Best GPUs for AI and ML Training (2026)

TL;DR

• Best overall: msi Gaming RTX 5090 32G Ventus 3X OC Graphics Card (32GB GDDR7, 512-bit, NVIDIA Blackwell Architecture) — Check price on Amazon
• Best value: Gigabyte 24GB NVIDIA GeForce RTX 3090 Turbo GDDR6X Graphics Card Model GV-N3090TURBO-24GD (Renewed) — $1,894.99 renewed (as of 11/06/2026 14:10 UTC - Details) · Check price on Amazon
• Best budget: ASUS Dual NVIDIA GeForce RTX 5060 Ti 16GB GDDR7 OC Edition Graphics Card — Check price on Amazon

GPUs for AI training split into two worlds: consumer cards that individual researchers and small teams buy, and datacenter accelerators that are rented by the hour. On a single consumer card, training mostly means fine-tuning with LoRA or QLoRA adapters. Compute Market says a 24GB card can fine-tune 7B to 13B models with 4-bit quantization, and that full fine-tuning or training from scratch above 7B needs enterprise GPUs or multi-GPU setups; RunPod's arithmetic puts full Adam fine-tuning of a 7B model at about 112GB before activations. [src1, src12]

The RTX 5090 stays the top pick. Puget Systems tested it for AI and calls it "theoretically the best consumer GPU for AI workloads that money can buy", although it found software support immature at launch. StorageReview measured it 16% ahead of the RTX 4090 in AI text generation and 56% ahead in image generation. Most published GPU tests measure inference; the training evidence comes from Nikolay Falaleev's PyTorch benchmarks, where the 5090 was about 44% faster than the RTX 4090 and about 132% faster than the RTX 3090 on average across his vision models. [src3, src4, src5]

The RTX 3090 stays the value pick. It has the same 24GB as the RTX 4090 and costs about $700 to $1,000 used: Compute Market calls it "the budget king of AI hardware", and Hardware Corner "the most cost-effective second-hand entry point" for local LLM work. Two of them give 48GB; ServeTheHome found a pair with an NVLink bridge nearly doubled a single card's deep learning results. [src1, src2, src10, src11]

The 16GB cards — RTX 5080, RTX 5070 Ti and RTX 4080 SUPER — fit the same models: Compute Market says 16GB limits the RTX 5080 and RTX 5070 Ti to the model sizes of the RTX 4080 SUPER. Tom's Hardware found the RTX 5070 Ti ahead of the 4080 SUPER in its AI tests at a lower launch price, and called the RTX 5080 disappointing for minor gains over the 4080 SUPER and for sticking with 16GB. [src1, src8]

The RTX 5060 Ti 16GB is the new budget pick: the cheapest new card that fits a 7B QLoRA run. Compute Market recommends it over a used RTX 3090 for 7B–13B work if you want a warranty and lower power draw, and Hardware Corner calls it the "value king" for entry-level local LLM work; Tom's Hardware measured it 40% faster than the RTX 4060 Ti 16GB in AI text generation. [src15, src16, src17] It replaces the RTX 4070 Ti SUPER, this card's budget pick until now: no outlet we read recommends the 4070 Ti SUPER for AI training in 2026, and Tom's Hardware says the RTX 5070 Ti took over where it left off. [src8]

Prices are the problem in 2026. Tom's Hardware wrote in April 2026 that a new RTX 5090 cost nearly twice its $1,999 launch price. [src6, src9] In our checks on 2026-10-07 the RTX 5090 listing we link had no offer, the RTX 5060 Ti was the only pick sold by Amazon.com itself, and every other pick was offered only by marketplace sellers, at prices far above the list prices outlets quote.

For more than 32GB per GPU, rent. RunPod lists the 80GB H100 SXM at $3.49 an hour, the 141GB H200 at $4.59, the B200 (about 180GB usable) at $6.79 and a 48GB L40S at $1.09. [src12] Workstation cards have their own card, and they have become expensive: Tom's Hardware reported in August 2026 that NVIDIA lists the 96GB RTX PRO 6000 Blackwell at $16,000. [src14]

Top Pick: NVIDIA GeForce RTX 5090 (32GB; no purchasable offer at Amazon in our last three daily checks, through 2026-10-07) — Check price

