學習RDNA3精神-TIME SPY EXTREME
YUJIHW一共在TIME SPY EXTREME中嘗試了數十個不同的使用設定的組合, 並記錄低其中約二十個不同的使用設定在運行TIME SPY EXTREM (TSE)時的數據。
DEFAULT(500-3025//2498//100%//1.15v//fan46%)意思是DRIVER內的設定值/目標值, 第一樣要認識的地方就是DRIVER設定值絕對不等如實測中的實際值。
- DEFAULT是該組使用設定的稱呼。
- 500-3025所指的是GPU MIN CLOCK=500以及GPU MAX CLOCK=3025。
- 2498所指的是VRAM CLOCK。
- 100%所指的是POWER LIMIT。
- 1.15V所指的是VOLTAGE。
- FAN46%所指的是MAX FAN RPM目標是46%(即亦DEFAULT水平)。
部份使用設定包含VRAM FAST, 意思是將MEMORY TIMING設定為FAST, 沒有提及即是MEMORY TIMING維持在DEFAULT。
∇ 下面所收集的硬件數據全是HWINFO64內的最大實際值, 事實上的確會與在DRIVER內所設定的目標值有一定的距離。OC1/2/3/4….這些名字都是亂改的不必理會。
以下分段講解上面所列出的數據:
∇ OC1的目的是透過將GPU CLOCK(最小和最大)由預設的500-3025目標值增加至2500-3500目標值, 以及將VRAM從預設的2498目標值增加至2700目標值, 還有將POWER LIMIT由預設的100%增加至115%目標值後, 再觀察TIME SPY EXTREME GRAPHICS SCORE(TSE GS)分數有否合理地上升。115%的POWER LIMIT目標值使實際的最大功耗由378W急增至432W。實際的PEAK POWER一開始並沒有記錄下來, 但往後大概徘徊在500~560W(最大實際值)。GPU的最大實際頻率(2645)未有如DRIVER內設定的目標值(3500)般大幅地等比上升, 反映RX7900XTX的GPU頻率控制相當虛無縹緲, 其演算法過份深奧難明。溫度變化方面, 核心溫度還好, GDDR6溫度不理想, 而GPU HOT SPOT溫度亦有點太高。據AMD文件解釋, GPU HOT SPOT溫度安全上限約在110C, 而GPU頻率會因應HOT SPOT溫度而自動調整。
timespy extreme | stock (500-3025//2498//100%//1.15v//fan46%) |
oc1 (2500-3500//2700/115%//1.15v//fan46%) |
graphics score | 14522 | 15421 |
gpu temp | 57c | 60c |
memory temp | 82c | 84c |
gpu hot spot | 84c | 94c |
gpu fan | 1261rpm | 1496rpm |
total board power | 378w | 432w |
gpu power peak | 479w | |
front end clock | 2645mhz | 2692mhz |
front end effective | 2850mhz | 2914mhz |
shader clock | 2642mhz | 2698mhz |
shader effective | 2642mhz | 2698mhz |
memory clock | 2485mhz | 2686mhz |
core voltage | 0.921v | 0.941v |
gddr6 usage | 6701mb | 4316mb |
ddr5 usage | 10679mb | 11455mb |
gpu mcd6 | 68.2c | 71.7c |
∇ OC1由於在DRIVER內增加GPU MAX CLOCK的實際效果不算理想, 之後我們嘗試在OC2中再增加GPU MIN CLOCK(由2500提升至2900)和風扇轉速目標(由預設46%增加至70%), 但TSE GS反而倒跌。
OC3則將GPU MIN CLOCK由2900降至2800, OC3的TSE GS終於回升, 但仍然低於OC1(2500 MIN CLOCK)。我們認為DRIVER內的GPU MIN CLOCK設定不能太高, 2500頂盡。而風扇轉速對性能的影響似乎微乎其微。
timespy extreme | oc1 (2500-3500//2700/115%//1.15v//fan46%) |
oc2 (2900-3500//2700//115%//1.15v//fan70%) |
oc3 (2800-3500//2700//115%//1.15v//fan70%) |
graphics score | 15421 | 13711 | 14381 |
gpu temp | 60c | 53c | 54c |
memory temp | 84c | 78c | 78c |
gpu hot spot | 94c | 91c | 91c |
gpu fan | 1496rpm | 2248rpm | 2257rpm |
total board power | 432w | 433w | 435w |
gpu power peak | 554w | ||
front end clock | 2692mhz | 2874mhz | 2791mhz |
front end effective | 2914mhz | 3103mhz | 3012mhz |
