Pemalite said:
Bofferbrauer2 said:
I wonder why you measure with dual channel DDR5-4000 as if that were the limit. Did mix up the speeds with those from DDR4X? After all, DDR5 is supposed to start at 4800, and go all the way to 6400 - for now (higher speeds are planned for later by JEDEC). By that point, you have 51.2GB/s per channel, or 102.4GB/s on a standard dual channel board, quite a bit more than just 64GB/s. Unless you meant on a single channel, which will certainly be possible shortly before DDR6 comes along... Of course, that would still just be half of the bandwidth of the RX 470. However, the RX 470 gets beaten in performance by the RX 5300XT despite the latter only having 112GB/s. So either the bandwidth of the RX 470 was oversized, or RDNA needs much less of it than GCN4 to push pixels around. | |
I am basing that on initial launch support.
DDR4 3200Mhz is pretty ubiquitous and rigs with DDR4 3600Mhz is pretty common as it is the optimal DRAM for AMD's infinity fabric.
Not sure where you got DDR5-4000 from.
DDR5 is certainly meant to start at 4800mhz and will be what APU's start at... And DDR5 4800mhz is 64GB/s of bandwidth.
DDR5 6400mhz will happen later once manufacturers get an idea of yields and scaling.
I mean... We didn't get DDR2 with 1333mhz memory modules, that happened years after the DRAMs release, we didn't get DDR3 1866mhz memory on release, that came years later... And consequently, we didn't get DDR4 3600mhz memory straight away either.
So whilst yes faster speeds are planned for later, we don't know if systems will be compatible or how long it will take for that memory to come out, we might have already gone through a few upgrade cycles.
**********
You are right that RDNA does need less bandwidth than GCN4, primitive shaders, draw stream rasterization, improved culling, improved delta colour compression and a far more efficient cache hierarchy and a re-balancing of compute resources meant it was a more efficient beast for rasterization.
Which is why I would have liked to have seen AMD abandon Vega in it's integrated graphics years ago... Because even with DDR4 2400mhz the uplifts would be pretty damn awesome.
So yes, whilst RDNA offers better *gaming* performance than GCN, in compute workloads GCN will still win, which is why AMD decided to branch off it's architectures as CDNA and RDNA.
I am unable to find any real deep-diving benchmarks with the 5300XT likely because it's an OEM part verses the RX 470.. But comparing the 5500XT against the RX 590 the RX 590 still comes out ahead despite there only being a 225GB vs 256GB/s bandwidth difference, which probably showcases architectural differences rather than bandwidth. |
Ah, I see your problem. You're on a wrong calculation, that's why you get 64GB/s. DDR5-4800 in dual channel doesn't result in 64GB/s, but in 76.8GB/s. To get 64GB/s, you would need DDR4-4000, hence why I was asking why you were calculating with such low speeds. You could have noted this yourself, as you said in your previous post:
"In saying that... Dual Channel DDR4 3600mhz can do 57.6GB/s... Sadly laptop APU's aren't pushing that though, they stop at 51.2GB/s."
The jump from DDR4-3200 to 3600 brings 6.4GB/s. you need 6.4GB/s more to get to 64GB/s. Which results in 4000, not 4800.
DDR5-6400 is supposed to reach 51.2GB/s in single-channel and thus 102.4Gb/s in dual channel. And they are already testing DDR5-8400, which would result in 134.4GB/s; more than most entry-level GPUs have (RX 560 and 5300XT have 112GB/s; the GTX 1650 originally had 128GB/s).