Puget Systems tested the RTX 5090 Founders Edition for AI and concluded: "With 32GB of GDDR7 running at 1.79TB/s and impressive compute capabilities, the RTX 5090 is theoretically the best consumer GPU for AI workloads that money can buy." In its llama.cpp tests the 5090 led the RTX 4090 by about 29% in token generation, but Puget found the software support for the new card too immature at launch, in February 2025, to give a full assessment. [src3] StorageReview measured the RTX 5090 16% ahead of the RTX 4090 across Procyon's AI text generation tests and 56% ahead in image generation, and ahead of the RTX 6000 Ada in every one of them. [src4] For training itself, Nikolay Falaleev's PyTorch benchmarks put the RTX 5090 about 44% ahead of the RTX 4090 on average across his vision models, at about 28% more power: in FP16 it trained ResNet-50 at 1,623.9 images per second, against 1,360.6 on the RTX 4090 and 888.7 on the RTX 3090. [src5] Compute Market ranks it first and calls it "the most powerful consumer GPU ever built for AI workloads"; it notes that the 575W card needs a power supply of 1,000W or more. [src1] Its 32GB sits between RunPod's 24GB class for 13B QLoRA and its 48GB class for 30B, and RunPod rates the RTX 5090 "strong for LoRA, diffusion training and fast iteration". [src12] Price: it launched at $1,999, and Tom's Hardware wrote in April 2026 that a new 5090 cost nearly twice that. [src6, src9] The one RTX 5090 offer our checks found on 2026-10-07, for an ASUS ROG Astral RTX 5090 listed on another of our cards, was $6,898.99 from a marketplace seller. Availability: Puget, StorageReview and Tom's Hardware tested NVIDIA's Founders Edition; the listing we link is MSI's Ventus 3X OC version of the same GPU, and it had no purchasable offer in our last three daily checks, through 2026-10-07. That is an availability note, not a change to the ranking. Best you can buy today instead: the RTX 3090 below, in stock as a renewed card at $1,794.99 from a marketplace seller when we checked on 2026-10-07; for more speed at the same 24GB, the RTX 4090 was $4,499.00 from a marketplace seller, with low stock, that day.

Best Value: NVIDIA GeForce RTX 3090 (24GB, used or renewed; $1,794.99 renewed from a marketplace seller in our check on 2026-10-07) — Check price

Compute Market calls the RTX 3090 "the budget king of AI hardware" and lists it at $699 to $999 used; its budget guide goes further: "The RTX 3090 is, without question, the best value GPU for AI in 2026". Compute Market puts LoRA and QLoRA fine-tuning of models up to 13B within the 3090's 24GB. [src1, src2] Hardware Corner, which tested it with current models, calls the RTX 3090 "the most cost-effective second-hand entry point for anyone serious about local LLM work"; it ran 30B to 35B models at 4-bit and puts the card's average US market value at $1,000. [src11] It is the slowest of the three cards Falaleev tested: in his FP16 tests the RTX 3090 trained ResNet-50 at 888.7 images per second, against 1,360.6 on the RTX 4090 and 1,623.9 on the RTX 5090. [src5] Two of them scale well: ServeTheHome's dual RTX 3090 tests with an NVLink bridge, which include TensorFlow ResNet-50 training, found that "Deep Learning is where a dual GeForce RTX 3090 configuration will shine", with nearly double the results of a single card. [src10] The RTX 3090 launched at $1,499 in September 2020, with 24GB of GDDR6X at 936 GB/s and a 350W board power. [src6] Buying used: Compute Market warns that cards used for mining may have been thermally stressed, and advises testing the fans, running a stress test and buying where you can return the card. [src1] Price and availability: the listing we link is Gigabyte's blower-style RTX 3090 Turbo, renewed. It was $1,794.99 from QyTech, a marketplace seller, in stock, in our check on 2026-10-07: about $75 per GB of VRAM, against about $29 to $42 per GB at the used prices Compute Market lists.

