Hi Lev!
Thanks very much. As a bit of an addendum, I’ve been looking at the stutter that occurs as well, and I think it’s a symptom of the same problem. Beyond the cards being sort of “globally” underfed, there are moments where the pipeline just freezes up. I wanted to get more data to look at this more closely, but alas. The following table is from my last few test runs:
| capture | missions | gameplay | frames | median | 3× threshold | ≥3× | ≥2× | ≥33ms | ≥50ms | ≥100ms |
|---|---|---|---|---|---|---|---|---|---|---|
| Test 33 | 3 | 49.6 min | 275,155 | 10.77ms | 32.3ms | 85 | 528 | 75 | 6 | 0 |
| Test 36 | 4 | 60.7 min | 365,074 | 9.79ms | 29.4ms | 79 | 703 | 42 | 6 | 0 |
| Test 37 | 4 | 93.7 min | 545,148 | 10.12ms | 30.4ms | 159 | 1,148 | 105 | 20 | 1 |
So, 323 “events” total that are 3x median frame time or more. If we include frames at double the median frame time, we have 2379 events, or 11.7 per minute. 3x median frames occur on average 1.5 times a minute, which lines up roughly with my “feeling” from playing. One note of interest is that in test 33 I played with higher graphics settings than I normally do. Test 33 ran at 1.71 stutters/min against Test 37’s 1.70 — despite test 33 having 28W more power and 400MHz more clock on average. (Test 36 came in lower at 1.30, so there’s run-to-run variation, but it’s interesting that pushing the card harder didn’t reduce the stutter rate.)
Each event shows a similar pattern of behaviour, and this pattern is preserved across vendor (5070ti and 3070ti) with a few caveats (different boost behaviour). Here’s my 9070xt (when looking at the % of frame that is work, keep in mind that the card is underclocking heavily to match the load). Table is averaged over these 3x median frame time events.
| frame offset | frame time | queue depth | render→present | % of frame that is GPU work |
|---|---|---|---|---|
| −8 | 11.42 | 0.91 | 10.64 | 91% |
| −4 | 11.55 | 0.90 | 10.38 | 90% |
| −2 | 11.87 | 0.87 | 10.65 | 90% |
| −1 | 11.53 | 0.92 | 10.96 | 95% |
| 0 | 34.47 | 0.33 | 11.73 | 34% |
| +1 | 11.24 | 1.19 | 13.44 | 120% |
| +2 | 12.75 | 1.05 | 12.93 | 101% |
| +4 | 12.05 | 1.20 | 13.84 | 115% |
| +8 | 12.26 | 1.24 | 14.04 | 115% |
Things look reasonably stable, then a big spike in frame time, yet the render > present time comparatively moves little. The card’s not slower, it just has nothing to chew on at that moment. We can see the queue depth is much shallower. The following frames are back to normal and come with a chunkier render queue the card works on over the next several frames. The more severe the hang, the more noticeable this behaviour. See following table.
| Severity band | n | median frame | median GPU work | % of frame that is GPU work |
|---|---|---|---|---|
| 3–4× | 261 | 33.6ms | 11.7ms | 34.6% |
| 4–5× | 30 | 43.4ms | 12.3ms | 27.2% |
| 5–7× | 20 | 57.7ms | 11.5ms | 20.4% |
| 7×+ | 12 | 75.3ms | 8.9ms (the mean is 11.6ms. Weird spread of values, and few samples.) | 10.6% |
Nvidia is the same, though they don’t downclock in response to the low load, so the 5070 devours what it gets lightning fast and spends the rest of the time idle. 31.74ms frame time and the card renders the frame and passes it off in 2.64ms. 29ms of idle time! Keep in mind this is on a 1080p monitor (I know, I know), so the GPU is absolutely not the weak link. It’s paired with a 7800X3D as well, so hard to say the CPU is underspec either. It also suffers a higher stutter rate than me (2.2 3x median frames per minute vs my 1.7), although it is a default .ini file.
5070ti:
| frame offset | frame time | queue depth | render→present | % of frame that is GPU work |
|---|---|---|---|---|
| −8 | 8.26 | 0.300 | 2.44 | 30% |
| −4 | 8.14 | 0.314 | 2.62 | 32% |
| −2 | 8.55 | 0.281 | 2.47 | 29% |
| −1 | 8.36 | 0.311 | 2.60 | 31% |
| 0 | 31.74 | 0.080 | 2.64 | 8% |
| +1 | 8.80 | 0.270 | 2.52 | 29% |
| +2 | 10.02 | 0.250 | 2.51 | 25% |
| +4 | 9.23 | 0.270 | 2.44 | 26% |
| +8 | 8.76 | 0.271 | 2.40 | 27% |
| Severity band | n | median frame | median GPU work | % of frame that is GPU work |
|---|---|---|---|---|
| 3–4× | 63 | 27.6ms | 2.5ms | 9.0% |
| 4–5× | 18 | 34.6ms | 2.8ms | 8.4% |
| 5–7× | 21 | 51.4ms | 3.2ms | 6.1% |
| 7×+ | 7 | 63.6ms | 2.7ms | 3.8% |
The 5070ti is barely breaking a sweat. Referring to my earlier post, the thing averages under half its power limit, yet it’s still spending two thirds of its time idling and waiting on work, and has the same frame time stutter as I do on my 9070xt. The 3070 Ti (1440p) is the same as the 5070 over three missions: 93 events, queue depth 0.85 > 0.20, render-present 14.1ms > 11.5ms while frame time went 14.6ms > 58.0ms. GPU work drops from 85% of the frame to 20%. The 3070 IS working hard - it averaged 241w out of 290 (you can see it in my original post) and still has the stutter.
So yeah - an addition to the power draw stuff in my original post. I’m pretty sure there’s an issue here with the stutter. Both it and AND powerdraw seem to affect both vendors equally (provided the AMD user uses Vizra’s config. Without it… hell nah). Again, I’m not a pro, happy for criticism, but the rate the 3x median stutters occur seems about right in my experience. Once or twice a minute a bit of a freeze. The 2x median would also explain why it just doesn’t feel nice even with the framerate is apparently high. I used to play on a 3050 laptop, so aiming for 60fps was about the best I could hope for in games. Here I’m apparently averaging 90, yet it feels rough.
Hope this is in someway helpful to someone. I’m going to go kill some heretics.
Edit: One thing I haven’t included is correlation with heavy combat. Mostly because this is not a game where heretics give you time to note down the start time of a particularly nasty encounter so I don’t have good data. But my theory is that the heavier the combat, the more stressed the engine is and it stands to reason that the stutter would also be worse during periods of intense activity.
The dark ritual map has that cinematic when you enter the cathedral. I can find in the log and sync with the performance data. Safe to assume everything after that point is pretty intense. Sure enough, for me and the 3070 the number of 2x median frames roughly doubled in 3/4 sets of data. Both times for me, once for the 3070, and in one run he was “stable” at his normal stutter rate.
Another moment I had was all 3 other players crashed in the middle of a fairly spicy Auric Maelstrom run and I had to survive alone while they restarted the game and rejoined. Can use the disconnect/rejoin lines in the console log to identify the time window. Before and after this 90 second window I was getting about five 2x frametime events per minute, during it I was getting 16.
Again, not enough good data to make firm judgements. It’s just not enough confirmed time in combat to try make comments about the rarer 3x median frames. But if I were a betting man…
Does have the plus that all my tests have been done on the same version though.