In this article
Short answer
This build puts a 32 GB GeForce RTX 5090 at the centre, with a Core Ultra 7 265K, 96 GB of DDR5 and a 4 TB NVMe drive. VRAM decides which language models run at speed, and 32 GB holds 32B-class models at 4-bit. It suits people who run large local models daily and want long context.
- ComputingForGeeks puts a 32B model at 4-bit near 20 GB. That fits here and does not fit a 16 GB card.
- INNO3D lists a 1000 W minimum system power for the card, so the build uses a 1000 W supply with a native 16-pin cable.
- INNO3D lists the card's length as 333, with a width of 60, and gives no unit. Check the maker's sheet against the case, which was chosen around a card of that size.
- 96 GB of system memory gives room for layers that spill out of VRAM when you try the very largest models.
- CPUIntel Core Ultra 7 265K LGA1851 Desktop Processor
- GPUINNO3D GeForce RTX 5090 X3 OC 32GB
The parts list#
| Component | Part | Why |
|---|---|---|
| CPU | Intel Core Ultra 7 265K LGA1851 Desktop Processor | 20 cores, DDR5 to 6400 MT/s and PCIe 5.0 for the card and drive |
| Motherboard | ASUS ROG Strix Z890-A Gaming WiFi ATX Motherboard | PCIe 5.0 x16 slot, five M.2 slots, four DIMM slots |
| GPU | INNO3D GeForce RTX 5090 X3 OC 32GB | 32 GB of GDDR7 and 1792 GB/s of bandwidth |
| RAM | TeamGroup T-Create Expert 96GB (2x48GB) DDR5-6800 CL36 Kit | 96 GB in two modules for spill-over and large contexts |
| Storage | Samsung 990 PRO 4TB PCIe 4.0 NVMe M.2 SSD | Room for dozens of model files, 2400 TBW endurance |
| PSU | Corsair RM1000x 1000W ATX 3.1 Gold Power Supply | Meets INNO3D's 1000 W minimum, native 12V-2x6 cable |
| CPU cooler | Noctua NH-D15 G2 Dual-Tower Air CPU Cooler | 168 mm dual tower rated for LGA1851 |
| Case | Fractal Design Meshify 2 Black Light TG ATX Case | Takes cards up to 470 mm and coolers up to 185 mm |
Why these parts#
CPU
The Core Ultra 7 265K (opens in a new tab) has 20 cores, made of 8 performance cores and 12 efficiency cores, and 20 threads. Intel lists a 5.5 GHz maximum turbo, a 125 W processor base power and a 250 W maximum turbo power. It uses socket LGA1851 and supports DDR5 up to 6400 MT/s with a 256 GB maximum.
For local language models the processor matters less than the card, but it still handles tokenising, scheduling and any layers that run on the CPU. Intel also lists 24 PCIe lanes in revisions 5.0 and 4.0, which covers a graphics card and fast storage. The chip includes an NPU rated at 13 peak TOPS, which is separate from the GPU and not used by most local LLM tools.
Intel's page gives a thermal solution specification but does not say a cooler is bundled, so the build includes one. For other options see our guide to the best CPU for AI and machine learning.
Motherboard
The ASUS ROG Strix Z890-A Gaming WiFi (opens in a new tab) uses Intel's Z890 chipset. ASUS lists support for Intel Core Ultra processors (Series 2), and the 265K belongs to that group. ASUS refers buyers to its CPU support list, so check it for the BIOS version on your board. ASUS also notes that Core Ultra 200S processors need Windows 10 21H2 or later, or Windows 11 22H2 or later.
It offers four DIMM slots and a 256 GB ceiling, with DDR5 listed by ASUS at up to 9066+ MT/s (OC), one PCIe 5.0 x16 slot from the CPU and a second PCIe 4.0 slot at x4 from the chipset. There are five M.2 slots, with M.2_1 at PCIe 5.0 x4 and the rest at PCIe 4.0 x4. ASUS lists Intel 2.5 Gb Ethernet, Wi-Fi 7 and two Thunderbolt 4 ports.
