1. Introduction
The GPU market in 2026 is broader and more fragmented than at any point in its history. NVIDIA's Blackwell architecture dominates the high end, AMD has restructured its entire GPU lineup with RDNA 4 and CDNA 4, Intel has finally entered the discrete desktop GPU market with Battlemage, and workstation-class accelerators now compete directly with datacenter hardware in the same physical enclosure.
This article compares every major GPU across four tiers: consumer gaming, workstation, datacenter accelerators, and the new category that blurs the line between them. The focus is on hard numbers: VRAM, memory bandwidth, performance per dollar, and GB-per-dollar ratios. Where benchmarks exist, they are cited. Where they do not, the article states the limitation clearly.
Primary sources used: NVIDIA product pages, AMD product pages, TechPowerUp GPU database, Tom's Hardware, ServeTheHome, Klymentiev.com, IntuitionLabs, GPU.fm, and WCCFTech. All pricing is USD MSRP at launch unless noted. All benchmark data is sourced from the cited articles. Vendor claims are explicitly labeled as such. Independent benchmarks are separately identified.
Qwen3.8-27B: Architectural Study of a Dense Multimodal LLM examined how AI models are driving GPU demand. The hardware side of that equation starts here.
2. Consumer Gaming GPUs: NVIDIA GeForce RTX 50 Series
2.1. RTX 5090 (Founders Edition)
| Spec | Value |
|---|---|
| Architecture | Blackwell (GB202) |
| CUDA Cores | 21,760 |
| VRAM | 32 GB GDDR7 |
| Memory Bus | 512-bit |
| Memory Bandwidth | 1,792 GB/s |
| TDP | 575W |
| Launch Price | $1,999 |
| GB per Dollar | 16.0 |
| GB/s per Dollar | 0.896 |
The RTX 5090 is the undisputed consumer performance king. Its 32 GB of GDDR7 on a 512-bit bus delivers 1,792 GB/s of bandwidth, nearly double the RTX 4090's 1,008 GB/s. At 575W TDP, it is also the most power-hungry consumer GPU ever.
Strengths: Maximum consumer VRAM and bandwidth, best 4K gaming, strongest ray-tracing performance. Weaknesses: $1,999 price, 575W TDP requires robust PSU and cooling, no significant advantage over 5080 at 1440p.
2.2. RTX 5080 (Founders Edition)
| Spec | Value |
|---|---|
| Architecture | Blackwell (GB203) |
| VRAM | 16 GB GDDR7 |
| Memory Bus | 256-bit |
| Memory Bandwidth | 960 GB/s |
| TDP | ~360W (est.) |
| Launch Price | $999 |
| GB per Dollar | 16.0 |
| GB/s per Dollar | 0.96 |
The RTX 5080 lands at the same GB-per-dollar ratio as the 5090 (16.0), but at half the VRAM capacity. The 256-bit bus limits bandwidth to 960 GB/s, which is adequate for 4K but may become a bottleneck for AI workloads that need to load large models.
Strengths: Strong 4K gaming at half the 5090 price, good GB/s per dollar. Weaknesses: 16 GB VRAM limits AI/model workloads compared to 32 GB on the 5090.
2.3. RTX 5070 Ti
| Spec | Value |
|---|---|
| Architecture | Blackwell |
| VRAM | 12 GB GDDR7 |
| Memory Bus | 192-bit |
| Memory Bandwidth | ~672 GB/s |
| Launch Price | $749 |
| GB per Dollar | 16.0 |
| GB/s per Dollar | 0.90 |
2.4. RTX 5070
| Spec | Value |
|---|---|
| Architecture | Blackwell |
| VRAM | 12 GB GDDR7 |
| Memory Bandwidth | 672 GB/s |
| Launch Price | $549 |
| GB per Dollar | 21.9 |
| GB/s per Dollar | 1.22 |
The RTX 5070 at $549 offers the best GB-per-dollar ratio in the RTX 50 series at 21.9, though the 12 GB VRAM cap limits its usefulness for AI workloads.
3. Consumer Gaming GPUs: AMD Radeon RX 9000 Series (RDNA 4)
3.1. RX 9070 XT
| Spec | Value |
|---|---|
| Architecture | RDNA 4 |
| Compute Units | 512 (4,096 unified) |
| VRAM | 16 GB GDDR6 |
| Memory Bus | 256-bit |
| Memory Speed | 20 Gbps |
| Infinity Cache | 64 MB |
| Memory Bandwidth | ~640 GB/s (est.) |
| Launch Price | ~$549 |
| GB per Dollar | 29.1 |
| GB/s per Dollar | ~1.17 |
AMD's RX 9070 XT represents a significant shift: 16 GB of VRAM at $549 gives it the highest GB-per-dollar ratio in the high-end consumer segment at 29.1. It uses GDDR6 rather than GDDR7, which limits bandwidth compared to the RTX 5070 Ti's 672 GB/s, but the larger VRAM pool and lower price make it competitive.
