Intel is Hiding Data from Us: What GPU Benchmarks Don't Tell You
video description
Date: 2022-07-17
Related videos
Comments and reviews: 10
NoNae
Btw, I don't like the way you show frame times.
First fps is an inferior unit compared to ms. You would say that they only differ by 1/x but there is a second difference. FrameS per seconD vs secondS per framE. They both assume taht you measure something else. FPS implies that the constant workload are the #frames and you measure ms. Whereas SPF implies that the time is constant and frames are measured. Example measuring in FPS implies that it is useful to open a second instance of CSGO on my laptop to get an additional 40fps. For Blender this is accurate because I can render half of the frames on my laptop and the other half on my desktop. But that's not how CS works. You can not only add ms, it makes also sense to subtract them to compare performance. The difference between 1 and 21 fps is the same as between 200 and 220. The engaged reader can now compare the measurements in ms as an exercise.
But do not plot ms against frame number. These small spikes don't reflect the experience. Plot them against time. That way one 1000 ms frame cant be easily be overshadowed by 1000 1ms frames.
reply
Btw, I don't like the way you show frame times.
First fps is an inferior unit compared to ms. You would say that they only differ by 1/x but there is a second difference. FrameS per seconD vs secondS per framE. They both assume taht you measure something else. FPS implies that the constant workload are the #frames and you measure ms. Whereas SPF implies that the time is constant and frames are measured. Example measuring in FPS implies that it is useful to open a second instance of CSGO on my laptop to get an additional 40fps. For Blender this is accurate because I can render half of the frames on my laptop and the other half on my desktop. But that's not how CS works. You can not only add ms, it makes also sense to subtract them to compare performance. The difference between 1 and 21 fps is the same as between 200 and 220. The engaged reader can now compare the measurements in ms as an exercise.
But do not plot ms against frame number. These small spikes don't reflect the experience. Plot them against time. That way one 1000 ms frame cant be easily be overshadowed by 1000 1ms frames.
reply
Frosty
So speed sync where he says he doesn't know why people haven't done this before is essentially vsync with triple buffering, which has been done for decades. In triple buffering you have 3 frame buffers, one displays the latest frame to the monitor, and the gpu alternates writing to the 2 back buffers and when the monitor needs the next frame the latest full frame is put into the front buffer. Giving you zero tearing and less latency. I don't know why he's pretending like no one has done this before. From memory I think triple buffering is OpenGL only, but DirectX has something similar called render ahead which uses a fixed number of buffers you can specify.
reply
So speed sync where he says he doesn't know why people haven't done this before is essentially vsync with triple buffering, which has been done for decades. In triple buffering you have 3 frame buffers, one displays the latest frame to the monitor, and the gpu alternates writing to the 2 back buffers and when the monitor needs the next frame the latest full frame is put into the front buffer. Giving you zero tearing and less latency. I don't know why he's pretending like no one has done this before. From memory I think triple buffering is OpenGL only, but DirectX has something similar called render ahead which uses a fixed number of buffers you can specify.
reply
Eric
Another awesome lecture ! Putting the game pipeline vs time on the chart is one of the best way to visualise a bottleneck. Of course, it would be even better if the chart is even more granular like when the game pipeline goes into VRAM (via PCIe Rebar) for texture decompression and stuff like that and why Rebar buys such high framerate for Intel Arc vs AMD Navi. Is it all due to texture load/decomp via DirectStorage (or normal reads)? Or is it due to the Raytracing units coming into play (in reference to the F1 benchmark a few days ago)?
reply
Another awesome lecture ! Putting the game pipeline vs time on the chart is one of the best way to visualise a bottleneck. Of course, it would be even better if the chart is even more granular like when the game pipeline goes into VRAM (via PCIe Rebar) for texture decompression and stuff like that and why Rebar buys such high framerate for Intel Arc vs AMD Navi. Is it all due to texture load/decomp via DirectStorage (or normal reads)? Or is it due to the Raytracing units coming into play (in reference to the F1 benchmark a few days ago)?
reply
lordsqueak
Heh , the funny thing is , despite all these different technologies, the base problem is exactly the same, tearing. It's just different ideas on how to deal with it.
It was mentioned that some of these ideas seems obvious in hindsight, and it just occurred to me that despite all the work that has gone into it, we still haven't solved it to satisfaction.
I suspect tearing is never really going to get solved fully, but instead we will eventually get so high frame rates that it no longer matters.
reply
Heh , the funny thing is , despite all these different technologies, the base problem is exactly the same, tearing. It's just different ideas on how to deal with it.
