rtx-5090-benchmarks

Rig: one RTX 5090 32GB + 64GB DDR5 (96GB addressable) + 2TB NVMe — the consumer-sovereignty test Model: zai-org/GLM-5.2 — 743B-param / 39B-active MoE (256 routed experts, 8 + 1 shared active), 78 layers, MIT licence, 1M context. Architecture glm_moe_dsa: MLA (Multi-head Latent Attention, compressed-latent KV) + DSA (DeepSeek Sparse Attention). Released 2026-06-16, the largest open-weights drop in months. This is an autopsy, not a benchmark. GLM-5.2 cannot be served on the rig at any quantisation. The point is why, by arithmetic from the published config.json, and what the architecture genuinely buys.

Credit first: MLA makes the KV cache ~57x cheaper

GLM-5.2's KV cache is MLA-compressed. Per token, per layer, it stores a 512-wide latent (kv_lora_rank) plus a 64-wide decoupled RoPE key (qk_rope_head_dim) = 576 values, shared across all 64 heads. A standard attention with these head dims (K 256 + V 256 per head × 64 heads) would store 32,768 values. That is a ~57x smaller KV cache — real engineering, not a press release.

The payoff is concrete (576 × 78 layers × 2 bytes × tokens):

context	MLA KV cache	naive KV would be
64K	~5.5 GiB	~312 GiB
256K	~22 GiB	~1.22 TiB
1M	~88 GiB	~4.9 TiB

At 64K, MLA makes the KV trivial. The 1M headline is where the arithmetic bites.

Wall 1 — the weights never fit addressable memory, at any quant

743B params. In bf16 that is 1.49 TB. To fit 743B into the rig's 96 GB addressable memory (32 VRAM + 64 RAM) you would need 1.03 bits per weight — below any coherent quantisation. The smallest usable quant (unsloth Dynamic 1.58-bit) is **147 GB**, already 1.5x the whole box. So GLM-5.2's weights cannot live in RAM+VRAM at any quant; the only consumer path is NVMe mmap streaming off the 2TB SSD, at well under 1–3 tok/s.

Wall 2 — the 1M KV alone eats the whole machine

Grant MLA its 57x win and the 1M cache is still ~88 GiB ≈ the rig's entire 96 GB working memory, leaving 8 GB for the model's active compute. So even the desperate NVMe-mmap hack — which can limp the weights at small context — is impossible at 1M: the KV cache alone fills the box. Serving GLM-5.2 at its headline context needs **147 GB weights + ~88 GiB KV ≈ 235 GB resident**, past a single H200 (141 GB) and into multi-GPU datacenter territory.

Wall 3 — DSA saves compute, not memory

glm_moe_dsa adds DeepSeek Sparse Attention: a lightning indexer selects a top-k of tokens to attend, cutting the O(N²) attention compute. But the full MLA KV still has to be resident — you cannot attend to what you have evicted — and reference sparse-attention kernels materialise the indexer's N×N scores (the exact wall measured on the Keye-VL-2.0 autopsy). DSA is a throughput win at datacenter scale, not a memory key that unlocks consumer hardware. And the 39B active params (of 743B) cut compute per token, not the memory floor: every expert must still be loadable, because the next token may route to any of them.

The verdict

GLM-5.2 is the largest open-weights release in months, and the team did the genuinely hard work — MLA plus DSA — to make 1M context cheap. It is 57x cheaper. It is still ~235 GB of weights and KV to use that context, on a model whose weights cannot fit a consumer box at any quant. Clever is not the same as consumer. GLM-5.2 is a datacenter model with an open licence, not a model you run on a 5090. The home-lab move is to wait for a GLM-5.2-Air — a ~100–130B variant would be the first that even NVMe-streams usefully — or rent the iron.

Method

All figures derived from the published config.json: num_hidden_layers 78, kv_lora_rank 512, qk_rope_head_dim 64, num_key_value_heads 64, qk_nope_head_dim 192, v_head_dim 256, max_position_embeddings 1048576; plus the published 743B-total / 39B-active counts. KV cache = (kv_lora_rank + qk_rope_head_dim) × layers × 2 bytes × tokens; naive KV = 2 × heads × head_dim × layers × 2 bytes × tokens. Quant floor scaled from unsloth's measured DeepSeek-R1 671B 1.58-bit GGUF (131 GB). No weights were downloaded or served.