Yesterday on AMD’s Reddit subboard, user FortCrypto reported finding a post on the Eternal Gaming-Age Forum site (EGAF) where a new user named “Bright Iron” provided information about AMD’s latest RDNA 3 architecture Navi 31 GPU theoretical bandwidth speed. He speculates that the AMD Radeon RX 7900 XT with the RDNA 3 ‘Navi 31’ GPU could have an effective bandwidth of over 6 TB/s.
According to the data, the AMD Radeon RX 7900 XT graphics card can deliver more than 7 TB per second of bandwidth. While the discussion is classified as speculative or theoretical, in this case, it is no less than a fascinating insight into AMD’s new graphics card.
In the post, the Bright Iron user does his calculations using the information found in AMD’s RDNA 3 Navi 31 flagship GPU, which offers 6 MCDs and up to 384 MB of Infinity Cache in a 3D V-Cache package.
His theory is based on information already available, such as the AMD Radeon RX 6950 XT, which is the best performer in the company’s current GPU lineup, offering 16 GB of GDDR6 memory with the help of a graphics card clocked at 18 Gbps 256-bit bus. This combination of technologies results in a bandwidth speed of 576 GB/s.
Bright Iron goes on to discuss bandwidth dwarfed by NVIDIA’s GeForce RTX 3090 GPU, which offers 936 GB/s, compared to 1,008 GB/s for the company’s RTX 3090 Ti.
Bright Iron went on to explain that AMD’s bandwidth performance is quite impressive compared to NVIDIA because of the company’s use of Infinity Cache technology. Infinity Cache is a memory that is separate from the main memory located on the AMD GPU. It is typically small in size compared to the main memory on the graphics card, but due to its proximity to the chip, Infinity Cache can be pulled at a higher rate.
MD Infinity Cache provides up to 128 MB of memory capacity and 2 TB/s of memory bandwidth. The AMD Radeon RX 6850 XT offers 512 GB/s of bandwidth as standard until you add 128 MB of Infinity Cache, increasing the bandwidth by another 1,281 GB/s and providing users with a total of 1,793 GB/s of bandwidth, far more than NVIDIA can currently produce.
The newer RDNA 3 architecture will allow Infinity Cache to reach 384 MB, which is three times higher than the previous RDNA 2 architecture. The chip can generate 192 MB from the standard Cache while utilizing another 192 MB of 3D V-Cache packaging to achieve 384 MB for the entire Infinity Cache.
Bright Iron has calculated that this new amount of Infinity Cache will yield three times the bandwidth (excluding the main memory bandwidth) assuming it runs at the same speed as RDNA 2. RDNA 2’s 128MB Infinity Cache has a maximum bandwidth of 2 TB/s (terabytes per second), so we see that RDNA 3’s 384MB is 6 TB/s.
However, it could potentially run much faster, so 6 TB/s is probably the minimum requirement; again, this does not include the additional bandwidth of main memory, which is claimed to be 24GB of GDDR6 and a 384-bit bus (don’t confuse this number with 384MB of Infinity Cache).
Memory @ 16Gbps = 16,000,000,000 bps (bits per second)
16,000,000,000 bps x 384 = 6,144,000,000,000 bps
6,144,000,000,000 bps ÷ 8 bits = 768,000,000,000 Bps (Bytes per second)
768,000,000,000 Bps ÷ 1,000,000,000,000 bytes = 0.768 TB/s
0.768 TB/s + 6 TB/s = 6.768 TB/s
