AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

[CarstenS]

More likely 32 GB of mem and only 250W TDP limits the petformance

Can you elaborate? Do you mean that the 32 GByte eat too much into the TDP budget or that 32 GByte is too little for ViewPerf 13?
FWIW, during ViewPerf 2020 (UHD), I frequently saw clocks beyond the advertised 2300 MHz. Didn't check yet with the older VP13.

 

[DegustatoR]

More likely 32 GB of mem and only 250W TDP limits the petformance

I dunno about that. Ampere is well known to show relatively higher performance gains over Turing in proviz applications. I'd expect this to be the main reason here.

 

[Kaotik]

Can you elaborate? Do you mean that the 32 GByte eat too much into the TDP budget or that 32 GByte is too little for ViewPerf 13?
FWIW, during ViewPerf 2020 (UHD), I frequently saw clocks beyond the advertised 2300 MHz. Didn't check yet with the older VP13.

I meant the power consumption part, with only 250W TDP 32 GB eats already notable chunk

 

[CarstenS]

I meant the power consumption part, with only 250W TDP 32 GB eats already notable chunk

Maybe, but it does not seem too far out there, considering RTX A5000 has only 230 watts of TDP with it's 24 GByte of GDDR6 (20 watts for 8 GByte?).

 

[Pressure]

Navi 23 (Dimgrey Cavefish) die from the W6600 press deck, was this posted already?

237 mm², 11.06BT, TSMC 7nm, 32 CUs, 32 MB IC, 128-bit, PCIe 4.0

 

[dskneo]

Navi 23 (Dimgrey Cavefish) die from the W6600 press deck, was this posted already?

Yes but I only posted the main specs. 100watts for 28 active CU at 2900mhz.

Its ~5700xt performance at 100watts which is nuts. Navi23 was dubbed "nvidia killer" internally at AMD if leaks are to be believed.

 

[Bondrewd]

100watts for 28 active CU at 2900mhz.

Unicorn bin and nowhere near all w/ls but yeah.

Its ~5700xt performance at 100watts which is nuts.

That one is actually cute yeah.

Navi23 was dubbed "nvidia killer" internally at AMD if leaks are to be believed.

Every part is designed to snap some greenie necks off.
Usually a cointoss on the things doing it sans some (N31) exceptions.

 

[Putas]

Less than 2/3 of Navi 22 transistors with so many CUs?

 

[Bondrewd]

Less than 2/3 of Navi 22 transistors with so many CUs?

SRAM went poof! and thus you're here.
IC isn't cheap which is why it's going off-die for throw-it-en-masse duties the very next year.

 

[tunafish]

And as a caveat, with only 32MB of IC and a 128-bit interface feeding all that compute, perf will probably fall off a cliff if you try using resolutions higher than FHD.

That is, of course, entirely acceptable for a part in this segment.

 

[no-X]

Navi 23 has more bandwidth per FLOP than GeForce RTX 3070 (I'm not even taking infinity cache into account). Anyway, gaming arithmetical performance isn't limited by bandwidth. Fill-rate is. And fill-rate is already quite limitedy because of (just) 32 ROPs.

 

[Bondrewd]

And fill-rate is already quite limitedy because of (just) 32 ROPs.

The thing has 64.

 

[no-X]

Interesting, all previous leaks said 32. In that case the Radeon RX 6600 XT shouldn't be slower than Radeon RX 5700 (assuming the ~2,7GHz boost is correct).

 

[Putas]

SRAM went poof! and thus you're here.
IC isn't cheap which is why it's going off-die for throw-it-en-masse duties the very next year.

If I subtract IC transistors I am still getting 2/3 of Navi 22 transistors.

 

[CarstenS]

If I subtract IC transistors I am still getting 2/3 of Navi 22 transistors.

What numbers do you use? Navi 22 has 17.2 bln x-tors (51,2M/mm²), Navi 23 11.06 (46,8M/mm²), that's 64% of the former. And it most likely has 32 Compute Units (if the artists representations and the lack of a harvested variant are anything to go by), so 80% of the 40 in Navi 21.

What's most likely different as well: 32 MB ∞$, 4 Display Engines (instead of 6), PCIe x8 (instead of 16), no ECC for Memory (smaller Memory Controllers), lower-specced memory Interface (could be utilized for smaller memory controllers as well, see Hawaii vs. Tahiti).

IMHO, that could work out.

 

[Putas]

I assume 6T cell for IC.
Wasn't aware of those other cuts. Maybe it can work out.

 

[CarstenS]

Your calculation got me interested and I started looking. Please note though, that apart from the smaller ∞$, the other cuts are most likely, but not confirmed by AMD.

So you calculated with 384M x-tors for the additional 64 Megs of ∞$? I think you'd need probably additional FlipFlops in some of the longer data paths, additional tags and whatnot (have not much clue about caches).

 

[Putas]

Let's say 3.3 billion for 64 MB.

 

[CarstenS]

Oh, yes, of course.

 

[DavidGraham]

RX 6600M is slower than RTX 3060 laptop, which is found to be 8% faster @1080p, 13% faster @1440p, 37% faster at Ray Tracing, over 100% faster in Blender Ray Tracing.

One reason for this results is the 128-bit memory bus coupled with 32MB of Infinity cache, severely limiting the performance of this RDNA2 GPU.

Another point:

both Radeon RX 6600M and RX 6700M require two separate motherboards designs, which will greatly increase required R&D spending for OEMs. In comparison, NVIDIA’s whole series ranging from GTX 1650 to RTX 3080 only requires two board designs

https://videocardz.com/newz/amd-radeon-rx-6600m-navi-23-mobile-gpu-gets-tested