Cheapest-per-GB VRAM is not value: two join artifacts produced a 60x and an 80x false ranking, and the real answer is that only five accelerators in this corpus can be ranked at all

Asked for the cheapest configuration that can genuinely serve VRAM, the corpus produced a table where a $310 supply record paired with a 96GB RTX PRO 6000 Blackwell and came out as the best value in the database at $5,636 per 1000 tok/s, and where the RTX 4090 came out as the worst at $456,621 per 1000 tok/s - roughly 80x the RTX 5090's figure, the opposite of every measurement the corpus actually holds. Both figures come from real records. Both are artefacts of the join, not of the hardware.

Two records attach a number to an accelerator the number does not describe. ARTEFACT 1 - PRICE OVER-ATTRIBUTION. data/supply/microcenter-in-store-gpus.json carried accelerator_ids ['intel-arc-b580', 'intel-arc-pro-b60', 'nvidia-rtx-pro-6000-blackwell'] against a single price of 309.99. The price_basis names exactly one item: an ASRock Intel Arc B580 Challenger OC 12GB bare card. The other two ids are parts that appeared on the same listing page. A join that pairs price x benchmark therefore divided a 12GB card's price by a 600W 96GB workstation card's throughput (55 tok/s on Llama 3.3 70B Q8_0) and reported the RTX PRO 6000 as 60x better value than it is. ARTEFACT 2 - THROUGHPUT UNDER-ATTRIBUTION. data/benchmarks/comm-4090-llamacpp-rpc-3-host-decode-tok-s.json held 4.38 tok/s with accelerator_ids ['nvidia-rtx-4090']. That figure is a THREE-HOST llama.cpp RPC cluster totalling 56GB - one RTX 4090 plus two RTX 4060 Ti - on the reporter's first attempt, single request, n_ctx_slot 8192. It is not a per-card 4090 figure. Filed under the 4090's id it became the 4090's only decode benchmark and dragged the card to the bottom of the ranking.

Both records were corrected on 2026-10-11 and the ranking recomputed. The join rules that follow from them are the real deliverable: RULE A - a supply record's price attaches to the item its price_basis NAMES, not to every accelerator the page mentions. A listing page covering several parts needs either one accelerator_id plus a basis that names that part, or per-part prices. RULE B - a per-card tok/s is the denominator of every value figure in this corpus. A cluster, RPC, or multi-host figure must not be filed under one card's accelerator_ids. Where no record type fits a mixed configuration, leave the list empty and name the machines in the methodology text. With both artefacts removed - and a third swept on 2026-10-12, when the RTX 5080's $6,549.99 'price' turned out to be the RTX 5090's list band over-attributed to it - the honest local value table (cheapest non-cloud purchase price per card, batch size 1 in every row, from records in this corpus). The first column is CAPEX PER 1000 tok/s of measured decode throughput - capex / tok/s x 1000 - so LOWER is better: $ per 1000 tok/s accelerator capex tok/s model (VRAM) 2976 intel-arc-b580 $ 200 67.2 R1-Distill-Qwen 7B Q4_K_M (12GB) 5618 intel-arc-b580 $ 200 35.6 R1-Distill-Qwen 14B Q4_K_M (12GB) 8573 amd-radeon-rx-9070-xt $ 829 96.7 Qwen3 8B Q4 (16GB) 14279 nvidia-rtx-5090 $ 1,999 140.0 Qwen3-14B-AWQ (32GB) 23518 nvidia-rtx-5090 $ 1,999 85.0 Qwen3 32B Q4_K_M (32GB) 29867 amd-radeon-rx-7900-xtx $ 1,406 47.06 Qwen3.8-27B Q4_K_XL (24GB) 44422 nvidia-rtx-5090 $ 1,999 45.0 Llama 3.3 70B Q4_K_M (32GB) 302254 nvidia-rtx-pro-6000-bl $ 16,624 55.0 Llama 3.3 70B Q8_0 (96GB) Read the ordering carefully - it is dominated by which model each card was measured on, not by the cards. The Arc B580's leading figure is a refurb-tier price ($199.99 Newegg Refreshed) on a 7B distilled model, and the RTX PRO 6000's trailing figure comes from running a 70B at Q8_0 at 55 tok/s on a $16.6k card - neither says much about the card alone. Cross-model comparisons within this table are not value findings. THE HONEST LIMIT: these eight rows across FIVE accelerators are the only honest capex/decode joins in the entire corpus. 