AMD Big Navi Speculation: The RTX 2080 Ti Destroyer?

I was waiting for AMD to share some details on Big Navi & RDNA 2 before doing the performance speculation.

On AMD Financial Analyst Day, I got exactly what I was looking for.

Yes, you heard that right. The Radeon Technologies Group is not holding back. RDNA 2 could be AMD’s Zen 2 moment for Graphics Cards.

Earlier on, AMD also claimed a 50% higher performance-per-watt for RDNA 1 and we are going to verify the same. We will check if RX 5700 XT’s performance-per-watt figure is 50% better than Vega 64 or not. That should give us a rough idea for RDNA 2.

With RDNA 1, it was possible for AMD to launch an RTX 2080 Ti competitor that consumes more power. RDNA 2 will allow them to go way beyond the 2080 Ti into new horizons.

And this time around, they are not planning to repeat the same mistakes as they did with Vega. AMD has officially announced that they are working on two separate GPU architectures, RDNA & CDNA.

RDNA will be optimized for Gaming & CDNA for Compute.

AMD also revealed the roadmap for their GPU architectures. RDNA 2 & CDNA are coming in 2020 while RDNA 3 & CDNA 2 will arrive in 2021.

RDNA 2 will be fabricated using one of the variants of TSMC’s 7nm Process (most likely N7+ instead of N7P or N7). Unlike Zen 4 based EPYC Genoa, they decided not to reveal the fabrication node of RDNA 3. But as Zen 4 is confirmed to be on TSMC’s 5nm HPC node, I am estimating the same for RDNA 3.

These are the three core features that RDNA 2 brings to the table:

Let’s take a deep dive into Radeon’s 2nd Gen DNA and estimate the performance numbers for AMD’s upcoming flagship card.

    RDNA 2

In 2020, AMD will be refreshing its RX 5000 lineup with the RX 6000 based on RDNA 2. The new chips will are codenamed Navi 2x with Navi 21 being the anticipated ‘Big Navi’.

c

AMD is claiming that next-gen Navi will be extremely powerful while being power-efficient. The second statement ‘Top-of-Stack GPUs with Uncompromising 4K Gaming’ gives a clear hint that there will be a flagship chip this time around.

RDNA 2 based GPUs will also support Hardware Accelerated Ray Tracing & Variable Rate Shading.

Fabrication Process

AMD has already revealed long back that RDNA 2 based GPUs will be fabricated using TSMC’s 7nm+ (N7+ Process).

The Navi 10 was fabricated using TSMC’s N7P process while Radeon VII used the N7 Process. Remember, N7, N7P, and N7+ are different from each other and N7+ is the one based on EUV.

Compared to the N7 Process, N7+ uses EUV for some of the critical layers and it is 15-20% denser than N7. This new node helps with either a 15% power reduction or 10% higher performance. You can get more information on different fabrication nodes here.

According to TSMC, the current defect rate for N7+ is on par with N7, i.e. 0.09. So, this shouldn’t be much of an issue for the yields of the bigger chips.

Image Credits: TSMC

Power Efficiency

AMD is claiming a 50% improvement in Performance-per-watt.

This would suggest that at the same power consumption as the RX 5700 XT, an RDNA 2 GPU should perform 50% better.

AMD also compared the performance-per-watt of their CPU architectures and as per this graph, RDNA 1’s performance/watt was somewhere in between the Zen & Zen 2.

But RDNA 2’s performance/watt has even surpassed Zen 2. The image also states what made this possible:

Ray Tracing

AMD also showcased Ray Tracing on their upcoming GPUs. They will use a common Ray Tracing architecture across the consoles and PC and that would be extremely helpful for the game developers as well.

While it is hard to speculate how well Ray Tracing will perform on RDNA 2, it could end up being somewhere in between Turing & Ampere.

    Performance Speculation

Let’s first look at the specification of AMD’s flagship consumer chips in the last few generations. The specification of Big Navi is based on rumors and some calculations.

If the RX 5700 XT was on N7+ (15 to 20% density improvement), its die size would be around 213 to 220 mm². Doubling the compute units would result in a die size of 426 to 440 mm². Most of the remaining die (65 to 79 mm²) would be reserved for Ray Tracing hardware.

This will bring the estimated transistor count of Navi 21 around 23.7 to 24.8 billion.

That means Big Navi will have 27 to 33% more transistors than the RTX 2080 Ti while being 33% smaller in size.

GPU Name	Vega 64 (Vega 10)	Radeon VII (Vega 20)	RX 5700 XT (Navi 10)	Rumoured Big Navi (Navi 21)
Architecture	5th Gen GCN	5th Gen GCN	RDNA	RDNA 2
Fabrication Process	GlobalFoundries 14nm LPP	TSMC 7nm (N7)	TSMC 7nm (N7P)	TSMC 7nm EUV (N7+)
Die Size	486 mm²	331 mm²	251 mm²	505 mm²
Number of Transistors	12.5 billion	13.2 billion	10.3 billion	23.7 – 24.8 billion
Compute Units/Shaders	64 / 4096	60 / 3840	40 / 2560	80 / 5120
ROPs	64	64	64	96
TMUs	256	240	160	320
Base/Boost Clock	1247 MHz / 1546 MHz	1400 MHz / 1750 MHz	1605 MHz / 1905 MHz	1700 MHz / 1850 MHz
Memory Type	HBM2	HBM2	GDDR6	GDDR6 / HBM2e
Memory Size	8 GB	16 GB	8 GB	16 GB / 24 GB
Memory Bandwidth	483.8 Gbps	1028 Gbps	448 Gbps	–
Memory Bus Width	2048-bit	4096-bit	256-bit	512-bit (GDDR6) / 4096-bit (HBM2e)
TDP	295 W	300 W	225 W	–
Avg. Power Consumption (Gaming)	288 W	291 W	218 W	–
Relative Performance (Gaming)	100%	123.3%	118.3%	–

The relative gaming performance in the above table is based on our gaming benchmarks.