Best Budget: NVIDIA GeForce RTX 5060 Ti 16GB ($792.99 sold by Amazon.com in our check on 2026-10-07) — Check price

Compute Market's budget guide puts "QLoRA on 7B-13B models with reasonable batch sizes" in its 16GB tier and names the RTX 5060 Ti 16GB for it, at the $429 to $479 it lists. [src2] Its head-to-head with the used RTX 3090 recommends the RTX 5060 Ti, with its 180W TDP, if you mainly work with 7B–13B models, want a full manufacturer warranty and care about power and noise; the 3090 is its choice for 30B+ models. [src15] Hardware Corner, which tested the RTX 5060 Ti 16GB, calls it the "value king" for entry-level local LLM work; it ran models up to 20B and puts the card's average market value at $549. [src17] Tom's Hardware tested Asus and PNY versions and measured the RTX 5060 Ti 16GB 40% faster than the RTX 4060 Ti 16GB in MLPerf Client's Llama 2 7B text generation; it strongly advises buying the 16GB model rather than the 8GB one. [src16] The RTX 5060 Ti 16GB is a training card for small models only: 16GB of GDDR7 at 448 GB/s, less than half the RTX 3090's bandwidth, and 23.7 FP32 TFLOPS. [src6, src16] RunPod's QLoRA budget for a 7B model is about 16GB. [src12] Price and availability: the RTX 5060 Ti 16GB launched at $429 in April 2025. The ASUS Dual OC listing we link was $792.99, sold by Amazon.com, in stock, in our check on 2026-10-07 — about 85% above launch, and inside the $699 to $999 range Compute Market lists for a used RTX 3090 with 24GB. [src1, src16] The RTX 4070 Ti SUPER that held this slot before is dropped: no outlet we read recommends it for AI training in 2026, and Tom's Hardware says the RTX 5070 Ti takes over "from where the 4070 Ti Super left off". [src8]

Previous Flagship, 24GB: NVIDIA GeForce RTX 4090 ($4,499.00 from a marketplace seller, low stock, in our check on 2026-10-07) — Check price

Compute Market calls the RTX 4090 "the GPU that made local AI mainstream, and it remains the card we recommend most often", and ranks it its best value at the $1,599 to $1,999 it lists. [src1] RunPod calls the 24GB RTX 4090 "the cheapest way to get real training throughput" in its cloud. [src12] It is the second-fastest card here for training: Falaleev measured the RTX 4090 at 1,360.6 images per second in FP16 ResNet-50 training, 53% ahead of the RTX 3090 and behind the RTX 5090. Puget measured the RTX 5090 about 29% ahead of it in token generation, and StorageReview 16% to 56% ahead in AI text and image generation. [src3, src4, src5] The RTX 4090 has 24GB of GDDR6X at 1,008 GB/s and a 450W board power, and launched at $1,599 in October 2022. [src6] Tom's Hardware notes that local-LLM buyers compete for used 4090s. [src9] Price and availability: the MSI Gaming X Trio listing we link was $4,499.00 from Everlast Computers, a marketplace seller, with low stock, in our check on 2026-10-07 — more than twice the price range Compute Market lists for it. At that price it is not a value pick: buy it only if you find one far cheaper and 24GB is enough.

Fastest 16GB Card: NVIDIA GeForce RTX 5080 ($2,379.99 from a marketplace seller in our check on 2026-10-07) — Check price

Puget Systems tested the RTX 5080 Founders Edition, the model we link. In llama.cpp token generation it found the RTX 5080 "landing just short of the RTX 4090's performance"; in SD.Next image generation the RTX 5080 and RTX 4080 SUPER "achieve practically identical scores"; and it notes that the RTX 5080 "has disappointed some by maintaining a total of 16GB of memory". [src3] Tom's Hardware called the RTX 5080 "disappointing for only offering minor performance improvements over the existing 4080 / 4080 Super, and for sticking with 16GB of VRAM". [src8] Its 16GB of GDDR7 runs at 960 GB/s; it launched at $999 in January 2025 with a 360W board power. [src7] Compute Market says 16GB limits the RTX 5080 and RTX 5070 Ti to the same model sizes as the RTX 4080 SUPER, and RunPod lists the RTX 5080 for small LoRA and QLoRA jobs. [src1, src12] Price and availability: $2,379.99 from Grandis Goods, a marketplace seller, in stock, in our check on 2026-10-07 — $850.99 more than the RTX 5070 Ti that day, for the same 16GB.

Newest 16GB Card for Less: NVIDIA GeForce RTX 5070 Ti ($1,529.00 from a marketplace seller, low stock, in our check on 2026-10-07) — Check price

Tom's Hardware tested the RTX 5070 Ti, an Asus Prime card, and found it ahead of the RTX 4080 SUPER in both AI tests that ran on it (Procyon did not yet support RTX 50-series cards). In MLPerf Client's Llama 2 7B text generation it was 17% faster than the RTX 4070 Ti SUPER, and in SPECworkstation's AI inference test 18% faster. Tom's Hardware says the RTX 5070 Ti takes over "from where the 4070 Ti Super left off" and "can get away with 16GB by virtue of costing $749 — again, in theory". It has 16GB of GDDR7 at 896 GB/s and a 300W board power. [src8] Compute Market's verdict on the RTX 5070 Ti and RTX 5080 is that 16GB limits them to the same model sizes as the RTX 4080 SUPER. [src1] Price and availability: the listing we link is MSI's Gaming Trio OC Plus version. It was $1,529.00 from GH Electronics, a marketplace seller, with low stock, in our check on 2026-10-07 — about twice its $749 launch price.