shader clock | 2698mhz | 2869mhz | 2787mhz |
shader effective | 2698mhz | 2868mhz | 2787mhz |
memory clock | 2686mhz | 2686mhz | 2686mhz |
core voltage | 0.941v | 1.004v | 0.975v |
gddr6 usage | 4316mb | 4924mb | 6232mb |
ddr5 usage | 11455mb | 11245mb | 11378mb |
gpu mcd6 | 71.7c | 61.4c | 60.7c |
∇ 再之後的OC4和OC5分別透過設定不同的更低GPU MIN CLOCK, 成功將TSE GS創新高。我們再次發現將風扇轉速降回至預設狀態並未有對性能造成任何損失。我們最後得出在DRIVER內將GPU MIN CLOCK設定為1000~2000最為理想, 當中以2000最好。
OC6則嘗試再在DRIVER內增加GPU MAX CLOCK, 我們最終發現GPU MAX CLOCK從3500再增加至4000後系統會變得不穩定。所以GPU MAX CLOCK在DRIVER內最好設定在3000-3500之間。
還有, OC3與OC4相比, 兩者的GPU頻率接近一樣, 但是TSE GS分數差距甚大…看來在DRIVER內的GPU MIN CLOCK設定對實際性能有重大影響, 縱使看不出來。
此時我們建議將GPU MIN CLOCK設定為2000則可; MAX CLOCK則不要高於3500。
timespy extreme | oc3 (2800-3500//2700//115%//1.15v//fan70%) |
oc4 (2000-3500//2700//115%//1.15v//fan46%) |
oc5 (1000-3500//2700//115%//1.15v//fan46%) |
oc6 (2000-4000//2700//115%//1.15v//fan46%) |
graphics score | 14381 | 15485 | 15481 | failed |
gpu temp | 54c | 59c | 58c | failed |
memory temp | 78c | 84c | 84c | failed |
gpu hot spot | 91c | 90c | 91c | failed |
gpu fan | 2257rpm | 1492rpm | 1486rpm | failed |
total board power | 435w | 434w | 434w | failed |
gpu power peak | 554w | 510w | 557w | failed |
front end clock | 2791mhz | 2719mhz | 2775mhz | failed |
front end effective | 3012mhz | 2920mhz | 2920mhz | failed |
shader clock | 2787mhz | 2702mhz | 2703mhz | failed |
shader effective | 2787mhz | 2701mhz | 2703mhz | failed |
memory clock | 2686mhz | 2686mhz | 2686mhz | failed |
core voltage | 0.975v | 0.942v | 0.942v | failed |
gddr6 usage | 6232mb | 6934mb | 7551mb | failed |
ddr5 usage | 11378mb | 11413mb | 11438mb | failed |
gpu mcd6 | 60.7c | 69.9c | 70.1c | failed |
∇ 與OC5相比, OC7除了將GPU MIN CLOCK由1000增加至2000外, 還有將VRAM TIMING設定由DEFAULT改為FAST, 但TSE GS沒有改變。
此時我們不建議將VRAM TIMING由DEFAULT提升至FAST。
timespy extreme | oc5 (1000-3500//2700// 115%//1.15v//fan46%) |
oc6 (2000-4000//2700// 115%//1.15v//fan46%) |
oc7 (2000-3500//2700// 115%//1.15v//fan46%//vramfast) |
graphics score | 15481 | failed | 15479 |
gpu temp | 58c | failed | 60c |
memory temp | 84c | failed | 84c |
gpu hot spot | 91c | failed | 93c |
gpu fan | 1486rpm | failed | 1503rpm |
total board power | 434w | failed | 433w |
gpu power peak | 557w | failed | 544w |
front end clock | 2775mhz | failed | 2681mhz |
front end effective | 2920mhz | failed | 2912mhz |
shader clock | 2703mhz | failed | 2697mhz |