We are most likely to see GDDR6 on Big Navi but with higher bandwidth and bus-width. A 512-bit bus-width which is double to that of RX 5700 XT is expected.

Even if a 384-bit Memory bus suffices for the 80 Compute Units, it won’t be enough with Ray Tracing on. For better Ray Tracing Performance, Big Navi will require more bandwidth and a lot more cache. We should expect 512-bit GDDR6 on Big Navi and if there’s still a bottleneck, AMD will use HBM2e.

Efficiency & Performance

Let’s revisit how AMD improved performance on 1st Gen RDNA.

Around 25% of the total perf/watt improvement was due to the 7nm Process. The rest 75% came from Architectural improvements.

With 7nm+ for RDNA 2, we are expecting 15% lower power at the same clock speeds. The rest of the improvements will come from architectural improvements.

Compared to 14nm Vega 64, AMD claimed a 1.5X Performance-Per-Watt for 1st Gen Navi.

Let’s verify AMD’s official numbers.

In the specs table from the previous section, we have listed the power consumption and relative performance of Vega 64, Radeon VII, and RX 5700 XT.

In gaming, Radeon VII is 23% faster than Vega 64 while consuming a similar amount of power. That’s a performance-per-watt jump of nearly 23%.

RX 5700 XT is 3.25% slower than Radeon VII in gaming while consuming 25% less power.

Let’s calculate performance-per-watt for each card and also multiply the numbers with 100 to improve readability.

Vega 64: 100/288*100 = 34.72

Radeon VII: 123.3/291*100 = 42.37  (22% higher performance-per-watt from Vega 64)

RX 5700 XT: 118.3/218*100 = 54.27 (28% higher performance-per-watt from Radeon VII & 56.3% higher than Vega 64)

Big Navi: 81.4 (50% improvement vs RX 5700 XT as AMD is claiming)

Looks like AMD’s performance-per-watt numbers for RX 5700 XT vs Vega 64 are even better than what they are claiming (56.3% vs 50%).

Here’s what AMD expects from RDNA 2.

Yes, there’s another 50% bump in performance/watt. And this time around the efficiency gains from moving from N7P to N7+ aren’t even close to what AMD gained by moving from GloFo’s 14nm to TSMC’s N7P.

If what AMD is claiming turns out to be true, we are looking at an entirely different architecture with huge IPC gains. These numbers are hard to believe but then again this is the same company that developed Zen 2.

With these numbers, AMD wouldn’t even require 80 Compute Units to defeat RTX 2080 Ti.

Big Navi = 2x RX 5700 XT with 50% higher performance-per-watt

This should theoretically make Big Navi 100% faster than RX 5700 XT if there are no bottlenecks from other components of the GPU.

But performance doesn’t always scale linearly as Buildzoid has explained in this video. However, he didn’t factor in efficiency improvements that would come with RDNA 2.

If Big Navi was on RDNA 1 with 80 CU, it would be around 35-40% faster than RTX 2080 Ti while consuming around 420 Watts and that’s not feasible and would require a significant reduction in clock speeds. In short, an RDNA 1 based 80 CU card is out of the question.

Let’s bring in RDNA 2 and reduce the power consumption of Big Navi to 300 Watts. Remember, earlier we arrived at a perf/watt figure of 81.4 for Big Navi while assuming 50% perf/watt improvement vs 5700 XT (AMD’s claims).

Thus, Big Navi’s relative performance vs Vega 64 = 81.4*300/100 = 244.2%

That means Big Navi will be around 144.2% faster than Vega 64.

⇒ Big Navi could be around 106% faster than RX 5700 XT.

The 2x Performance Rumors vs RX 5700 XT could actually end up being true as RDNA 2’s efficiency solves the biggest issue.

Yes, that figure sounds unrealistic. But unless AMD’s claims are false or my calculations were wrong, this should be close to the actual performance theoretically.

There could be other bottlenecks as well that would hamper the performance but even in the worst-case scenario, Big Navi should be at least 35-40% faster than RTX 2080 Ti for gaming.

Also, don’t forget that the part about 80 CU and 300W Power Consumption is based on rumors & assumptions. Big Navi could end up being a 72 CU part at 250W and then the performance numbers will be different. Also, the difference in clock speeds would influence the final performance as well.

AMD could also launch a lower CU card first that defeats RTX 2080 Ti and then later launch a bigger GPU to compete against Ampere.

    Big Navi vs RTX 3080 Ti

The question is not whether Big Navi will be able to beat RTX 3080 Ti, but whether RTX 3080 Ti will be able to beat Big Navi.

In my speculation article for Nvidia Ampere, I arrived at a figure of 35 to 45% improvement on GA102 (RTX 3080 Ti) when compared to TU102 (RTX 2080 Ti). However, that estimation was without any calculation as Nvidia is yet to reveal the details on the architecture. It was all based on rumors & leaks that would likely turn out to be false.

I’ll be updating the article after GTC 2020 where Nvidia is likely to share the first details on Ampere.

As per the current estimations, Big Navi could be on par or slightly faster than RTX 3080 Ti.

But after turning Ray Tracing On, I expect Nvidia’s Ampere Flagship to deliver higher FPS than Big Navi. I don’t think Big Navi would bother Jensen Huang as long as Nvidia has the fastest Ray Tracing GPU.