CDNA2 is a disaster? How exactly do you come to that conclusion? Just looking at the Top500 list, I see there a system with CDNA2 accelerators sitting comfortably in the top spot. It delivers >7x the performance of the best nVidia equipped system at only twice the power consumption. Alternatively compare the positions 3 and 4. The third place (again with MI250X) is slightly faster than the 4th place (nV) at 29% of the power consumption. It looks like it quite efficiently does what it was designed for.
- Status
He most probably meant that despite large two chiplets, TR count was surprisingly low. I was also surprised. Maybe this is side effects of full rate 64 FP?
Bondrewd
Veteran
Yea and that what makes xtor counts into meaningless metrics those days.
They're more about dark silicon flexing at this point than literally anything else.
MI250X was basically expressly designed for high efficiency FP64 performance. There it delivers (it clocks higher and basically doubles FP64 performance [in some cases with matrix ops it quadruples it] at the same power consumption and does so while spending only a few transistors more and with a relatively small process upgrade [7nm => 6nm]). That is the important metric and I don't see anything one would qualify there as an "disaster". Who cares how many transistors the chips have as long as the performance comes out at an acceptable cost (die size) in an acceptable power envelope.
And I fail to see the connection to the RDNA3 discussion (quite different targets). I think you give the "argument" of Troyan too much credit and are trying to give it some sense which wasn't there to start with. In the end, it may even work against Troyan's main argument, that N31 may be too small and spends too few transistors on something as Aldebaran actually shows that large gains can be had for relatively small investments.
Last edited:
Rootax
Veteran
I hope (and I'm pretty sure) it's not just brute force and you have some space to make your hardware smarter or faster than x or y...
Frenetic Pony
Veteran
Supposed to be the 7900 something, not the XTX which is supposed to have 3 8 pin. Which is supported AMDs official statements of moving to higher power usage.
So, 7900 or 7900 xt.
Bondrewd
Veteran
It is XTX.
The new ref is great, shame AMD would rather not sell many of those to not piss AIBs off (you know well enough why after the recent EVGA debacle).
Nope.
Frenetic Pony
Veteran
Have to disagree, the official AMD stance is
Straight from an interview, and this pic looks to be from the same leaker saying they've seen XTX cards with 3 8 pin connectors.
Bondrewd
Veteran
Oh yeah babeeeh the fun is starting.
I'm just saying what it is.
Yes the AIB ones.
As consumer i might give AMD their "fun" and expect mine in return if the MSRP is real instead of just a PR label and some price tag next to "out of stock" notification somewhere hard to find in AMD's website probably with just some link to closeby partners outside of some countries that end up nowhere, meanwhile the AIB's scalp the final consumer with +25% or so slice on top. I wonder what excuse we'll hear now for the prices without the mining scapegoat to fallback to, oh yeah, "inflation".
As im planing to purchase a new PC this december things don't look good from my pov, even more so since only high end get released 1st and mid range (with stupid high prices but not as ridiculous and "enthusiast") gets typically set back for a whole quarter.
Seeing as with their green competitor the situation is the same, and AMD loves glueing their MSRP noses up their competitors arse i guess intel is looking mighty interesting as a stop gap alternative.
As im planing to purchase a new PC this december things don't look good from my pov, even more so since only high end get released 1st and mid range (with stupid high prices but not as ridiculous and "enthusiast") gets typically set back for a whole quarter.
Seeing as with their green competitor the situation is the same, and AMD loves glueing their MSRP noses up their competitors arse i guess intel is looking mighty interesting as a stop gap alternative.
digitalwanderer
Legend
Have we heard prices yet? People seem awful judgemental and hateful a week away from getting our first real information about the new GPUs, it's sort of cracking me up.
Based on previous releases with only the Rx4/580 series being the most "recent" exception i don't find my judgement/expectations awfull and well within reason imo, and just because there's still no real inf. atm doesn't somehow prevent that scenario, it wouldn't be the 1st time happening would it?
Subtlesnake
Regular
Were the Radeon 5000 series selling for well above MSRP?
Bondrewd
Veteran
No.
Only Vega (cuz mining) and RDNA2 (yea mining) were that.
itsmydamnation
Veteran
I got a vega 56 on launch day for 630 AUD , a month later they where selling ~ 1500 AUD , was super lucky as my 280 went pop 2 days before launch, if it had of been a week later I would have been very sad panda.
Frenetic Pony
Veteran
Shrug, suppose it's possible if its binned super tightly and running 2x8 and the 75 watts from the slot. Seems too tight for variable draw though.
Yes, but the bigger problem is that this goes both ways, as the 'hateful prejudgement' is not a result of having fun with that, but from frustration dating back years already.
It's quite hard to find some middle ground and POV from which things look reasonable and acceptable.
It rather looks the whole train has gone off rails already, we await a crash, and after that maybe things build up slowly again more reasonable.
That's why i hope for a quick move towards game ready APUs for mainstream, to prevent the crash. I really have no better idea than this, and sacrificing enthusiast level would be a small price to pay.
But maybe a wonder happens and AMD decides to attack competition with low prices. However - everything becomes more expensive right now, so why should they.
Bondrewd
Veteran
It's not much higher than 6950XT and that does just fine given its transients.
Jawed
Legend
Looking at the cache hit rates in this image:
from:
github.com
it seems L1 hit rate is pretty miserable. This is a single frame rendered with a geometry buffer first created, then followed by some pretty advanced ray tracing techniques.
L1 in RDNA 2 is 128KB for a shader array. I suppose with it doubling in size in RDNA 3 and seemingly with more shader arrays, each L1 will service less WGPs. All of that would be expected to help with the L1 hit rate.
from:
kajiya/gi-overview.md at main · EmbarkStudios/kajiya
💡 Experimental real-time global illumination renderer 🦀 - kajiya/gi-overview.md at main · EmbarkStudios/kajiya
github.com
it seems L1 hit rate is pretty miserable. This is a single frame rendered with a geometry buffer first created, then followed by some pretty advanced ray tracing techniques.
L1 in RDNA 2 is 128KB for a shader array. I suppose with it doubling in size in RDNA 3 and seemingly with more shader arrays, each L1 will service less WGPs. All of that would be expected to help with the L1 hit rate.
- Status