shader effective | 2703mhz | failed | 2696mhz |
memory clock | 2686mhz | failed | 2686mhz |
core voltage | 0.942v | failed | 0.941v |
gddr6 usage | 7551mb | failed | 4233mb |
ddr5 usage | 11438mb | failed | 10798mb |
gpu mcd6 | 70.1c | failed | 71.4c |
∇ 在OC8我們開始嘗試手動降壓, 在DRIVER內由預設的1.150V降至1.050V, 其餘不變。最終TSE GS再創新高, 雖然實際的最高核心電壓變化不大。OC9將風扇轉速目標由46%增加至75%, 但實際的溫度變化不大, 性能亦然。OC10將VRAM TIMING在DRIVER內從FAST降回至DEFAULT設定, TSE GS降了約80分。另外值得留意實際的核心頻率, 當實際的FRONT END CLOCK上升至2900或以上, TSE GS亦跟隨上升。
此刻我們建議手動在DRIVER內降壓, 而VRAM TIMING雖然開始對性能有影響但是仍然不太建議將其設定為FAST。另一調較重點是如何提高實際的FRONT END CLOCK。RX7900XTX的實際最高電壓似乎不太受DRIVER的目標電壓值設定所影響。
timespy extreme | oc7 (2000-3500//2700// 115%//1.15v//fan46%//vramfast) |
oc8 (2000-3500//2700// 115%//1.05v//fan46%//vramfast) |
oc9 (2000-3500//2700// 115%//1.05v//fan75%//fast timing) |
oc10 (2000-3500//2700// 115%/1.05v//fan75%) |
graphics score | 15479 | 16337 | 16345 | 16265 |
gpu temp | 60c | 59c | 55c | 54c |
memory temp | 84c | 84c | 80c | 80c |
gpu hot spot | 93c | 91c | 86c | 87c |
gpu fan | 1503rpm | 1486rpm | 2406rpm | 2403rpm |
total board power | 433w | 433w | 433w | 433w |
gpu power peak | 544w | 545w | 514w | 518w |
front end clock | 2681mhz | 2918mhz | 2932mhz | 2937mhz |
front end effective | 2912mhz | 3148mhz | 3177mhz | 3173mhz |
shader clock | 2697mhz | 2920mhz | 2945mhz | 2942mhz |
shader effective | 2696mhz | 2920mhz | 2945mhz | 2942mhz |
memory clock | 2686mhz | 2686mhz | 2686mhz | 2686mhz |
core voltage | 0.941v | 0.921v | 0.933v | 0.933v |
gddr6 usage | 4233mb | 4901mb | 4133mb | 4135mb |
ddr5 usage | 10798mb | 10668mb | 10709mb | 10723mb |
gpu mcd6 | 71.4c | 70.7c | 65.2c | 64.8c |
∇ OC11嘗試進一步下探降壓空間, 由1.05V再降至0.95V, 同時我們偷懶直接開啟VRAM FAST TIMING, 結果系統不穩。在OC12中將DRIVER內的電壓目標值從0.95V回增至1.05V, 系統仍然不穩定, 兇手似乎就是VRAM TIMING FAST。OC13是亂來的, 我們竟然將DRIVER內的GPU MAX CLOCK由3500再增加至3800並同時將VRAM由2700增加至2800, 結果死得更快。OC14將DRIVER內的頻率目標定在2000-3500//2700(VRAM), 但同時將POWER LIMIT由一直以來的115%降回至100%, 最後同樣不穩。
此時我們發現VRAM FAST不是好東西, 時得時唔得。而降壓(即從預設的1.15V目標值降至1.05V目標值)又不增加POWER LIMIT(不增加至115%)好像不太好。
timespy extreme | oc10 (2000-3500//2700// 115%/1.05v//fan75%) |
oc11 (2000-3500//2700// 115%//0.95v// fan75%//vramfast) |
oc12 (2000-3500//2700// 115%//1.05v// fan75%//vramfast) |
oc13 (2000-3800//2800// 115%//1.05v// fan75%//vramfast) |
oc14 (2000-3500//2700// 100%//1.05v// fan75%//vramfast) |
graphics score | 16265 | failed | failed | failed | failed |