Strengths: Excellent GB-per-dollar (29.1), 16 GB VRAM at $549, strong 1440p performance. Weaknesses: Lower bandwidth than GDDR7 competitors, FSR 4.1 still trails DLSS 4 in quality, ray-tracing performance lags NVIDIA.
3.2. RX 9070 (non-XT)
| Spec | Value |
|---|---|
| VRAM | 16 GB GDDR6 |
| Memory Bandwidth | ~512 GB/s (est.) |
| Launch Price | ~$449 |
| GB per Dollar | 35.6 |
| GB/s per Dollar | ~1.14 |
The non-XT variant offers even better value at 35.6 GB per dollar, targeting 1080p/1440p gaming.
4. Consumer Gaming GPUs: Intel Arc Battlemage (B-Series)
4.1. Intel Arc B580
| Spec | Value |
|---|---|
| Architecture | Battlemage (BGI) |
| Xe-cores | 22 |
| VRAM | 12 GB GDDR6 |
| Memory Bus | 192-bit |
| Launch Price | $249 |
| GB per Dollar | 48.2 |
| GB/s per Dollar | ~1.64 |
The Arc B580 at $249 is the highest GB-per-dollar consumer GPU at 48.2. It targets the budget-midrange segment and competes with the RTX 4060 in raw gaming performance while offering more VRAM. Intel's XeSS upscaling has improved significantly since the Arc Alchemist launch.
4.2. Intel Arc B570
| Spec | Value |
|---|---|
| VRAM | 10 GB GDDR6 |
| Launch Price | $219 |
| GB per Dollar | 45.7 |
4.3. Intel Arc B380
| Spec | Value |
|---|---|
| Architecture | Battlemage |
| Launch Price | ~$150 (est.) |
| VRAM | 8 GB GDDR6 |
Intel's Battlemage lineup is still maturing. The B770 and B750 were reportedly cancelled due to VRAM cost pressures, leaving Intel with a gap in the $400-$500 range that AMD and NVIDIA fill.
5. Workstation and Hybrid AI GPUs
5.1. AMD Instinct MI350P (CDNA 4)
The MI350P is the bridge between consumer and datacenter: a PCIe-form-factor accelerator that brings datacenter-grade HBM to a standard PCIe slot.
| Spec | MI350P (Single GPU) | MI350P (Dual GPU) |
|---|---|---|
| Architecture | CDNA 4 | CDNA 4 |
| VRAM | 144 GB HBM3E | 288 GB HBM3E |
| Memory Bandwidth | 4 TB/s | 8 TB/s |
| Matrix Perf (MXFP8) | 2.3 PFLOPS | 4.6 PFLOPS |
| Price | ~$30,000 | ~$60,000 |
| GB per Dollar | 4.8 | 4.8 |
| TB/s per Dollar | 0.133 | 0.133 |
The MI350P's 144 GB HBM3E at $30K gives it a GB-per-dollar of 4.8, far lower than consumer cards, but this is expected. HBM3E is dramatically more expensive per GB than GDDR7. However, its 4 TB/s bandwidth is 2-3x that of the RTX 5090, and the MXFP8 compute performance of 2.3 PFLOPS makes it competitive with NVIDIA's H200 for specific workloads.
AMD claims the MI350P is 40% faster than the H200 NVL in AI workloads at a similar price point. This claim has not been independently verified at publication time.
5.2. AMD MI355X / MI350X (Datacenter OAM)
The MI350X series uses 8 GPU modules with 2.3 TB aggregate HBM3E memory and 64 TB/s peak aggregate bandwidth. The MI355X is the higher-end variant. These are rack-scale accelerators, not PCIe cards.
6. Datacenter GPUs: NVIDIA Blackwell
6.1. NVIDIA B200
| Spec | Value |
|---|---|
| Architecture | Blackwell |
| VRAM | 192 GB HBM3e |
| Memory Bandwidth | 8 TB/s |
| Transistors | 208 billion |
| Cloud Price | $4-6 / GPU-hour |
| Unit Price | ~$30,000-40,000 (est.) |
| GB per Dollar | 4.8-6.4 |
The B200 is NVIDIA's flagship datacenter GPU. Its 192 GB of HBM3e at 8 TB/s bandwidth is the highest of any single GPU. At $30K-40K per unit, the GB-per-dollar ratio is 4.8-6.4, comparable to the MI350P, which makes sense as they compete in the same segment.