GPU
The INNO3D GeForce RTX 5090 X3 OC (opens in a new tab) carries 32 GB of GDDR7 on a 512-bit bus, which INNO3D lists at 1792 GB/s. NVIDIA lists 21,760 CUDA cores and a 575 W total graphics power for the chip. INNO3D lists the card dimensions as length 333, height 137 and width 60, without stating a unit, so check the maker's sheet against your case.
Capacity is the reason to buy this card. At 4-bit, a 32B model's weights take roughly 20 GB by that measurement, and a longer context adds more. Bandwidth matters as well. Puget Systems explains that bandwidth drives generation speed more than anything else, and this card has about twice the bandwidth of an RTX 5070 Ti.
INNO3D lists a 1000 W minimum system power and a power cable choice of one 600 W PCIe Gen 5 connector or four 8-pin connectors through an adapter. It also lists a card stand built into the cooler. NVIDIA lists NVLink as unsupported for this GPU, so a second card would be a separate device. See our best GPU for local LLM guide for other cards.
RAM
The TeamGroup T-Create Expert (opens in a new tab) kit is 96 GB as two 48 GB modules, with a 6800 MT/s CL36 profile at 1.4 V. TeamGroup lists Intel XMP 3.0 and AMD EXPO profiles and the module height at 32 mm, so it clears most coolers.
The speed is above Intel's listed DDR5 6400 MT/s for this processor, so it runs as an overclock profile. TeamGroup states that XMP and EXPO profiles count as overclocking and that final speed depends on the board, processor and BIOS. ASUS's DIMM Flex memory list names this board with a TeamGroup T-Create Expert 96 GB 6800 kit, though it gives no voltage or date. If the system will not boot stably, drop to a lower profile in the BIOS. Why 96 GB at all? When a model is larger than VRAM, llama.cpp places the remaining layers in system memory. More memory means more of the model stays loaded, and a larger context cache can live there too. Our guide to how much RAM for a local LLM explains the numbers.
Storage
The Samsung 990 PRO 4TB (opens in a new tab) uses a PCIe 4.0 x4 NVMe interface with up to 7,450 MB/s reads, up to 6,900 MB/s writes, a 4 GB cache and 2400 TBW of endurance. Quantised model files range from a few GB to tens of GB, so a collection grows quickly. A fast drive also cuts the time to load a model into VRAM.
Samsung sells the drive without a heatsink, so use the heatsink on the board's M.2 socket. ASUS lists M.2_1 at PCIe 5.0 x4, but this drive is PCIe 4.0, so any of the board's sockets suits it.
Power supply
INNO3D lists a 1000 W minimum for this card, and NVIDIA says its system power figure is based on a Ryzen 9 9950X system and recommends a PCIe CEM 5.1 compliant supply. The Corsair RM1000x (opens in a new tab) is ATX 3.1 and fully modular, with a ten-year warranty.
Corsair lists one 12V-2x6 connector, four PCIe 8-pin connectors and two EPS connectors, and a depth of 160 mm. That one 12V-2x6 cable is the one the card needs. The supply matches INNO3D's minimum rather than exceeding it, so it suits a single-card build. If you add other high-draw parts, check the combined load against the makers' figures.
CPU cooler
The Noctua NH-D15 G2 (opens in a new tab) is a dual-tower air cooler with eight heat pipes and two 140 mm NF-A14x25r G2 fans. Noctua lists AM5, LGA1851 and LGA1700 support, a height of 168 mm and a weight of 1525 g with fans. Noctua says the case needs about 170 mm of clearance, and the Meshify 2 allows 185 mm.
Noctua does not publish a TDP figure on the specification page, and does not list memory clearance in its headline specs. The kit chosen here is 32 mm tall, which suits a large dual tower. Intel lists the 265K's maximum turbo power at 250 W, which is why the build pairs it with a large dual-tower cooler.
Case
The Fractal Design Meshify 2 (opens in a new tab) lists GPU clearance of 470 mm, cooler clearance of 185 mm and power supply length up to 250 mm. The card's listed length is 333 (INNO3D gives no unit), the cooler is 168 mm and the supply 160 mm, so all three look comfortable against Fractal's limits. It supports E-ATX, ATX, micro-ATX and mini-ITX boards and has seven expansion slots.