gpu temp | 54c | failed | failed | failed | failed |
memory temp | 80c | failed | failed | failed | failed |
gpu hot spot | 87c | failed | failed | failed | failed |
gpu fan | 2403rpm | failed | failed | failed | failed |
total board power | 433w | failed | failed | failed | failed |
gpu power peak | 518w | failed | failed | failed | failed |
front end clock | 2937mhz | failed | failed | failed | failed |
front end effective | 3173mhz | failed | failed | failed | failed |
shader clock | 2942mhz | failed | failed | failed | failed |
shader effective | 2942mhz | failed | failed | failed | failed |
memory clock | 2686mhz | failed | failed | failed | failed |
core voltage | 0.933v | failed | failed | failed | failed |
gddr6 usage | 4135mb | failed | failed | failed | failed |
ddr5 usage | 10723mb | failed | failed | failed | failed |
gpu mcd6 | 64.8c | failed | failed | failed | failed |
∇ OC15將降壓的目標值由以往的1.05V回升至1.1V(仍然比預設的1.15V為低), 並將POWER LIMIT目標值由115%降回至預設的100%, OC15的TSE GS比不上OC1(1.15V 115%), 就算OC15有偷開FAST TIMING。OC16進一步將POWER LIMIT目標值降至90%, 與OC15的100% POWER LIMIT目標值相比, OC16的TSE GS更低。同樣維持在POWER LIMIT 90%目標值, OC17將電壓目標值由1.1V降回1.05V, TSE GS分數有所回升。
我們還是認為VRAM FAST不如不開。還有, OC10的TSE GS與OC14-17的TSE GS有一段距離。為了最高性能, POWER LIMIT目標值應該被設定為115%, 電壓目標值似乎亦是越低越好除非不穩。1.05V應該是個好目標。
如果有留意前面的TOTAL BOARD POWER(TBP), 應該都會發現除了DEFAULT外, 其餘的OC設定(1至10)不論VOLTAGE的目標值設定在多少, 實際的TBP清一色是~435W, 因為POWER LIMIT的目標都被設定為115%。要真正降功耗(比預設的實際水平370W更低), POWER LIMIT的目標值不應設定在100%以上。
timespy extreme | oc1 (2500-3500//2700// 115%//1.15v// fan46%) |
oc10 (2000-3500//2700// 115%/1.05v//fan75%) |
oc14 (2000-3500//2700// 100%//1.05v// fan75%//vramfast) |
oc15 (2000-3500//2700// 100%/1.1v// fan75%//vramfast) |
oc16 (2000-3500//2700// 90%//1.1v// fan75%//vramfast) |
oc17 (2000-3500// 2700//90%//1.05v// fan75%//vramfast) |
graphics score | 15421 | 16265 | failed | 15302 | 14689 | 15187 |
gpu temp | 60c | 54c | failed | 54c | 50c | 50c |
memory temp | 84c | 80c | failed | 80c | 76c | 76c |
gpu hot spot | 94c | 87c | failed | 81c | 72c | 72c |
gpu fan | 1496rpm | 2403rpm | failed | 1908rpm | 1819rpm | 1809rpm |
total board power | 432w | 433w | failed | 377w | 339w | 344w |
gpu power peak | 518w | failed | 502w | 549w | 532w | |
front end clock | 2692mhz | 2937mhz | failed | 2626mhz | 2570mhz | 2762mhz |
front end effective | 2914mhz | 3173mhz | failed | 2867mhz | 2728mhz | 2825mhz |