6.2. NVIDIA H200
| Spec | Value |
|---|---|
| VRAM | 141 GB HBM3e |
| Memory Bandwidth | 4.8 TB/s |
| Unit Price | ~$25,000-35,000 |
| GB per Dollar | 4.0-5.6 |
The H200 succeeded the H100 and is widely deployed. Its 141 GB HBM3e and 4.8 TB/s bandwidth make it the workhorse GPU for inference and training workloads.
6.3. NVIDIA H100
| Spec | Value |
|---|---|
| VRAM | 80 GB HBM3 |
| Memory Bandwidth | 3.35 TB/s |
| Unit Price | ~$25,000-40,000 (secondary market) |
| GB per Dollar | 2.0-3.2 |
6.4. NVIDIA GB200 NVL72
The GB200 NVL72 is a full rack solution with 72 B200 GPUs interconnected. Total VRAM approaches 13.8 TB HBM3e. Pricing is on the order of $30M-40M per rack.
7. Datacenter GPUs: AMD Instinct
7.1. MI300X
| Spec | Value |
|---|---|
| VRAM | 192 GB HBM3 |
| Memory Bandwidth | 5.3 TB/s |
| Unit Price | ~$15,000-20,000 |
| GB per Dollar | 9.6-12.8 |
The MI300X is AMD's best value datacenter GPU at ~$15-20K, offering 192 GB HBM3 at nearly twice the GB-per-dollar of the H100. It is widely considered the best price-to-performance datacenter GPU.
7.2. MI325X / MI355X
| Spec | Value |
|---|---|
| VRAM | 288 GB HBM3e (MI325X) |
| Unit Price | ~$30,000-40,000 |
| GB per Dollar | 7.2-9.6 |
The MI325X upgrades to 288 GB HBM3e, and the MI355X is the next-generation variant. Both compete directly with the H200 and B200.
8. GB-per-Dollar Analysis
This is where the landscape gets interesting. GB-per-dollar is a rough proxy for AI model capacity at a given budget, but bandwidth and compute matter for throughput.
8.1. Consumer Segment (GB per dollar)
| GPU | Price | VRAM | GB/$ | Rank |
|---|---|---|---|---|
| Intel Arc B580 | $249 | 12 GB | 48.2 | 1 |
| Intel Arc B570 | $219 | 10 GB | 45.7 | 2 |
| AMD RX 9070 | ~$449 | 16 GB | 35.6 | 3 |
| AMD RX 9070 XT | ~$549 | 16 GB | 29.1 | 4 |
| NVIDIA RTX 5070 | $549 | 12 GB | 21.9 | 5 |
| NVIDIA RTX 5090 | $1,999 | 32 GB | 16.0 | 6 |
| NVIDIA RTX 5080 | $999 | 16 GB | 16.0 | 7 |
| NVIDIA RTX 5070 Ti | $749 | 12 GB | 16.0 | 8 |
In the consumer segment, Intel's Battlemage dominates GB-per-dollar at the budget end, AMD's RDNA 4 leads in the high-end, and NVIDIA's Blackwell is at the bottom of the ranking despite offering the best raw performance.
8.2. Datacenter Segment (GB per dollar)
| GPU | Price | VRAM | GB/$ | Rank |
|---|---|---|---|---|
| AMD MI300X | ~$15K | 192 GB | 12.8 | 1 |
| AMD MI325X | ~$30K | 288 GB | 9.6 | 2 |
| NVIDIA B200 | ~$30K | 192 GB | 6.4 | 3 |
| AMD MI350P | ~$30K | 144 GB | 4.8 | 4 |
| NVIDIA H200 | ~$25K | 141 GB | 5.6 | 5 |
| NVIDIA H100 | ~$25K | 80 GB | 3.2 | 6 |
The MI300X is the datacenter value champion by a wide margin. Its 192 GB at $15-20K is nearly double the GB-per-dollar of NVIDIA's comparable offerings.
