with realistic resolution/details - answering the question: Is my CPU fast enought in this game to not significantly limit my GPU?
You could also think of a two-dimensional chart combining a selection of GPU(-classes) with the selection of CPUs, showing the trend, which CPU/GPU combinations are reasonable for this game... the GPUbusyKeepsTestersBusy -chart
reply
You wanted to highlight the differences in CPU in these charts - got it.
But for future GPU-Busy analyses I would much more be interested in Benchmarks with realistic resolution/details - answering the question: Is my CPU fast enought in this game to not significantly limit my GPU?
You could also think of a two-dimensional chart combining a selection of GPU(-classes) with the selection of CPUs, showing the trend, which CPU/GPU combinations are reasonable for this game... the GPUbusyKeepsTestersBusy -chart
reply
gamersnexus
What is really interesting is if you have an ARC GPU they actually have a GPU usage stat and a GPU Render/Comp usae stat which run differentl percentages. For example in some scenarios the GPU usage can be at 99% while the render usage could be at 90%. This paired with GPU busy, frametime and being able to get readouts for the read and write bandwidth on the GPU starts to paint a very interesting picture and it is awesome Intel gives access to those sensors on ARC GPU's. This needs to become standard.
reply
What is really interesting is if you have an ARC GPU they actually have a GPU usage stat and a GPU Render/Comp usae stat which run differentl percentages. For example in some scenarios the GPU usage can be at 99% while the render usage could be at 90%. This paired with GPU busy, frametime and being able to get readouts for the read and write bandwidth on the GPU starts to paint a very interesting picture and it is awesome Intel gives access to those sensors on ARC GPU's. This needs to become standard.
reply
LaymensLament
the memory part confused me. according to your overview a year ago i went with ddr4 4000 memory for my 12600k (there was no performance difference in games between ddr4 and 5). is there now a benefit to ddr5? dies something on the 13 gen change that makes ddr5 better for those cpu specifically? on a quick glance it doesnt seem that ddr5 is strictly performing better even tho it has higher clocks. is there anyway to understand when your cpu is ram limited independant of cross component testing?
reply
the memory part confused me. according to your overview a year ago i went with ddr4 4000 memory for my 12600k (there was no performance difference in games between ddr4 and 5). is there now a benefit to ddr5? dies something on the 13 gen change that makes ddr5 better for those cpu specifically? on a quick glance it doesnt seem that ddr5 is strictly performing better even tho it has higher clocks. is there anyway to understand when your cpu is ram limited independant of cross component testing?
reply
Zumzifero
On Intel CPUs, a testing with both DDR4 3600 and DDR5 6000 would be very interesting: there is a Huge difference between the Rizen 5800X3D and 7800X3D and memory could have some part in it
Also, the Intel 13k series seems to do better then the X3D, meaning the infinity cash is too small for the kind of data the game is loading and unloading and the infinity is not doing a very good job.
There is a lot of stuff worth digging in these banchmatk results
reply
On Intel CPUs, a testing with both DDR4 3600 and DDR5 6000 would be very interesting: there is a Huge difference between the Rizen 5800X3D and 7800X3D and memory could have some part in it
Also, the Intel 13k series seems to do better then the X3D, meaning the infinity cash is too small for the kind of data the game is loading and unloading and the infinity is not doing a very good job.
There is a lot of stuff worth digging in these banchmatk results
reply
Ricardo
One possible application of GPU busy for the home user is trying to get the most out of their graphics settings when CPU limited. When the distance from the GPU busy to Frame render time is large, meaning CPU limited, you can start cranking up the graphics settings so that GPU busy comes closer to Frame render time, while trying to keep FPS average the same. When you start dropping FPS then it would mean that you have gone over the sweet spot, I guess?
reply
One possible application of GPU busy for the home user is trying to get the most out of their graphics settings when CPU limited. When the distance from the GPU busy to Frame render time is large, meaning CPU limited, you can start cranking up the graphics settings so that GPU busy comes closer to Frame render time, while trying to keep FPS average the same. When you start dropping FPS then it would mean that you have gone over the sweet spot, I guess?