shader clock | 2698mhz | 2942mhz | failed | 2657mhz | 2534mhz | 2623mhz |
shader effective | 2698mhz | 2942mhz | failed | 2657mhz | 2533mhz | 2623mhz |
memory clock | 2686mhz | 2686mhz | failed | 2686mhz | 2686mhz | 2686mhz |
core voltage | 0.941v | 0.933v | failed | 0.874v | 0.827v | 0.815v |
gddr6 usage | 4316mb | 4135mb | failed | 4208mb | 4858mb | 5484mb |
ddr5 usage | 11455mb | 10723mb | failed | 10653mb | 10675mb | 10719mb |
gpu mcd6 | 71.7c | 64.8c | failed | 66.4c | 61.6c | 61.0c |
∇ 由於我們發現了實際的TBP在POWER LIMIT目標值都設定在115%的時候仍然會衝到435W(實際的TBP功耗), 其TSE GS分數介乎163XX~137XX不等, 所以YUJIHW認為如果TSE GS不能突破16000, 不如乖乖將實際的TBP盡量維持在DEFAULT時的實際TBP水平, 亦即~370W。換句話說如果RX7900XTX在預設下能夠跑出TSE GS 145XX而其實際的TBP是~380W, 那麼在TSE GS 145XX~153XX的區間中, 其實有辦法在實際的TBP維持在同一水平(~370W)甚至更低(~340W)的情況下做得到TSE GS ~15XXX。
所以由UV1起, 調整的方向會以不增加預設的POWER LIMIT目標值(100%)為先。故此我們在UV1內直接將POWER LIMIT目標值降至90%, 同時將降壓目標值維持在1.05V, 再降低GPU的目標頻率和VRAM的目標頻率至預設水平500-3000//2498, 結果UV1實際的TBP功耗首次做出最低的340W水平, 而UV1的TSE GS分數14818亦比DEFAULT時的14522略高。這種做法更像是在預設的實際TBP功耗水平(~370W)下細調各項目標設定以獲得更高的性能。
UV2在UV1的基礎上更改了GPU MAX CLOCK目標值, 從UV1的3000 GPU MAX目標值降至UV2的2500 GPU MAX目標值, 結果UV2的TSE進一步下降(比DEFAULT時更低), 可是UV2的實際TBP亦降至新低322W。
要真正使功耗變得好看, 例如低於300W TBP(預設狀態的實際TBP大約是380W), UV3這樣極度進取的做法以犧牲性能為手段, 將GPU MAX CLOCK目標值由之前的2500再降至2000, 同時再將電壓目標值由之前的1.05V降至1.00V, 最終做出實際的276W TBP(比預設狀態時低了約100W)而TSE GS是12455(比預設狀態時低了約2000分)。
timespy extreme | oc1 (2500-3500//2700// 115%//1.15v// fan46%) |
oc10 (2000-3500//2700// 115%/1.05v//fan75%) |
oc17 (2000-3500// 2700//90%//1.05v// fan75%//vramfast) |
uv1 (500-3000//2498/ /90%//1.05v// fan46%) |
uv2 (500-2500//2498// 90%//1.05v// fan46%) |
uv3 (500-2000//2498// 90%//1.00v// fan46%) |
ocpassed (2000-3100//2600// 115%//1.125v// fan75%) |
graphics score | 15421 | 16265 | 15187 | 14818 | 14299 | 12455 | 15612 |
gpu temp | 60c | 54c | 50c | 52c | 48c | 45c | 56c |
memory temp | 84c | 80c | 76c | 78c | 74c | 72c | 80c |
gpu hot spot | 94c | 87c | 72c | 75c | 70c | 61c | 87c |
gpu fan | 1496rpm | 2403rpm | 1809rpm | 1476rpm | 1481rpm | 1475rpm | 2407rpm |
total board power | 432w | 433w | 344w | 340w | 322w | 276w | 365w |
gpu power peak | 518w | 532w | 463w | 357w | 315w | 434w | |
front end clock | 2692mhz | 2937mhz | 2762mhz | 2753mhz | 2519mhz | 2029mhz | 2801mhz |
front end effective | 2914mhz | 3173mhz | 2825mhz | 2946mhz | 2722mhz | 2195mhz | 3030mhz |
shader clock | 2698mhz | 2942mhz | 2623mhz | 2732mhz | 2525mhz | 2030mhz | 2807mhz |
