4 ms·
Coming I hope, though I wouldn't have huge expectations. He already did the general compute benchmark of the two 9070 cards here[1], and between poorly optimiz
by magicalhippo 2y ago
Coming I hope, though I wouldn't have huge expectations.
He already did the general compute benchmark of the two 9070 cards here[1], and between poorly optimized drivers and GDDR6's lower memory bandwidth, I wouldn't expect any great scores.
In terms of memory bandwidth they're between a 4070 and a 4070 Ti SUPER, and given that LLMs are very memory-bandwidth constrained as I mentioned in another comment, at best you'd expect the LLM score to end up between the 4070 and the 4070 Ti SUPER.
[1]: https://www.phoronix.com/review/amd-radeon-rx9070-linux-compute https://www.phoronix.com/review/amd-radeon-rx9070-linux-comp...
- Joel_Mckay 2y agoThe 9070 cards are hard to find, and inflated 30% above MSRP right now. Apparently the supply of fuse-crippled 9070 cards was just used to up-sell 9070 XT (easier to find.) Have a look at the GamersNexus YT rant about it... They make a fair argument in that price range an older model used RTX nvidia card may be a better value. Depends on your use-case, as rtx 5090 nvidia AI frame interpolation is dog crap hype for CGI or CUDA accelerated ML libraries. Personally, I would go with the rtx 4090 or even an rtx 3090 with 24G vram for ML and CGI workstation, as CUDA+Optix has better software support. For just gaming, the 9070 XT is a better deal when the MSRP is within range. Depends how willing you are to get ripped off by scalper prices right now. lol =3