'Price' here means a non-cloud purchase price - not a per-GPU-hour rental, and not the $/kWh us-electricity-industrial-tariff-2024 energy record, which is not a supply price at all. Seven accelerators carry a non-cloud purchase price but no decode benchmark and are therefore unknowns, not picks: RTX 3090 ($1,100 refurb, 24GB), RTX 4090 ($2,000, 24GB), A100 40GB SXM4 ($2,499 used module, 40GB), A100 80GB SXM4 ($10,450, 80GB), A100 80GB PCIe ($12,099, 80GB), H100 PCIe ($43,000, 80GB), H800 80GB ($41,221). The RTX 5080 is the mirror-image unknown: it has a batch-1 decode benchmark (65 tok/s on Qwen3 32B, local-rtx-5080-32b-q4-decode) but no surviving purchase price - its only priced record was the $6,549.99 5090 band over-attributed to it. The RTX 5070 Ti and RTX 4080 lost their only prices to the same class of over-attribution and now have neither half, and the 3090 Ti, H100/H200 SXM, B200 and B300 are priced by nothing but the energy tariff or by nothing at all. The corpus cannot say which of them is good value, only which has been measured.

FILED 2026-10-11 (parent, in place of a subagent that died on HTTP 429 before making a tool call). The question asked was what the cheapest genuinely-usable VRAM configuration is, comparing 2x Tesla M40 (48GB) against 1x P40 (24GB) against a single 4090 (32GB). That comparison cannot be answered from this corpus, and the reason is itself the finding: the M40 and P40 have supply records and no decode benchmark of any kind, so the corpus holds their price and nothing about what they run. Tesla M40 24GB at $189 is $7.88/GB and is the cheapest VRAM in the database; it is also tensor-coreless Maxwell with compute capability 5.2, so it cannot run fast decode, and the corpus has no tok/s number that would let a reader see that for themselves. The one thing the corpus CAN say from real paired records: on Strix Halo (gfx1151, 128GB LPDDR5X, 256 GB/s bus), dense Qwen2.5-72B Q5_K_M runs at 4.1 tok/s while GPT-OSS 120B MXFP4 MoE reaches 49.0 tok/s on the same machine - see llamacpp-strix-halo-8060s-qwen2-5-72b-q5-k-m-decode and llamacpp-strix-halo-8060s-gpt-oss-120b-mxfp4-decode. A 12x difference decided entirely by whether the model is dense, because only active experts are streamed. That is the strongest value-shaped result the corpus holds, and it is about model choice rather than hardware choice. PICKS BY MODEL SIZE, as the value brief asked. Same corpus records, split by which model class the card has actually been measured on, because a card's rank depends entirely on which row you read (figures are $ per 1000 tok/s; lower is better): SMALL/MEDIUM MODELS (7-32B) - the cheap cards lead on the pure ratio, on small models: intel-arc-b580 $ 200 67.2 tok/s R1-Distill-Qwen 7B Q4_K_M batch 1 = $ 2,976 per 1000 tok/s intel-arc-b580 $ 200 35.6 tok/s R1-Distill-Qwen 14B Q4_K_M batch 1 = $ 5,618 per 1000 tok/s amd-radeon-rx-9070-xt $ 829 96.7 tok/s Qwen3 8B Q4 batch 1 = $ 8,573 per 1000 tok/s nvidia-rtx-5090 $ 1,999 140.0 tok/s Qwen3-14B-AWQ (awq_marlin) batch 1 = $ 14,279 per 1000 tok/s nvidia-rtx-5090 $ 1,999 85.0 tok/s Qwen3 32B Q4_K_M batch 1 = $ 23,518 per 1000 tok/s amd-radeon-rx-7900-xtx $ 1,406 47.06 tok/s Qwen3.8-27B Q4_K_XL batch 1 = $ 29,867 per 1000 tok/s On the pure ratio the Arc B580 leads this class - but at a refurb price on a 7B distill, and the cross-model caveat applies in full. The 9070 XT is the strongest new-purchase ratio. The 7900 XTX remains the only AMD card in the corpus with a decode benchmark on a mid-size (27B) model. LARGE MODELS (70B): nvidia-rtx-5090 $ 1,999 45.0 tok/s Llama 3.3 70B Q4_K_M batch 1 = $ 44,422 per 1000 tok/s nvidia-rtx-pro-6000-bl $16,624 55.0 tok/s Llama 3.3 70B Q8_0 batch 1 = $302,254 per 1000 tok/s The 5090 runs 70B Q4_K_M at 45 tok/s on 32GB. The RTX PRO 6000 runs 70B at Q8_0 - a materially higher-quality quant - on 96GB, and its poor ratio is a consequence of paying $16.6k for headroom and precision, not of being slow. These two are not competing for the same buyer. VRAM GATES, which decide the class before any ratio does: nvidia-rtx-5090 32GB - fits 70B Q4_K_M, not 70B at Q8 amd-radeon-rx-7900-xtx 24GB - fits 27B Q4_K_XL (17.5GB), not 70B nvidia-rtx-5080 16GB - 32B Q4 only, no 70B row exists intel-arc-b580 12GB - fits 8-14B only amd-radeon-rx-9070-xt 16GB - fits 32B Q4 A card that cannot fit the model is not a candidate at any price, so the VRAM gate is applied before the tok/s-per-dollar split rather than after. | PRICE-ATTRIBUTION CORRECTION 2026-10-11 (crawler/w17-dc-capex): the RTX 4080 capex ($16,624) and the RTX 5070 Ti capex ($6,550) quoted in this gotcha were NOT prices for those cards. Both were over-attributions from supply records whose price_basis names a different part (the RTX PRO 6000 Blackwell, and the RTX 5090 respectively). The offending accelerator_ids have been removed from bh-photo-used-workstation-gpus, bh-photo-workstation-gpus and new-retail-rtx-5090-amazon. After the fix NEITHER the RTX 4080 NOR the RTX 5070 Ti has any non-cloud price_usd in the corpus, so neither can be ranked on value at all. | UNIT RELABEL 2026-10-12 (repo-hygiene wave, retiring ledger D12): the value column was headed 'tok/s per $1000' but every value is capex / tok/s x 1000 - dollars per 1000 tok/s of decode throughput, lower is better. Relabeled throughout. The ordering conclusion was re-read after the relabel and corrected: the earlier claim that the 7900 XTX was 'the best value in this class on a pure ratio' read the inverted label at face value - on the true ratio the cheap cards on small models lead, and in the same class the 5090 beats the 7900 XTX. The table was also recomputed against the current corpus: the RTX 5080's $6,550 'price' was the 5090's band over-attributed (row removed; the card is now a benchmark without a price), and the RX 9070 XT and Arc B580 rows join - their batch-1 decode benchmarks (local-rx9070xt-qwen3-8b-q4-decode-b1, local-arc-b580-r1-qwen-7b/14b-q4-decode-b1) landed in the same wave-17 change that fixed the 4080/5070 Ti prices, one commit after this gotcha was filed. The same pass removed the remaining RULE A over-attributions the joins still carried: the 5090 off the B&H workstation record, H100/H200 SXM off the eBay A100 record, the 3090 Ti off the Micro Center refurb record (its real $1,239.99 tier was already in the basis), the 4090 and B580 off the Newegg 3090 record, the 3090 off the PicClick 4090-median record, and the eBay 'a100-40gb' record re-pointed at the 40GB accelerator its own basis names.