W7800 vs RX 7900 XTX
I ran the identical Q4_K_M blob and Ollama version on both cards: the consumer 7900 XTX is 29% faster at generation and 44% faster at prefill. The gap follows compute, not bandwidth — what the W7800 really offers is its 48 GB.
- Tested
- Setup
Radeon Pro W7800 48 GB — bare-metal workstationBare metal, Linux, ROCm (Ollama) · AMD Ryzen 5 7600 (6C/12T, AVX-512) · 64 GB DDR5-6400
Radeon RX 7900 XTX 24 GB — bare metalBare metal, Ollama (ROCm)
- Runtime
- Ollama 0.32.14 on both, flash attention on, q8_0 KV cache
- Models
qwen3.8:27bQ4_K_Mtested
Verdict
If the model fits in 24 GB, buy the 7900 XTX — it wins every speed measurement. Buy the W7800 for capacity: twice the VRAM is its only advantage, and on low-bit quants its 70 CUs struggle with unpacking the weights.
Works
- 7900 XTX: +44% prefill, +29% short and +32% long-context generation
- Prefill gain (+44%) matches CU × clock (+45%) almost exactly
- 7900 XTX results stable across two model families and two Ollama versions (33.6 → 33.2 tok/s)
Doesn’t
- num_ctx differed (262K vs 32K) — measured earlier to not affect speed, but not identical
- Single model and quant; Q8_0 does not fit the 24 GB card
- Prefill gain, 7900 XTX
- +43.9%
- Generation gain (short)
- +29.2%
- Bandwidth difference
- +11%
- CU × clock difference
- +45%
Generation speed
higher is better- W7800 48 GB
- RX 7900 XTX 24 GB
Data
| W7800 48 GB | RX 7900 XTX 24 GB | |
|---|---|---|
| Short prompt (21 tok) | 25.68 tok/s | 33.19 tok/s |
| Long prompt (31,073 tok) | 20.75 tok/s | 27.49 tok/s |
Prefill, 31,073-token prompt
higher is better- W7800 48 GB
- RX 7900 XTX 24 GB
Data
| W7800 48 GB | RX 7900 XTX 24 GB | |
|---|---|---|
| Prefill | 498.0 tok/s | 716.5 tok/s |
Where the advantage comes from
7900 XTX over W7800, percent
Data
| 7900 XTX advantage | |
|---|---|
| VRAM bandwidth (spec) | 11% |
| Measured generation | 29% |
| Measured prefill | 44% |
| CU × clock (spec) | 45% |
Setup — one variable
My July comparison mixed up the GPU and the quantization (Q8_0 on the W7800, Q4_K_M on the 7900 XTX), so there was no way to tell what the card itself contributed. This time I held everything constant except the card:
| Held constant | Value |
|---|---|
| Model | qwen3.8:27b-q4_K_M — the same blob, SHA-256 verified |
| Runtime | Ollama 0.32.14 on both (the 7900 XTX box was upgraded for this test) |
| Settings | FLASH_ATTENTION=1, KV_CACHE_TYPE=q8_0 |
| Prompts | Same short (21 tok) and long (31,073 tok) prompts |
| Offload | Both 100% GPU, verified |
The one asymmetry: num_ctx was 262,144 on the W7800 and 32,768 on the 7900 XTX (262K doesn’t fit in 24 GB). I’d
already measured that allocated context doesn’t change speed, so the comparison stands.
The cards
| W7800 | RX 7900 XTX | Difference | |
|---|---|---|---|
| Architecture | Navi 31 / gfx1100 | Navi 31 / gfx1100 | same |
| Compute units | 70 | 96 | +37% |
| VRAM | 48 GB | 24 GB | −50% |
| Bandwidth | 864 GB/s | 960 GB/s | +11% |
| Clock under load | avg 2,352 / max 2,495 MHz | max 2,639 MHz | ~+6% |
Results
| W7800 | 7900 XTX | 7900 XTX advantage | |
|---|---|---|---|
| Prefill (31,073 tok) | 498.0 tok/s (62.4 s) | 716.5 tok/s (43.4 s) | +43.9% |
| Generation — short | 25.68 tok/s | 33.19 tok/s | +29.2% |
| Generation — long | 20.75 tok/s | 27.49 tok/s | +32.5% |
| Long-context slowdown | −19.2% | −17.2% | similar |
| Bandwidth efficiency (short) | 52.7% | 61.3% |
The 7900 XTX short-prompt runs came in at 33.32 / 33.23 / 33.03 — very little variance.
The gap is compute, not bandwidth
| Measure | Difference | What it says |
|---|---|---|
| VRAM bandwidth | +11% | the pure bandwidth advantage |
| Measured generation | +29% | more than twice the bandwidth gap — the rest is compute |
| Measured prefill | +44% | prefill is pure compute |
| CU × clock | 1.37 × 1.06 = +45% | matches prefill almost exactly |
- Prefill is compute-bound: I measured +44% against +45% of theoretical compute.
- Generation is bandwidth-bound, but not purely: +29% lands between +11% (bandwidth) and +45% (compute). The difference is Q4_K_M unpacking — every token’s weights get unpacked on the CUs as they’re read.
This independently explains why Q4_K_M efficiency drops to 53% on the W7800: there are fewer CUs to absorb the unpacking cost. The same card holds 69% on Q8_0, which is cheap to unpack.
What it means if you’re buying
- The model fits in 24 GB → the 7900 XTX is clearly the better card.
- The model needs more than 24 GB → the W7800 is the only single-card option, and the speed loss is the price you pay for capacity.
A side note on moving models between machines
ollama pull on the second box crawled along at ~4.6 Mbps (I didn’t dig into why). Streaming the five blobs over the
LAN instead moved the 16 GB layer in 150 s (106 MB/s), and I checked each blob against its SHA-256 file name.