The AMD RDNA 3 Navi 31 flagship GPU is said to have up to 384 MB of Infinity Cache in a 3D V-Cache package and six MCDs.


@ Kepler_L2 reveals entirely new information about AMD’s RDNA 3 flagship GPU, the Navi 31, which says the chip can carry up to 384MB of Infinity Cache in standard and 3D packages.

Suspicion of shaking Infinity Cache up to 384MB with AMD RDNA 3 “Navi 31” MCM GPU, 3D V-Cache flavor emerges

This information is specific to AMD RDNA 3’s flagship Navi 31 chip.

So far, there is no information about the top chip known only as the Navi 3x, so the Navi 31 will remain in its flagship position when launched this year and will be replaced by faster variations next year. maybe.

AMD Navi 31 GPUs have been identified as adopting an advanced chiplet packaging design with next-generation Infinity Cache.

According to @ Kepler_L2, who dug up various patches and logs in AMD’s driver, the Navi 31 GPU seems to have up to 6 MCDs (multi-cache dies).

Each MCD is equipped with a standard cache of up to 32MB, and a model with 3D V-Cache is also available.

In this case, the standard cache will be 192MB, but some versions will have 3D V-Cache, which will have up to 384MB of Infinity Cache, insiders said.

The Infinity Cache is 32MB in size, half the size of Zen 3’s 3D V-Cache-equipped product (64MB), but with six MCDs, the 3D package can theoretically provide an additional 384MB of Infinity Cache.

In other words, the standard 192MB and 3D V-Cache 192MB can provide a total of 384MB of Infinity Cache.

* Click the image to enlarge it in another window / tab.

That’s three times the AMD RDNA 2 flagship Navi 21, with a total Infinity Cache of only 128MB.

The Infinity Cache architecture allows AMD to send more bandwidth to the GPU, eliminate memory bottlenecks, and provide fast cache near the GPU instead of going through VRAM.

AMD Radeon RX 7900 XT graphics card specifications

AMD Navi 31 GPU, the flagship chip of RDNA 3, will drive next-generation enthusiast cards such as the Radeon RX 7900 XT graphics card.

AMD has heard that next-generation RDNA 3 GPUs will stop CU (Compute Units) and adopt WGP (Work Group Processors).

Each WGP is equipped with dual CUs (Compute Units), but RDNA 2 doubles the SIMD32 cluster, which was only two in each CU.

According to the latest information, AMD Navi 31 GPUs with RDNA 3 architecture will offer a single GCD with 48 WGP, 12 SA and 6 SE.

This will provide a total of 12,288 stream processors, less than the previous number.

This will also reduce overall computing performance unless AMD is crazy with flagship parts at clock frequencies above 3.0GHz.

The Navi 31 GPU seems to provide a chip with a 384-bit bus interface, equipped with 6 MCDs with 32MB of Infinity Cache per die, and a 64-bit (32-bit x 2) memory controller. be.

In terms of clocks, AMD Navi 31 GPUs are said to offer clock speeds that reach or exceed 3GHz.

It is said that NVIDIA’s flagship GPU also achieves a clock speed close to 2.8 GHz, but AMD has had a clear advantage in clock speed over NVIDIA in the past generation, so it will continue to do so. Is expected to continue.

The clock of 3GHz means that the latest flagship can be expected to have FP32 performance of 75TFLOPs or more, which is 2.3 times that of the current RDNA 2 flagship RX 6950 XT.

AMD seems to have confirmed that its lineup of next-generation graphics cards will feature higher power consumption, but it will still be a more efficient option than what NVIDIA offers.

Since the AMD Radeon RX 6950 XT already has a TBP of 335W, the final TBP of the Radeon RX 7900 XT is expected to be close to 400-450W for more than double the performance improvement.

This card will be available later this year, as AMD itself confirmed.

AMD RDNA GPU (Intergenerational Comparison) Bulletin:

ソース:wccftech – AMD RDNA 3 Navi 31 Flagship GPU Allegedly Rocks Six MCDs With Up To 384 MB Infinity Cache With 3D V-Cache Packaging


This is a follow-up to Navi31 and RX7900XT, which are said to be 1GCD.

Infinity cache seems to be equipped with 384MB, which seems to be a model with 3D V-Cache.

Since the memory bus will be 384 bits, the memory capacity is expected to be 24 GB, which is the same as expected for RTX4090.

It is said that the FP32 calculation performance reaches 75TFLOPS at 3GHz, which is 2.3 times the performance of the RX6950XT.

The original table was still old, so I modified it here.

It is said that it will be a 6MCD + 1GCD configuration, and it seems that it will be a fairly high cost product even from the number of memory dies.

Even this does not have the same computing performance as the RTX4090.

Also, is Ada Lovelace’s game performance per TFLOPS higher than Ampere?

If it gets higher, the RTX4000 series will have considerably higher performance, but if it is about the same as Amepre, I wonder if the RX7900XT will be overwhelmingly superior in game performance even at 75TFLOPS. increase.

Unfortunately, TBP is expected to reach a maximum of about 500W, and I think there is a possibility that it will be equipped with a PCIe 5.0 standard compatible connector.

In addition, although it is not mentioned in the text, FSR 3.0, which is said to be released overlapping with the Navi 3X generation, is compatible with hardware AI accelerators, so it is finally Tensor for Radeon. It means that an arithmetic unit equivalent to the core will be installed.

AMD GPU Radeon series

Radeon RX 6000 series


¥167,698 (2022/07/13 12:51:25 As of Amazon research-detailed)


¥163,000 (2022/07/13 12:51:26 As of Amazon research-detailed)


¥121,000 (2022/07/13 12:53:08 As of Amazon research-detailed)


¥96,800 (2022/07/13 12:34:38 ​​As of Amazon research-detailed)


¥54,100 (2022/07/13 09:00:19 As of Amazon research-detailed)


¥44,000 (2022/07/13 08:42:40 As of Amazon research-detailed)


¥34,000 (2022/07/14 04:31:13 As of Amazon research-detailed)

* SAPPHIRE is a manufacturer specializing in AMD Radeon and is a reference manufacturer of Radeon.


Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Recent News

Editor's Pick