shader effective | 2698mhz | 2942mhz | 2623mhz | 2732mhz | 2524mhz | 2030mhz | 2807mhz |
memory clock | 2686mhz | 2686mhz | 2686mhz | 2485mhz | 2485mhz | 2485mhz | 2587mhz |
core voltage | 0.941v | 0.933v | 0.815v | 0.903v | 0.904v | 0.903v | 0.956v |
gddr6 usage | 4316mb | 4135mb | 5484mb | 6079mb | 8615mb | 9233mb | 8597mb |
ddr5 usage | 11455mb | 10723mb | 10719mb | 10714mb | 11694mb | 11606mb | 10889mb |
gpu mcd6 | 71.7c | 64.8c | 61.0c | 62.6c | 59c | 55.5c | 65.9c |
收集以上約二十個使用設定中得出的經驗, 要追求最高性能(16XXX TSE GS), 代價是~430W實際TBP, 溫度變化不明顯而性能未見受溫度影響。若要將預設狀態下的實際的380W TBP視為目標, OC15這一種將電壓目標由預設的1.15V降至1.05V就能做到(頻率目標設定為2000-3500//2700)。OC16和OC17在做出比DEFAULT時的更低的實際TBP的時候,仍然能夠跑贏DEFAULT時的TSE GS。最後是暴力降功耗, 要到UV3這種極致的做法才能做出-100W的TBP實際表現。
簡單來說, 想降低RX7900XTX的功耗並不容易, 除非願意放棄大量性能。
應用在PORT ROYAL上
∇ 以下抽選出在TSE中表現較好的使用設定, 並以PORT ROYAL作驗證。似乎靠得住。
port royal | stock (500-3025//2498// 100%//1.15v// fan46%) |
oc9 (2000-3500// 2700//115%//1.05v// fan75%// fast timing) |
oc15 (2000-3500//2700// 100%/1.1v// fan75%// vramfast) |
uv1 (500-3000//2498// 90%//1.05v// fan46%) |
uv2 (500-2500//2498// 90%//1.05v// fan46%) |
uv3 (500-2000//2498// 90%//1.00v// fan46%) |
ocpassed (2000-3100//2600// 115%//1.125v// fan75%) |
graphics score | 15937 | 17992 | 16879 | 16360 | 15620 | 13635 | 17190 |
gpu temp | 57c | 55c | 51c | 51c | 49c | 43c | 55c |
memory temp | 82c | 80c | 78c | 76c | 74c | 70c | 80c |
gpu hot spot | 63c | 87c | 78c | 74c | 67c | 58c | 88c |
gpu fan | 1262rpm | 2409rpm | 2107rpm | 1477rpm | 1478rpm | 1477rpm | 2411rpm |
total board power | 375w | 432w | 376w | 338w | 303w | 268w | 362w |
gpu power peak | 504w | 538w | 519w | 450w | 317w | 275w | 431w |
front end clock | 2748mhz | 3009mhz | 2774mhz | 2855mhz | 2529mhz | 2029mhz | 2888mhz |
front end effective | 2974mhz | 3252mhz | 2991mhz | 3070mhz | 2728mhz | 2200mhz | 3115mhz |
shader clock | 2753mhz | 3011mhz | 2770mhz | 2848mhz | 2532mhz | 2033mhz | 2886mhz |
shader effective | 2752mhz | 3010mhz | 2770mhz | 2847mhz | 2532mhz | 2033mhz | 2885mhz |
memory clock | 2485mhz | 2686mhz | 2686mhz | 2485mhz | 2485mhz | 2485mhz | 2587mhz |
core voltage | 0.961v | 0.974v | 0.916v | 0.904v | 0.903v | 0.904v | 0.988v |
gddr6 usage | 8182mb | 10491mb | 10535mb | 10580mb | 13626mb | 12950mb | 12326mb |
ddr5 usage | 6963mb | 6937mb | 6990mb | 6917mb | 8174mb | 7659mb | 7079mb |
gpu mcd6 | 66.8c | 65.1c | 61.5c | 60.3c | 57.9c | 52.1c | 64.9c |