What's the point of putting 3 8 pin connectors if one of pins is not required? If OC versions really need a 3rd pin then it either clocks like a champ or there is something really wrong with it's efficiency. I guess we'll know soon.
AMD RDNA4 Architecture Speculation
- Thread starter Frenetic Pony
- Start date
No, it doesn't mean either of those. We have countless examples from previous generations of cards with extra power connectors that offer no discernible benefits over default. Most likely it's just PR department telling engineers MOAR POWAH sells even when it does nothing.
According to IGN RX 9070 (non-XT) reaches 99 FPS in COD: Black Ops 6 at 4K Extreme settiings without FSR.
www.ign.com
I Benchmarked the AMD Radeon RX 9070 - IGN
Everything we know so far about the newest graphics cards from AMD.
DegustatoR
Legend
Not a very representative benchmark but puts it above 7970XTX.According to IGN RX 9070 (non-XT) reaches 99 FPS in COD: Black Ops 6 at 4K Extreme settiings without FSR.
I Benchmarked the AMD Radeon RX 9070 - IGN
Everything we know so far about the newest graphics cards from AMD.
Probably not a good idea to read much into it.
If it essentially gives you the performance of a 7900XT, which was going for ~$650 I think at it's lowest, why would you think they need to price it that low? It isn't a huge upgrade from a 6800XT yes but not everyone upgrades that often. The 7900XTX is still an option if someone really wants to go AMD.
Given that Nvidia is the market leader, naturally they can dictate pricing and AMD has almost no option but to follow. If AMD had announced a price and then had to announce a price cut due to Nvidia's aggressive pricing for the 5070, it would have been even worse.
Besides, they actually had great products in Kracken Point, Strix Halo and 9950X3D where they have leadership performance to talk about. I think they'd rather get more positive spin out of those than the slight negative spin RDNA4 might have gotten them.
I agree, the die size does seem too large, even if they beefed up the RT and ML hardware. The rumours were sub 300mm2 die. Let's wait for official news from AMD before we jump to any conclusions but it does seem larger than expected.
Not a very representative benchmark but puts it above 7970XTX.
Probably not a good idea to read much into it.
Yea too many variables to read into it. They might have used a totally different scene than TPU. And the CPU used is a 9950X3D vs what I think is a 13900K for TPU.
Some others have measured with different known references and got < 350mm2.https://videocardz.com/newz/amd-nav...rx-9070-pictured-may-exceed-nvidia-ad103-size
~390mm2 ? Quite chunky
Hard to get solid measurements on such a wide picture at an angle, you can easily be off +10% on each side.
I would guess around 320-330mm2.
They supposedly doubled up RT and went to 4SE/128ROPs.
N31, and to some degree N32, were shader heavy designs.
N48 seems to be much better balanced, at least on paper.
Edit- Found your original post(from April '24) with the silly season die sizes from twitter for N48 and N44.
Going off my post/response, N48 was rumored to be ~240mm2 and N44 was ~120mm2 (i think).
Even then I thought ~280-320mm2 for N48 and ~140-150mm2 for N44 made more sense.
Last edited:
DavidGraham
Veteran
FSR4 hands on from HU, at 4K Performance the improvements vs FSR3 are very obvious.
NOTE: it's quite obvious AMD was ready to talk about a lot more things than they did, but they canceled all of it at the last minute.
NOTE: it's quite obvious AMD was ready to talk about a lot more things than they did, but they canceled all of it at the last minute.
DegustatoR
Legend
Seems like a solid enough leak.
7900XTX level in RT, >7900XT w/o RT.
TSE is skewed in RDNA's favor so the card being >4080S probably won't be representative overall.
Still the result suggests that it should be faster than 7900XT but slower than 7900XTX.
Combined with TDP increase it may mean that they are trying to reach some specific card on the Nvidia side - remains to be seen which one though.
7900XTX level in RT, >7900XT w/o RT.
TSE is skewed in RDNA's favor so the card being >4080S probably won't be representative overall.
Still the result suggests that it should be faster than 7900XT but slower than 7900XTX.
Combined with TDP increase it may mean that they are trying to reach some specific card on the Nvidia side - remains to be seen which one though.
Some of the stuff posted there is a bit suspect, like how GPU-Z and/or the driver seem to think it's a 7800XT.
In particular the Furmark shot stood out to me because it showed 329w power draw, with a GPU voltage of... 0.765v?