Previous-Generation 16GB: NVIDIA GeForce RTX 4080 SUPER ($2,099.99 from a marketplace seller in our check on 2026-10-07) — Check price

Compute Market names the RTX 4080 SUPER its best mid-range card: "The RTX 4080 SUPER occupies the pragmatic middle ground. At 16GB GDDR6X, it handles 7B–13B models comfortably and draws only 320W". [src1] Puget measured it practically level with the RTX 5080 in SD.Next image generation. [src3] In Tom's Hardware's AI tests the RTX 5070 Ti edged past the RTX 4080 SUPER. [src8] The RTX 4080 SUPER has 16GB of GDDR6X at 736 GB/s and launched at $999 in January 2024. [src7] Price and availability: $2,099.99 from Grandis Goods, a marketplace seller, in stock, in our check on 2026-10-07 — $570.99 more than the RTX 5070 Ti that day. Buy it only below the RTX 5070 Ti's price.

Large-Scale Training: rent an NVIDIA H100, H200 or B200

For full fine-tuning above 7B or for pre-training, rent datacenter GPUs by the hour. At RunPod's Secure Cloud rates, updated 2026-09-29: the H100 SXM costs $3.49 an hour ($2.89 for PCIe) and, with 80GB and FP8 through the Transformer Engine, is "still the workhorse if your stack is tuned for Hopper"; the H200 SXM costs $4.59 with 141GB at 4.8 TB/s, a gain RunPod calls "almost entirely memory, which is where most training is actually stuck"; and the B200 costs $6.79 with 192GB (about 180GB usable), 8 TB/s and FP4, which RunPod says puts it "clearly ahead on both memory-bound and compute-bound training". RunPod also lists the B300, with 288GB, from $6.94 an hour. [src12] Northflank gives the H100 3.35 TB/s of memory bandwidth. [src13]

Best AMD Option: AMD MI300X (cloud)

RunPod lists the MI300X at $2.39 an hour, checked on 31 August 2026: 192GB of HBM3 at about 5.3 TB/s. RunPod recommends it when you need 192GB per GPU and your stack is ROCm-ready. [src12, src13]

More Than 32GB on One Desk: see our workstation GPUs card

Workstation cards are covered on our workstation GPUs card, and they have become expensive: Tom's Hardware reported in August 2026 that NVIDIA lists the 96GB RTX PRO 6000 Blackwell at $16,000 on its US store, up from $8,565 at launch. [src14] The RTX 6000 Ada Generation that this card listed before, as "RTX A6000 Ada", is dropped: StorageReview measured the RTX 5090 ahead of the RTX 6000 Ada in every Procyon AI test it ran. [src4] To rent 48GB instead, RunPod lists an L40S at $1.09 an hour, which it calls "the sweet spot for mid-scale fine-tuning where 24–32GB is cramped". [src12]

Head-to-Head Comparisons

RTX 5090 vs RTX 4090

The RTX 5090 has 32GB of GDDR7 against the RTX 4090's 24GB of GDDR6X, 78% more memory bandwidth (1,792 against 1,008 GB/s) and a 575W board power against 450W. [src6] In Falaleev's PyTorch benchmarks the RTX 5090 was about 44% faster on average across his vision models, at about 28% more power; in FP16 ResNet-50 training it reached 1,623.9 images per second against 1,360.6. [src5] Puget measured the RTX 5090 about 29% faster in llama.cpp token generation, and StorageReview 16% faster in Procyon AI text generation and 56% faster in image generation. [src3, src4] In our check on 2026-10-07 the RTX 5090 listing had no offer and the RTX 4090 was $4,499.00 from a marketplace seller.

Pick RTX 5090 if: you need 32GB or the fastest single consumer card for training.
Pick RTX 4090 if: 24GB is enough and you find one well below an RTX 5090's price.

