Thread by @davidbepo

Navi 3 launch thread
first the most generic part, value perf and such

its no surprise to anyone that this card is another massive disappointment not just vs rumor mill but also vs AMDs absolutely memey data they provided on announcement event, based on that i was expecting close to a 60% gen on gen uplift but 50% wouldnt have disappointed me

and yet they not only failed to get to 50% they didnt even do 40... which is another marketing disaster
the card manages to beat 4080 by a bit while costing less, but this isnt a complete beating since 4080 consumes less and has DLSS(tho that doesnt justify 200€, AT ALL)

efficiency also massively missed targets with no way to sell this as a 50%
the cutdown card is an absolute meme with literal worse value than top end, it should have costed 850 at most and ideally 800

overall while i do think this cards are better than the alternative, they are very sad
and thats it for part 1, part 2 with all the technical stuff will be coming tomorrow because i delayed too much and now its late...

part 2 start, last things first
the mismatched stepping i aluded to here:

https://twitter.com/davidbepo/status/1603105379814371328

i have reached to TPU guys and along with other info i have collected the answer is clear
GPU stepping is indeed the A0 VBIOS reports...

and C8 is the board stepping which GPUz reads from the driver

this is absolutely shocking as Navi 3 has been delayed almost a year from the first internal target but somehow no time for a respin was doable (?)

about the clocks
from what i have gathered the clock reported by apps is the shader clock(im about 90% sure of this), frontend one can only be accesed by HWinfo, the division is a bit annoying but thankfully for me i get the clock i need for my math

and about the clocks itself, well ive never seen this level of variability and that variability is also in all aspects, workload, sample and inter SKU delta, for the first time in history i cant put an accurate AVG clock for my math for either of the cards

to illustrate this point here is the TPU and HWU clock numbers for 7900 XTX
techpowerup.com/review/amd-rad…

and ditto for XT
techpowerup.com/review/amd-rad…

as you can see the game clocks range massively but not only that, TPU shows the XT as clocking better while its the opposite on HWU
also nothing fits AMD stated clocks
and lets not talk about the furmark thing, i literally best that... on my 5500 XT, not joking ill show ya

Power draw behavior also seems to be completely broken, all volts are super low, never touching 1V, which is nice, Vf itself is actually pretty good, even better than expected, yet somehow regardless of V the card draws insane amounts of Amps(limit is over 400A btw)

needless to say thats unlike thats unlike any chip ive ever seen, i even dare say any chip ever
i have a theory of what is going on here, but it covers like 20% of what we are seeing at most, the rest i have no way to explain

my theory is that AMD went so fucking apeshit on density any heavy load is unsustainable on power delivery and heat density, this bodes well with the card nearing clock wall and exceeding 3 GHz on super light stuff while games get mid 2 GHz

and heavy stuff like furmark just dies, regressing to <OG Navi clocks
then again this does account for 20% or so of the range that we see the rest i just cant explain, the amperage behavior is also a huge question mark

side note, i have decided to use 2,45(XT) and 2,5(XTX) for AVG clocks in my DGPERF sheet because its close to AMD claims, fits with the avg of the results and works well with same IPC

anyways while clock/power is uber cursed it isnt far off what AMD claimed in announcement yet as i said in the beginning perf most definitely is which leads us to IPC and FP32 SPAM

im not gonna beat around the bushes, FP32 spam has utterly FAILED and IPC is only surpassed in sucking by the 4090
after some perf calibration i get 0,65 IPC for both cards this means that from a doubling in shaders AMD made each CU... 12% stronger at the same clocks... pathetic

even novideo got more with the amprere SPAM
in more correct shader or TFLOP terms each shader/TFLOP is now 1,78 times weaker
im sure area cost isnt much, but it has to be more than 12%... not to talk the power cost

admittedly a bit of this is due to slight CPU cap and higher shader scaling but still the result is memeworthy and i cant call the spam anything else than a failure with the results in hand

and then the technical back end is also an issue, VOPD has all kinds of pitfalls that make it not always work, even pure math programs cant get close to the theoretical peak perf
however in a pleasant surprise(literal only one) wave64 can do the full perf without VOPD

so GPU should be 12288 shaders more times than 6144 so to say

i was initially planning to stop the thread here (at least unless i got more stuff) but i believe a small talk about shipping as is vs delaying and fixing is in place

before i knew about this A0 stuff this was my stance

and i still believe that is the case HOWEVER depending on when AMD found A0 was a mess it could also be a case of blatant incompetence, if you see A0 sucks you dont just ship it, a respin done early barely means any delay

and now thats it for the thread, hope you found it helpful interesting or such

@threadreaderapp unroll