reply
TretrebuchPlay
No, it did officially launched. Just because Bethesda said it is early access, everyone is rolling with it? 1. september was the official launch day. They just call it early access, so people do not go up in arms for having to pay for the more expensive version, otherwise you have to wait for several days post launch to get access. WHY IS EVERYONE ACCEPTING THIS BLATANT ANTI CONSUMER PRACTICE?!?! I DON'T GET IT!
reply
No, it did officially launched. Just because Bethesda said it is early access, everyone is rolling with it? 1. september was the official launch day. They just call it early access, so people do not go up in arms for having to pay for the more expensive version, otherwise you have to wait for several days post launch to get access. WHY IS EVERYONE ACCEPTING THIS BLATANT ANTI CONSUMER PRACTICE?!?! I DON'T GET IT!
reply
KEVLAR2346
seems like gpu busy is going to give us, the viewers, a better understanding of what cpu's are going to bottleneck with new gpu launches. I'm excited to see more of these graphs especially with older cpu's on newer gpu's. For example, I'd like to see a 9900k with a 4070 and a 3080 ... something like that may have influenced my purchase rather than the a 4070 performs similarly to a 3080 argument.
reply
seems like gpu busy is going to give us, the viewers, a better understanding of what cpu's are going to bottleneck with new gpu launches. I'm excited to see more of these graphs especially with older cpu's on newer gpu's. For example, I'd like to see a 9900k with a 4070 and a 3080 ... something like that may have influenced my purchase rather than the a 4070 performs similarly to a 3080 argument.
reply
Dynasty2201
5600x and a 3080. DLSS mod installed. FPS is all over the place. 60-100 in dungeons and indoors, outside it can drop to the 40s easily at times especially during sunsets. Mostly medium settings, 1440p, 75% render. Absolute joke, and I don't see how Nvidia can help here as this is blatantly an AMD game sponsorship, but even high end AMD cards are having the same low FPS issues.
reply
5600x and a 3080. DLSS mod installed. FPS is all over the place. 60-100 in dungeons and indoors, outside it can drop to the 40s easily at times especially during sunsets. Mostly medium settings, 1440p, 75% render. Absolute joke, and I don't see how Nvidia can help here as this is blatantly an AMD game sponsorship, but even high end AMD cards are having the same low FPS issues.
reply
Mike
I think it's a better experience to be GPU-bound than CPU-bound. There is less stutter, maybe it shows up in frame time charts. It would be good to prove that either way. That leads to needing to know what the GPU Busy value is for a range of GPUs for a selected CPU. If buying a 13400 and a 4090 would result in being CPU-bound, is that a worse experience than buying a 4070?
reply
I think it's a better experience to be GPU-bound than CPU-bound. There is less stutter, maybe it shows up in frame time charts. It would be good to prove that either way. That leads to needing to know what the GPU Busy value is for a range of GPUs for a selected CPU. If buying a 13400 and a 4090 would result in being CPU-bound, is that a worse experience than buying a 4070?
reply
Lukas
Good to have an Intel CPU for Starfield. Seems to me that the game, other than most games, actually utilizes the E-Cores.
But what is it with the 7950X3D? Why is it even worse than the 7800X3D which is the same product line? Not to mention that the 7950X3D costs around 700 Euros while a 7800X3D costs the same as the 13700K around 380 Euros?
reply
Good to have an Intel CPU for Starfield. Seems to me that the game, other than most games, actually utilizes the E-Cores.
But what is it with the 7950X3D? Why is it even worse than the 7800X3D which is the same product line? Not to mention that the 7950X3D costs around 700 Euros while a 7800X3D costs the same as the 13700K around 380 Euros?
reply
Hammer
Is it possible for both the cpu and gpu to be running at 100% efficientcy? That would defy the there's always a bottleneck statment as your not bottlenecked, you're simply at max capacity of both components and neither is individually holding you back. I guess you could say you're equally bottlenecked, but that sounds rather preposterous.
reply
Is it possible for both the cpu and gpu to be running at 100% efficientcy? That would defy the there's always a bottleneck statment as your not bottlenecked, you're simply at max capacity of both components and neither is individually holding you back. I guess you could say you're equally bottlenecked, but that sounds rather preposterous.
reply
Add a review, comment
Other channel videos