RTX 5090 vs RTX 3090 (Used)

The RTX 5090 has 8GB more VRAM and nearly twice the memory bandwidth of the RTX 3090 (1,792 against 936 GB/s). [src6] Falaleev measured the RTX 5090 about 132% faster than the RTX 3090 on average across his vision models — 1,623.9 against 888.7 images per second in FP16 ResNet-50 training — and about twice as fast in LLM inference, at 210.79 against 100.53 tokens per second on Llama 3.1 8B. [src5] The RTX 3090 costs $699 to $999 used by Compute Market's figures, and two of them give 48GB in total: ServeTheHome found two RTX 3090s with an NVLink bridge nearly doubled a single card's deep learning results. [src1, src10]

Pick RTX 5090 if: you want the fastest single card and can find one.
Pick RTX 3090 (used) if: VRAM per dollar matters more than speed, or you plan to run two cards.

RTX 3090 (Used) vs RTX 5060 Ti 16GB

The used RTX 3090 has 24GB against the RTX 5060 Ti's 16GB and more than twice its memory bandwidth (936 against 448 GB/s), but draws 350W against 180W. [src6, src16] Compute Market's verdict: the RTX 3090 for 30B+ models, if you accept no warranty and possible mining wear; the RTX 5060 Ti for 7B–13B models with a full warranty, lower power and less noise. It lists $699 to $999 for a used 3090 and $429 to $479 for a new 5060 Ti. [src15] In our check on 2026-10-07 the RTX 5060 Ti listing we link was $792.99 from Amazon.com, and the renewed RTX 3090 listing $1,794.99 from a marketplace seller.

Pick RTX 3090 (used) if: you need 24GB, for example 13B QLoRA, and can buy from a seller who takes returns.
Pick RTX 5060 Ti 16GB if: your runs fit in 16GB and you want a new card with a warranty.

RTX 5080 vs RTX 5070 Ti

Both have 16GB of GDDR7, so they fit the same models. The RTX 5080 has 960 GB/s against the RTX 5070 Ti's 896 GB/s and 10,752 shaders against 8,960, and launched at $999 against $749. [src7, src8] Tom's Hardware called the RTX 5080 disappointing for minor gains over the RTX 4080 SUPER, and found the RTX 5070 Ti ahead of the 4080 SUPER in its AI tests. [src8] Puget measured the RTX 5080 just short of the RTX 4090 in llama.cpp token generation. [src3] In our check on 2026-10-07 the RTX 5080 was $2,379.99 and the RTX 5070 Ti $1,529.00, both from marketplace sellers.

Pick RTX 5080 if: you want the fastest 16GB card and the price gap is small.
Pick RTX 5070 Ti if: you want 16GB of current-generation memory for the least money.

H100 vs H200 (and the A100)

The H200 keeps the H100's Hopper compute and adds memory: 141GB of HBM3e at 4.8 TB/s against 80GB of HBM3 at 3.35 TB/s. [src12, src13] At RunPod it costs $4.59 an hour against $3.49 for the H100 SXM. The A100 SXM, at $1.59, has 80GB but no FP8, so RunPod says it is slower than Hopper but "often wins on cost per finished run". [src12]

Pick H100 if: your model and batch fit in 80GB and your stack is tuned for Hopper.
Pick H200 if: you keep running out of memory: long contexts, large batches or 70B+ fine-tuning.
Pick A100 if: your stack does not use FP8 and cost per finished run matters most.

H100 vs AMD MI300X

The MI300X has 192GB of HBM3 at about 5.3 TB/s, against the H100's 80GB at 3.35 TB/s, and costs $2.39 an hour at RunPod against $3.49 for the H100 SXM. RunPod recommends it only for ROCm-ready stacks; the H100's FP8 through the Transformer Engine is mature on Hopper-tuned stacks. [src12, src13]

Pick H100 if: you want the broadest software support.
Pick MI300X if: you need 192GB per GPU and your stack is ROCm-ready.