8.3. Bandwidth per Dollar
| GPU | Bandwidth | Price | GB/s per $ |
|---|---|---|---|
| Intel Arc B580 | ~500 GB/s | $249 | 2.01 |
| AMD RX 9070 XT | ~640 GB/s | $549 | 1.17 |
| NVIDIA RTX 5070 | 672 GB/s | $549 | 1.22 |
| NVIDIA RTX 5080 | 960 GB/s | $999 | 0.96 |
| NVIDIA RTX 5090 | 1,792 GB/s | $1,999 | 0.90 |
| AMD MI350P | 4 TB/s | $30K | 0.133 |
| NVIDIA B200 | 8 TB/s | $30K | 0.267 |
| AMD MI300X | 5.3 TB/s | $15K | 0.353 |
The MI300X is the bandwidth-per-dollar king of the datacenter segment. At $15K for 5.3 TB/s, it outperforms the B200 on this metric despite the B200 having higher raw bandwidth.
9. The Threadripper Halo Station
Unveiled at IFA 2026 on September 4, 2026, the Threadripper Halo Station is AMD's answer to NVIDIA's DGX Station. It is a liquid-cooled deskside workstation that packs datacenter performance into a form factor that fits in a room.
9.1. Specifications
| Component | Specification |
|---|---|
| CPU | Ryzen Threadripper PRO 9995WX (Zen 5) |
| CPU Cores | 96 cores / 192 threads |
| CPU Boost | 5.4 GHz |
| CPU Price | $11,700 |
| System Memory | Up to 2 TB DDR5 |
| GPU/Accelerator | Up to 4x AMD Instinct MI350P |
| GPU VRAM | Up to 576 GB HBM3E (4x144 GB) |
| GPU Bandwidth | Up to 16 TB/s aggregate (4x4 TB/s) |
| Cooling | Liquid-cooled |
| Street Price | $100,000+ (core components only) |
9.2. What This Means
At $100K+ for just the core components (CPU, memory, accelerators), the Halo Station is not a consumer product. It is a competitor to NVIDIA's DGX Station, which supports AI models up to 1 trillion parameters using NVIDIA's H100/H200 GPUs.
The Halo Station's 576 GB of HBM3E across 4 MI350P accelerators is the key number. At 576 GB, it can run multi-billion parameter models entirely in GPU memory without CPU offloading. For comparison:
| System | GPU Memory | VRAM Type | Est. Max LLM Params |
|---|---|---|---|
| Halo Station (4x MI350P) | 576 GB | HBM3E | ~288B (Q4) / ~115B (Q8) |
| Halo Station (2x MI350P) | 288 GB | HBM3E | ~144B (Q4) / ~57B (Q8) |
| DGX Station (8x H100) | 640 GB | HBM3 | ~320B (Q4) / ~160B (Q8) |
| RTX 5090 (single) | 32 GB | GDDR7 | ~16B (Q4) / ~8B (Q8) |
The 96-core Threadripper PRO 9995WX at $11,700 handles system memory (up to 2 TB DDR5) and orchestration. Combined with the MI350P accelerators, this system is designed for trillion-parameter model inference and fine-tuning at the edge.
Key takeaway: The Halo Station represents AMD's strategy of using CDNA 4 (MI350P) to compete with NVIDIA's Blackwell H200/B200 in the workstation segment, backed by a 96-core Zen 5 CPU. It is not a consumer product, but it signals that the line between workstation and datacenter is disappearing.
10. Adjacent Technologies
10.1. HBM vs GDDR: The Memory Divide
HBM3e costs approximately 5-10x more per GB than GDDR7, but delivers 2-4x the bandwidth. This is why consumer GPUs use GDDR7 (cost-effective at the 12-32 GB range) while datacenter GPUs use HBM3e (bandwidth is the bottleneck for large models).
| Memory Type | Cost per GB (est.) | Bandwidth | Use Case |
|---|---|---|---|
| GDDR6 | ~$5-8 | 400-640 GB/s | Budget consumer |
| GDDR7 | ~$10-15 | 672-1,792 GB/s | High-end consumer |
| HBM3 | ~$50-80 | 3.3-5.3 TB/s | Datacenter |
| HBM3e | ~$60-100 | 4-8 TB/s | Datacenter / Workstation |
10.2. AI-Specific Tensor/Matrix Cores
All modern GPUs include dedicated AI cores:
- NVIDIA: Tensor Cores (FP8, FP4, INT8)
- AMD: Matrix Cores (FP8, INT8, INT4)
- Intel: Xe-cores with DP4a
For local AI workloads, FP8 and INT4 precision support is critical for running larger models at acceptable speeds. The RTX 50 series and MI350P both support FP8 and INT4.
10.3. Ray Tracing
NVIDIA's Blackwell RT cores deliver approximately 2x the ray-tracing performance of previous-gen Ada Lovelace. AMD's RDNA 4 RT cores have narrowed the gap but still trail NVIDIA by roughly 30-40% in raw RT performance. This gap is less relevant for AI workloads but matters for gaming and rendering.