It was mentioned that some of these ideas seems obvious in hindsight, and it just occurred to me that despite all the work that has gone into it, we still haven't solved it to satisfaction.
I suspect tearing is never really going to get solved fully, but instead we will eventually get so high frame rates that it no longer matters.
reply
periurban
Surely the system doesn't bottleneck if the performance request is within operating parameters? It only bottlenecks if the request is more than the system can deliver? No? This also applies to the choice between tearing and hitching. If the V-sync frame rate is well within capability there will be no hitching. It seems this argument will become redundant at framerates beyond 250fps at high resolutions where the requests are all within capability.
reply
Surely the system doesn't bottleneck if the performance request is within operating parameters? It only bottlenecks if the request is more than the system can deliver? No? This also applies to the choice between tearing and hitching. If the V-sync frame rate is well within capability there will be no hitching. It seems this argument will become redundant at framerates beyond 250fps at high resolutions where the requests are all within capability.
reply
NukeDispenser
When this guy is not finding excuses for why Intel GPUs had such a bad benchmark it's nice to listen to him, sadly he does that way too often, probably higher-ups marketing strategy.
Btw, why is this comment section filled with 90% copy-paste replies? I know there was almost nothing of substance in this video other than smooth-sync advert, but come on, it feels like bots wrote 90% of the comments.
reply
When this guy is not finding excuses for why Intel GPUs had such a bad benchmark it's nice to listen to him, sadly he does that way too often, probably higher-ups marketing strategy.
Btw, why is this comment section filled with 90% copy-paste replies? I know there was almost nothing of substance in this video other than smooth-sync advert, but come on, it feels like bots wrote 90% of the comments.
reply
MrDiscountNinja
Really enjoying these interview with Tom - definitely doing a lot to improve my opinion of intel. I just hope that we continue to get genuine conversations with their engineers and that PR or marketing doesn't start to get involved. I understand that even the engineers when speaking publicly need to have some degree of business sensitivity but the key for me is that the interaction is genuine.
reply
Really enjoying these interview with Tom - definitely doing a lot to improve my opinion of intel. I just hope that we continue to get genuine conversations with their engineers and that PR or marketing doesn't start to get involved. I understand that even the engineers when speaking publicly need to have some degree of business sensitivity but the key for me is that the interaction is genuine.
reply
junethefirst
Speed Sync sounds like Nvidia's Fast Sync, where the GPU goes wild and just sends the latest fully rendered frame when the fixed refresh moment arrives.
Smooth Sync is a real new thing. Maybe the dithering bands would be better if they were even taller but it's all about seeing it in movement.
In any case, good for Intel bringing both competition and new features to the table.
reply
Speed Sync sounds like Nvidia's Fast Sync, where the GPU goes wild and just sends the latest fully rendered frame when the fixed refresh moment arrives.
Smooth Sync is a real new thing. Maybe the dithering bands would be better if they were even taller but it's all about seeing it in movement.
In any case, good for Intel bringing both competition and new features to the table.
reply
Chival
Can someone explain why they need Smooth Sync if there's Speed Sync?
As I get it both technologies work on displays without adaptive sync support - Speed Sync picks up the last completed frame, Smooth Sync doesn't do anything about tears, it just masks them.
Isn't Smooth Sync inferior to Speed Sync approach in every way? Is it all about minimizing input lag?
reply
Can someone explain why they need Smooth Sync if there's Speed Sync?
As I get it both technologies work on displays without adaptive sync support - Speed Sync picks up the last completed frame, Smooth Sync doesn't do anything about tears, it just masks them.
Isn't Smooth Sync inferior to Speed Sync approach in every way? Is it all about minimizing input lag?
reply
Dirk
I remember YEARS ago when g-sync was first being advertised, I kept asking, Wait why don't we just, not show the tears and only the next fully rendered frame? and my classmates just told me if it were that easy they'd already be doing it hurrdurr. But to me, having to shell out for a fancy g-sync monitor just felt like a slap in the face. Turns out, I was right.
reply
I remember YEARS ago when g-sync was first being advertised, I kept asking, Wait why don't we just, not show the tears and only the next fully rendered frame? and my classmates just told me if it were that easy they'd already be doing it hurrdurr. But to me, having to shell out for a fancy g-sync monitor just felt like a slap in the face. Turns out, I was right.
reply
Add a review, comment
Other channel videos