Unless that bit of telemetry's wrong, that's a strange part of the V/F curve to be hanging out in while still pulling down 330W of power, especially considering the assumption is 330W is significantly raised from their original targets, and the board would be less power constrained than it would have been initially.
Edit: I guess I'm used to looking at the data for NVidia cards more often, which don't throttle quite as bad in FurMark.
Going back to TPU's 7800XT review, the data looks similar, so I guess it could be legit: https://www.techpowerup.com/review/amd-radeon-rx-7800-xt/39.html
Last edited:
According to KitGuru PowerColor has said that current drivers limit performance around 20 % compared to release/media drivers
We also have to keep in mind that the 256 bit memory interface will not scale much from 6nm to 4nm and there is likely more control logic given that RDNA4 supposedly also supports GDDR7. I would be surprised if N44 is that low though, it's almost exactly half of N48 and we know there is a lot of "uncore" which dosen't scale linearly with specs. If N48 is ~320mm2, N44 should be around ~170-180 mm2.
DavidGraham
Veteran
AMD rep says performance of the 9070XT will be better than the leaks prior to CES.
Raster performance will improve a little bit, while RT and ML will improve considerably, because the ecosystem has evolved now to make use of these features.
He also says AMD will focus mainly on value, there is not gonna be a focus on tech showcases, or technology demos or heavy visual features or any over investment in features that will not be taken advantage of.
Raster performance will improve a little bit, while RT and ML will improve considerably, because the ecosystem has evolved now to make use of these features.
He also says AMD will focus mainly on value, there is not gonna be a focus on tech showcases, or technology demos or heavy visual features or any over investment in features that will not be taken advantage of.
DegustatoR
Legend
AMD Radeon RX 9070 XT Custom Model Listed By Retailer For $530, Specs Disclosed With Navi 48 GPU Clocked at Almost 3 GHz
AMD Radeon RX 9070 XT GPU has been listed by a retailer based in the Philippines for a price of around $530 US & a GPU clock of almost 3 GHz.
wccftech.com
AMD Radeon RX 9070 XT Custom Model Listed By Retailer For $530, Specs Disclosed With Navi 48 GPU Clocked at Almost 3 GHz
AMD Radeon RX 9070 XT GPU has been listed by a retailer based in the Philippines for a price of around $530 US & a GPU clock of almost 3 GHz.
If that thing has performance in the 4070Ti region then it would be offering a lot of value at say $499. It will presumably be faster than the 5070 in raster and maybe comparable in RT. Plus now with AI based FSR4, and I wouldn't be surprised to see AMD come out with their own version of MFG either. Hopefully this generation we'll see a bit more competition in the mid range.
It would also make a very compelling alternative to anyone with an older GPU considering getting a PS5 Pro.
DegustatoR
Legend
Supposedly the results are for OC - not clear if manual or factory.
DegustatoR
Legend
Probably. Doubtful that they'll launch the thing in 2 days.
N4 is just very slightly improved N5. No big area savings from that.
RDNA4 CU is probably quite different than RDNA3 CU.
What is already publicly known is that the TMUs will be much beefier, capable of twice the texturing speed in many (but not all) situations. The beefier TMUs will both give perf increase and will also be bigger.
And then about the shaders itself:
RDNA3 doubled the number of multipliers per CU, but could dual-issue multiplications very rarely in real-world code, leading to very small performance increase from those dual-multipliers.
What I'm expecting is that they will get rid of many of the bottlenecks which prevented utilizing these dual-multipliers often. Might mean more register file ports or beefier frontend on the CUs.
Whatever they'll do, it will also mean the CUs might be considerably beefier.
Also, they might have finally added dedicated HW for the BVH tree traversal, not going it on software in shader cores anymore. If added, this will also add more area.
And the final thing: Cache. The L3 cache ("inifinity cache") can consume lots of die area if big. AFAIK we do not know the size of the cache yet, but I'd love to see the same 128 MiB 6800XT had, even though I'm pessimisting and thinking 64 MiB is more probable than 64 MiB.
Big cache is expensive, but can both improve performance by helping the bandwidth bottleneck, and also increase energy-efficiency, as DRAM accesses consume quite much energy.
IMHO 32 improved DCUs (64 CU's) with (often) twice the texturing power, beefier register files and instruction frontend, dedicated BVH tree traversal engine and 128 MiB of L3 cache would be quite nice, balanced and not very small chip.
128MiB is (was) rather overkill. 64MiB with better hit rates and lower latency will yield better perf/area ratio.