AMD Radeon PRO W7900
The Radeon PRO W7900 is AMD's 2022-generation RDNA 3 workstation card with 48 GB of ECC GDDR6 on a 384-bit bus, 96 compute units, and a 295 W triple-slot actively cooled board on PCIe 4.0 x16. Its draw for AI work is capacity: 48 GB on one card at about $4,000, with no HBM or fabric. It lacks the FP8 matrix support of the newer R9700, so it is an FP16 and INT8 card. AMD's own page carries a footnote that the W6000 and W7000 series are not designed or recommended for datacenter use, and GPU Smith treats it as a workstation and developer part.
- 48 GB on a single card avoids multi-GPU splits for many 30B-class models at 8-bit and lets you hold a quantised 70B model with a small context.
- No FP8 matrix support in AMD's published specs. Throughput on FP8 checkpoints will not benefit the way it does on the R9700; plan for FP16, INT8 or INT4.
- Memory bandwidth (864 GB/s) beats the R9700's 640 GB/s, which matters for decode-bound inference; the R9700 wins on matrix compute.
- AMD states this series is not designed for datacenter use. For rack deployment, use an Instinct part or a vendor-qualified server GPU.
How much memory does the W7900 have?
48 GB of GDDR6 with ECC on a 384-bit bus, 864 GB/s peak bandwidth.
How much does the Radeon PRO W7900 cost?
About $3,999 in the US as of August 2026, per a price tracker (reported).
W7900 vs R9700 for LLM inference?
The W7900 has 48 GB against 32 GB and higher bandwidth. The R9700 is the newer RDNA 4 card with FP8 support and, in Phoronix testing, about 1.47x the W7900 performance, at about a third of the price.
Is the W7900 supported in datacenters?
AMD's page says W6000 and W7000 series cards are not designed or recommended for datacenter use.