AMD accidentally leaks Navi 31/RDNA 3 details through their latest Linux Patches

RDNA 3 will use GPU chiplets, and now we know they are using them

AMD accidentally leaks Navi 31/RDNA 3 details through their latest Linux Patches

AMD's RDNA 3 flagship should feature six memory MCDs and a 384-bit memory bus

AMD's due to release graphics cards based their next-generation RDNA 3 architecture later this year, but so far AMD has remained mostly silent when discussing their new architecture's benefits. Unlike their upcoming Zen 4 CPU architecture, AMD is holding information about their RDNA 3 graphics cards close to their chests, but now, thanks to AMD's AMDGPU Linux drivers, we have some new information about AMD's RDNA 3 flagship. 

AMD's AMDGPU driver, as spotted by @kepler_L2 om Twitter, seemingly confirms that AMD's RDNA 3 flagship (Navi 31) graphics card will feature up to six memory controller dies (MCDs), which should connect to a single Graphics Compute Die (GCD). This lines up with older rumours that Navi 31 will use seven chiplets. 

With each MCD reportedly giving AMD's RDNA 3 GPUs a 64-bit memory controller and 32MB of Infinity Cache, AMD's Navi 31 flagship should feature a 384-bit memory bus and 192MB of Infinity Cache. When compared to AMD's RDNA 2 flagship, the Radeon RX 6950 XT, Navi 31 should feature a 50% increase in memory bus size and a 50% increate in Infinity Cache size. 

AMD accidentally leaks Navi 31/RDNA 3 details through their latest Linux Patches

(AMD's mention of six memory MCDs on an RDNA 3 GPU)

V-Cache on a GPU?

According to @Kepler_L2, AMD has plans to add V-Cache to some RDNA 3 GPU models, adding an additional 32MB of Infinity Cache to each GPU MCD. This would increase AMD's MCD Infinity Cache size from 32MB to 64MB. For AMD's Navi 31 flagship, which appears to feature six MCDs, V-Cache would give AMD's RDNA 3 flagship 384MB of Infinity Cache. 

Adding V-Cache to RDNA 3 has two benefits. First, it increased the amount of data that AMD's Infinity Cache can hold. Infinity Cache offers lower latencies and more bandwidth than a GDDR6 memory bus, enabling higher performance levels. Making this cache larger allows AMD to rely on Infinity Cache more and rely less on its GDDR6 memory bus, freeing up GDDR6 bandwidth for other tasks while also giving giving AMD's RDNA 3 GPU core faster memory access.

The second benefit of additional Infinity Cache is that it will likely give AMD GPUs more Infinity Cache bandwidth, which could have its own performance benefits. With a larger cache size and more memory bandwidth, AMD can increase the hit rate of their cache and deliver larger performance benefits to users. This could help AMD deliver higher performance levels at ultra high resolutions like 4K and 8K.

AMD accidentally leaks Navi 31/RDNA 3 details through their latest Linux Patches

What are the benefits of chiplets?

Larger chips are more expensive to create, as the larger your chip is, the fewer chips you can fit on a single silicon wafer. Larger dies often result in more wasted space on silicon wafers, and manufacturing defects can impact a larger percentage of the chips produced from each wafer. These two factors make large processors and GPU dies expensive to produce, and chiplets can avoid this problem. 

By connecting several smaller chips (chiplets) together, companies can create products that act similarly to a larger chip design while being significantly more cost effective. With chiplets, companies like AMD can create larger processors by bringing several chips together, avoiding the pitfalls of creating larger and larger monolithic designs. These pricing benefits can be used to lower the consumer costs of products, and to increase AMD's margins, which is great news for both consumers and investors. 

The problem with chiplets is that chiplets require fast interconnects to function, and it can be hard to get multiple separate processors to run together as unified multi-chip product. Companies that are investing in advanced chip packaging technologies and chiplets need to make sure that the advantages of chiplets outweigh these technical challenges, because chiplet-based product designs are only impressive if they deliver strong performance levels without any apparent downsides.

With AMD investing in "advanced chiplet packaging" with RDNA 3, it is clear that AMD is taking a multi-chip approach with their next-generation graphics cards. This is a first for the "gaming" segment of the GPU market, a factor that may place AMD ahead of Nvidia.

You can join the discussion on AMD's latest Navi 31 leaks on the OC3D Forums.

«Prev 1 Next»

Most Recent Comments

13-07-2022, 15:59:33

Pretty cool, I'll end up getting whatever they call their top end GPU but not for a while, Likely around Feb or March in the new year.Quote

13-07-2022, 16:57:09

I'm a bit of an AMD fanboy, but I think the next gen won't interest me too much. My 6900XT seems MORE than enough for my meagre needs, and it seems to run fine at 2800mhz, so I can't see me needing any more power than that. Still cool to see the advances made though, I still have a Radeon VII that I love dearly in my home theatre PC, and is STILL a killer 1440p gaming card when OCed and water blocked. I just won't replace that thing until I'm forced to, love it.Quote

13-07-2022, 21:42:42

would love to see the days of 512 bus again with todays memory and GPU's frame times and 1% lows would be so much better, but prices are bad enough as it is so better not give them ideasQuote

Register for the OC3D Newsletter

Subscribing to the OC3D newsletter will keep you up-to-date on the latest technology reviews, competitions and goings-on at Overclock3D. We won't share your email address with ANYONE, and we will only email you with updates on site news, reviews, and competitions and you can unsubscribe easily at any time.

Simply enter your name and email address into the box below and be sure to click on the links in the confirmation emails that will arrive in your e-mail shortly after to complete the registration.

If you run into any problems, just drop us a message on the forums.