Fractal ships it with two 140 mm fans at the front and one at the rear. It supports radiators up to 420 mm at the top and 360 mm at the front, which leaves a route to liquid cooling later.
Expected performance#
Benchmarks for the RTX 5090 are the most plentiful in this tier. Hardware Corner ran llama-bench at 16K context on gpt-oss 20B, a model that fits on every card in the chart. The 5090 led the 5070 Ti and the 5060 Ti by a wide margin. The source did not use this board partner's card, so the figures are for the chip.
Token generation, gpt-oss-20b, 4-bit: token generation (tokens/s, higher is better)
3 products. Highest: INNO3D GeForce RTX 5090 X3 OC 32GB, 249.19 tokens/s. Lowest: ASUS Dual GeForce RTX 5060 Ti 16GB GDDR7 OC, 82.42 tokens/s.
| Product | tokens/s |
|---|---|
| INNO3D GeForce RTX 5090 X3 OC 32GB | 249.19 |
| ASUS Prime GeForce RTX 5070 Ti 16GB GDDR7 OC | 133.05 |
| ASUS Dual GeForce RTX 5060 Ti 16GB GDDR7 OC | 82.42 |
About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.
The 5090 generated about 249 tokens per second, against about 133 for the 5070 Ti and about 82 for the 5060 Ti. A second source, the llama.cpp community scoreboard, shows a similar order on Llama 2 7B at Q4_0, reported by users rather than a lab.
llama.cpp llama-bench Llama 2 7B Q4_0, token generation: tg128 (tokens/s, higher is better)
3 products. Highest: INNO3D GeForce RTX 5090 X3 OC 32GB, 300.4 tokens/s. Lowest: ASUS Dual GeForce RTX 5060 Ti 16GB GDDR7 OC, 93.46 tokens/s.
| Product | tokens/s |
|---|---|
| INNO3D GeForce RTX 5090 X3 OC 32GB | 300.4 |
| ASUS Prime GeForce RTX 5070 Ti 16GB GDDR7 OC | 182.43 |
| ASUS Dual GeForce RTX 5060 Ti 16GB GDDR7 OC | 93.46 |
About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.
There the 5090 reached about 300 tokens per second, against about 182 for the 5070 Ti and about 93 for the 5060 Ti. Those figures are community-reported and board partners and power limits are not stated.
Bigger models show why this card exists. Computer Vision Lab measured the 5090 at about 61 to 63 tokens per second on 32B models in Ollama at 4-bit, against about 37 to 38 on the RTX 4090. On 7B and 8B models the 5090 was about 1.7 times faster. Hardware Corner's table also shows generation slowing as context grows, with the 5090 on Qwen3 8B at about 145 tokens per second at 16K context and about 49 at 131K.
One caution comes from Puget Systems. Its early prompt-processing results for the 50-series looked low, which it suspected was a Windows driver issue, because Linux runs did not show it. Treat any single prompt-processing figure with care.
Local LLM PC Build for India (last checked November 2026).
Upgrade path#
The processor and board have a clear path. The LGA1851 socket takes a faster Core Ultra chip, and the board has room for a 256 GB memory ceiling with four modules, though four modules usually run slower than two. Add a second NVMe drive before you add more memory.
For the card, there is not a higher consumer tier in this range, so the realistic upgrade is a second device. That means a second NVMe or a second GPU, though NVIDIA lists NVLink as unsupported and a second large card needs more than this power supply gives. Two 16 GB cards can reach a similar total at lower cost, but they add capacity rather than speed, so the single card here is the simpler route.
Alternatives by budget band#
For a lower budget with the same 16 GB models in mind, the AI PC build under 3 lakh is the entry route. The deep learning PC build is the balanced middle, and the Stable Diffusion PC build suits image work.
Assembly and setup notes#
These notes are drawn from the makers' documentation, and the manuals give the exact steps.
- Power cable: INNO3D lists one 600 W PCIe Gen 5 cable or four 8-pin connectors through an adapter. Use the native 12V-2x6 cable from the Corsair supply.
- Card support: INNO3D lists a built-in card stand. Follow its instructions for the case.
