Quote from: Steve Nordquist on Today at 09:39:315090 gets 24 GiB GDDR7
It's a "5090 Laptop" / "5090 Mobile", using GB203 chips (256-bit memory bus width, 896 GB/s), but is using 3 GB per memory chip, instead of 2 GB, like on the 5070 Ti / 5080 desktop cards.
See en.wikipedia.org/wiki/Blackwell_(microarchitecture)#Blackwell_dies.
256-bit/32-bit per chip = 8 chips:
5070 Ti / 5080 desktop: 8 chips * 2 GB per chip = 16 GB VRAM.
5090 Laptop: 8 chips * 3 GB per chip = 24 GB VRAM.
512-bit/32-bit per chip = 16 chips:
5090 desktop: 16 chips * 2 GB per chip = 32 GB VRAM.
Now imagine: 16 chips * 3 GB per chip = 48 GB VRAM, but that would eat into their RTX PRO workstation cards.
For $6k one could get an actual 5090 desktop (512-bit memory bus width, 1792 GB/s) PC and a 5090 desktop also has 32 GB of VRAM, not 24.
192 GB RAM +24 GB VRAM can fit full DeekSeek-V4-Flash AI (huggingface.co/unsloth/DeepSeek-V4-Flash-GGUF) weights and still have plenty of memory left for the context.
Other big 300B AI LLM models quants that also fit into 192 GB RAM + 24 GB VRAM, and where 128 GB RAM are not enough (although dense Qwen3.6-27B stands out in performance and doesn't require a lot of memory, but it being 27B active parameters, putting (a quant) on a GPU's much faster VRAM, not system's RAM, is recommended, so it runs faster):
artificialanalysis.ai/?models=qwen3-5-122b-a10b-non-reasoning%2Cqwen3-5-122b-a10b%2Cmimo-v2-5-0424%2Cqwen3-6-27b-non-reasoning%2Cminimax-m2-7%2Cqwen3-6-35b-a3b-non-reasoning%2Cstep-3-7-flash%2Cqwen3-6-35b-a3b%2Cdeepseek-v4-flash-high%2Cqwen3-6-27b%2Cdeepseek-v4-flash-non-reasoning%2Chy3%2Cgemma-4-31b%2Cgemma-4-31b-non-reasoning%2Cdeepseek-v4-flash%2Cmistral-medium-3-5%2Cgpt-oss-120b%2Cnvidia-nemotron-3-super-120b-a12b&intelligence=artificial-analysis-intelligence-index
(the current DeekSeek-V4-Flash is a preview version and the final version is supposed so be released this month)
llama.cpp (that many GUI tools use as backend and are just wrappers, but llama.cpp has its own nice webUI) DeepSeek V4 support is coming along:
git log --pretty=format:"%h %ad %s" --date=short --grep='deepseek.*v4\|dsv4' -i
91d2fc387 2026-07-20 llama_dsv4: write only used rows in state (#25325)
4937ca83f 2026-07-18 llama-quant : exclude i32 ffn_gate_tid2eid routing table from quantization (#25787)
5d5306bf3 2026-07-17 tests : initialize all tensors in test_dsv4_hc to avoid NaNs in sentinel tensors (#25822)
0dc74e332 2026-07-17 DeepseekV4: Add fused hyper-connection ops (#25585)
33a75f41c 2026-07-15 DeepseekV4: reduce graph splits (#25702)
7f575c39d 2026-07-14 DeepseekV4: fix seq_rm (#25588)
13f2b28b0 2026-07-11 DeepseekV4: clear cache only for seq rather than full (#25521)
00f5442cc 2026-07-11 ggml : add GGML_OP_LIGHTNING_INDEXER that implements DeepSeek V3.2/V4 lightning indexer (#24231)
2ed3c1abb 2026-07-10 llama : make all KQ masks f16 if FA is used, remove zero attention bias, remove raw_k repeats in DeepSeek V4 (#25370)
024c46ae4 2026-07-07 llama: fix quantized kv-cache for dsv4 (#25202)
8c146a836 2026-06-29 DeepSeek V4 (#24162)
Deutsch
English
Español
Français
Italiano
Nederlands
Polski
Português
Русский
Türkçe
Svenska
Chinese
Magyar