11. The Local AI Perspective
For someone running local LLMs, the GPU landscape maps to clear tiers:
11.1. Entry Level (Under $500)
- Intel Arc B580 ($249): Best GB-per-dollar at 48.2. Can run 7B-13B models at Q4. Intel drivers have improved significantly.
- AMD RX 9070 (~$449): 16 GB VRAM at $449 gives 35.6 GB/$. Can run 13B-20B models at Q4. Better driver maturity than Intel.
11.2. Mid Range ($500-$1,000)
- NVIDIA RTX 5070 ($549): 12 GB VRAM, best CUDA ecosystem support. Runs 7B-13B models smoothly. Limited by 12 GB VRAM.
- AMD RX 9070 XT (~$549): 16 GB VRAM at 29.1 GB/$. Can run 20B-30B models at Q4. ROCm support is improving but still lags CUDA.
11.3. High End ($1,000-$2,000)
- NVIDIA RTX 5080 ($999): 16 GB VRAM, strong performance. Runs 30B models at Q4.
- NVIDIA RTX 5090 ($1,999): 32 GB VRAM, 1,792 GB/s bandwidth. Runs 70B models at Q2-Q4 (with quantization tricks). The only single-GPU option for serious local AI.
11.4. Workstation ($30,000+)
- AMD MI350P (~$30K): 144 GB HBM3E. Can run 100B+ models entirely in GPU memory. For enterprises and research labs.
- Threadripper Halo Station ($100K+): Up to 576 GB HBM3E across 4 accelerators. Can run trillion-parameter models. The closest thing to a consumer datacenter.
12. Market Share and Competition
NVIDIA holds approximately 94% of the discrete GPU market (including both consumer and datacenter) as of early 2026, with AMD at roughly 5%. Intel's Battlemage is a new entrant in the budget consumer segment only, it has no datacenter presence.
AMD's market share in the datacenter AI segment is growing. The MI300X at ~$15-20K is widely considered the best value datacenter GPU, and cloud providers are increasingly offering MI300X instances as a cheaper alternative to H100/H200 instances. NVIDIA's 94% share is skewed by its consumer gaming dominance; in the AI training segment, AMD's share is estimated at 10-15%.
13. Conclusion
The 2026 GPU landscape is defined by three trends:
- GDDR7 arrives in consumer GPUs, doubling bandwidth over GDDR6 but at a cost premium. The RTX 50 series and RX 9000 XT both use it.
- HBM3e bridges the workstation-datacenter gap. The MI350P and Threadripper Halo Station bring datacenter memory to a desk-form factor, making large-model inference accessible outside of cloud datacenters.
- GB-per-dollar is the dominant value metric for local AI. Intel's B580 (48.2), AMD's RX 9070 XT (29.1), and the MI300X (12.8) are the value champions at their respective tiers. NVIDIA's Blackwell offers the best raw performance but the lowest GB-per-dollar in the consumer segment.
For local AI users: the RTX 5090 at $1,999 remains the best single-GPU option for serious model work (32 GB GDDR7), but the MI350P at $30K is the gateway to models that simply cannot fit on consumer hardware. The Threadripper Halo Station is the future of edge AI, a workstation that can run trillion-parameter models entirely offline.
References
- NVIDIA GeForce RTX 50 Series, Official product page, specs and pricing
- AMD Radeon RX 9070 XT, Official AMD product page
- Intel Arc B580, TechPowerUp GPU database
- RTX 5090 vs 5080 vs 5070 Ti, Specification comparison
- AMD Instinct MI350P, ServeTheHome technical coverage
- AMD MI350P vs H200, Performance comparison and pricing
- Threadripper PRO 9995WX, TechPowerUp CPU database, $11,700 MSRP
- NVIDIA B200, Specs, pricing, and benchmarks
- Data Center GPU Pricing 2026, IntuitionLabs pricing guide
- AMD Threadripper Halo Station, Shattered.io coverage, IFA 2026
- AMD Threadripper Halo Station at IFA 2026, TechPowerUp coverage
- AMD $100K Desk Machine, ValueAddVC analysis
- GPU Price/Performance Comparison, Silicon Analysts value analysis
- Best GPU for AI in 2026, Klymentiev.com datacenter comparison
- NVIDIA Data Center GPU Specs, IntuitionLabs datacenter specs
- AMD vs NVIDIA GPU Market Share 2026, Silicon Analysts market share[truncated]