Decision Logic

If you can buy one card and want the fastest training

→ The RTX 5090: 32GB and about 44% faster than the RTX 4090 in Falaleev's vision-model benchmarks. The listing we link had no purchasable offer in our checks through 2026-10-07, so expect to pay well above its $1,999 launch price where you find one. [src5, src6, src9]

If you want the most VRAM for the money

→ A used RTX 3090: 24GB for $699 to $999 by Compute Market's figures, or about $1,000 on average by Hardware Corner's. Buy from a seller who accepts returns, and stress-test it. [src1, src11]

If you want a new card for the least money

→ The RTX 5060 Ti 16GB, enough for 7B QLoRA; the listing we link was $792.99 from Amazon.com in our check on 2026-10-07. At that price a used RTX 3090 with 24GB can cost about the same, by Compute Market's figures. [src1, src15]

If you need more than 24GB but cannot get an RTX 5090

→ Two used RTX 3090s (48GB in total), which ServeTheHome found nearly doubled a single card's deep learning results with an NVLink bridge — or rent a 48GB L40S at $1.09 an hour. [src10, src12]

If your models fit in 16GB (for example 7B QLoRA) and you want more speed than the RTX 5060 Ti

→ The RTX 5070 Ti, which beat the RTX 4080 SUPER in Tom's Hardware's AI tests. Pay more for the RTX 5080 only for its extra speed; buy the RTX 4080 SUPER only below the RTX 5070 Ti's price. [src8, src12]

If you are doing full fine-tuning above 7B or pre-training

→ Rent: an H100 if 80GB fits, an H200 for memory-bound or long-context training, a B200 for 70B+ runs. [src1, src12]

If CUDA compatibility is not required

→ The AMD MI300X in the cloud: 192GB per GPU at $2.39 an hour at RunPod, for ROCm-ready stacks. [src12]

Default recommendation

→ The RTX 5090 if you can find one at a price you accept; otherwise a used RTX 3090. [src1, src3]

• The memory shortage has pushed prices far above launch: Tom's Hardware wrote in April 2026 that a new RTX 5090 cost nearly twice its $1,999 launch price, and in August 2026 that NVIDIA lists the 96GB RTX PRO 6000 Blackwell at $16,000. In our checks on 2026-10-07 every pick on this card with an Amazon offer except the RTX 5060 Ti was sold by a marketplace seller, the RTX 5070 Ti at about twice its launch price and the RTX 5060 Ti at about 85% above its own. [src6, src9, src14, src16]
• Outlets' prices lag the market: Compute Market's guide, updated 2026-10-04, lists the RTX 5090 at $1,999 to $2,199 and the RTX 4090 at $1,599 to $1,999 — far below the offers we observed. [src1]
• 24GB cards are in demand: Tom's Hardware suggests gamers sell their RTX 3090s to LLM enthusiasts "while the market is hot", and Hardware Corner puts the RTX 3090's average US market value at $1,000. [src9, src11]
• The 16GB tier moved to Blackwell: Tom's Hardware says the RTX 5070 Ti took over from the RTX 4070 Ti SUPER at a $50 lower launch price, found the RTX 5080 only a minor step over the RTX 4080 SUPER, and measured the RTX 5060 Ti 16GB 40% ahead of the RTX 4060 Ti 16GB in AI text generation. [src8, src16]
• Blackwell is in the cloud: RunPod rents the B200 at $6.79 an hour and the 288GB B300 from $6.94. [src12]
• Most published GPU tests measure inference: Puget, StorageReview and Tom's Hardware tested the RTX 5090, RTX 5080, RTX 5070 Ti and RTX 5060 Ti on LLM and image generation; the training-throughput figures on this card come from vision-model tests: Falaleev's of the RTX 3090, RTX 4090 and RTX 5090, and ServeTheHome's of the RTX 3090. [src3, src4, src5, src8, src10, src16]

• Observed prices are the Amazon offers our checks found on 2026-10-07, all but one from marketplace sellers; they move daily, and outlets' list prices can be far below them. [src1]
• The training benchmarks on this card are vision models (ResNet-50 and others) in PyTorch and TensorFlow; the LLM figures are inference. Compute Market says it tested LoRA fine-tuning but publishes no training figures. [src1, src5, src10]
• A used or renewed RTX 3090 carries risk: Compute Market warns about mining wear and advises testing the fans, running a stress test and keeping a return option. [src1]
• The RTX 5090's 575W board power needs a power supply of 1,000W or more, by Compute Market's advice. [src1, src6]
• Full fine-tuning needs far more memory than the model's size: about 16 bytes per parameter before activations, about 112GB for a 7B model. [src12]
• The MI300X needs a ROCm-ready stack. [src12]
• FP4 on Blackwell helps only if your software supports it: Puget found support immature at the RTX 5090's launch, and RunPod says FP4 "can materially change throughput and effective batch sizes if your stack supports it". [src3, src12]