- BIOS: ASUS refers to its CPU support list for Core Ultra processors. Confirm the BIOS before you install the chip.
- Operating system: ASUS notes that Core Ultra 200S processors need Windows 10 21H2 or later, or Windows 11 22H2 or later.
- Cooler: Noctua lists a height of 168 mm and says the case needs about 170 mm of clearance. Install the memory first, since the cooler sits close to the DIMM slots.
- Memory profile: TeamGroup treats XMP as overclocking. Enable the profile in the BIOS and test the system before you rely on it.
Verdict#
This build exists to hold large models in fast memory. The RTX 5090 gives 32 GB of VRAM and the highest published throughput in the sources we checked. The supporting parts are matched to its size and power draw, and each has headroom under the makers' stated limits.
Choose it if you run 32B-class models often, or if long context is part of your work. Choose another build if a 16 GB card fits your models, since the platform around it then costs less. The best GPU for local LLM guide helps you decide.
A build like this draws steady power for hours, so Indian buyers on unstable mains should consider a good online UPS and check that the supply suits 230 V input before running long jobs.
Frequently asked questions#
How much VRAM do I need for a 32B model?
About 20 GB at 4-bit for the weights, according to ComputingForGeeks, plus extra for the context cache. A 32 GB card leaves that space, and a 16 GB card does not.
Is 96 GB of RAM necessary?
No. It is useful when a model is larger than VRAM and some layers run on the CPU. If every model you use fits in 32 GB of VRAM, 64 GB is likely enough.
Why Intel for this build?
The Core Ultra 7 265K and Z890 board give PCIe 5.0, DDR5 and plenty of M.2 slots. The AMD builds in this series use AM5 instead, and either works with the card.
Will the card fit my case?
INNO3D lists the card's length as 333 and its width as 60, without a unit. Check the maximum GPU length in your case's specification before you buy a different one.
Can I run this on a smaller power supply?
INNO3D lists a 1000 W minimum for the card. Use the makers' figures, and check them against your full parts list. A smaller supply risks shutdowns under load, so do not save money on this part of the build.
Does a second GPU double the speed?
No. NVIDIA lists NVLink as unsupported, and Hardware Corner found that two cards add capacity, not speed, on a model that fits on one. See the dual GPU build for a longer explanation.
Sources & credits#
We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.
- Intel, Core Ultra 7 processor 265K specifications (opens in a new tab)
- ASUS ROG India, ROG Strix Z890-A Gaming WiFi specifications (opens in a new tab)
- INNO3D, GeForce RTX 5090 X3 OC (opens in a new tab)
- NVIDIA India, GeForce RTX 5090 (opens in a new tab)
- TeamGroup, T-Create Expert DDR5 memory (opens in a new tab)
- Samsung India, 990 PRO 4TB (opens in a new tab)
- Corsair, RM1000x power supply (opens in a new tab)
- Noctua, NH-D15 G2 specifications (opens in a new tab)
- Fractal Design, Meshify 2 (opens in a new tab)
- Hardware Corner, The Definitive GPU Ranking for LLMs (opens in a new tab)
- llama.cpp GitHub Discussion 15013, Performance of llama.cpp on Nvidia CUDA (community-reported) (opens in a new tab)
- Computer Vision Lab, Benchmarking Nvidia RTX 5090 (opens in a new tab)
- Puget Systems, NVIDIA GeForce RTX 5090 & 5080 AI Review (opens in a new tab)
- ComputingForGeeks, RTX 5060 Ti vs 5070 Ti: Local AI Benchmarks (opens in a new tab)
Facts and figures in this article come from these sources. Articles are compiled from the makers' published specifications and the benchmarks named above, and checked by the NextAIGrid Editorial Team.
Disclaimer
This page is for information only. Specifications and availability change, so check the maker's page before you buy. Found an error? Tell us.
How was this pc build put together?
It draws on 10 named sources: maker specification pages and published third-party benchmarks, listed in the article. We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Specifications and sources were last checked on 17 November 2026. Warranty terms and availability differ by brand and seller in India, so check the maker's India page for the warranty period and buy from an authorised seller. Read how we research and our editorial policy.
Spotted an error? Tell us. Confirmed fixes are listed on the corrections page.