1. GPU Khong Phai CPU Nhieu Nhan: Kien Truc Song Song Chuyen Biet
CPU toi uu hoa do tre latency thap cho cac tac vu tuan tu phuc tap (it luong, ma nhan phuc tap). GPU toi uu hoa throughput toc do cao cho hang ngan tac vu don gian song song (nhieu luong, ma nhan don gian). RTX 4090 co 16.384 CUDA Core — tu moi Core la mot don vi tinh toan FP32 don gian, chay song song de tinh toan gia tri mau sac cua hang tram trieu pixel moi giay.
2. Ada Lovelace (RTX 40 Series): SM, CUDA Core, RT Core Gen 3, Tensor Core Gen 4
Ada Lovelace duoc san xuat tren TSMC 4nm (4N process). Don vi co ban la Streaming Multiprocessor (SM) — moi SM chua 128 CUDA Core FP32, 2 RT Core (Gen 3), 4 Tensor Core (Gen 4) va L1 Cache/Shared Memory 128KB. RTX 4090 co 128 SM = 16.384 CUDA Core.
| Don Vi | RTX 4090 (AD102) | RTX 4080 Super (AD103) | RTX 4070 Ti Super (AD103) | RX 7900 XTX (RDNA 3) |
|---|---|---|---|---|
| Shader Cores | 16.384 CUDA | 10.240 CUDA | 8.448 CUDA | 6.144 SP (12.288 eff) |
| RT Cores | 128 (Gen 3) | 80 (Gen 3) | 66 (Gen 3) | 96 RA (Gen 2) |
| Tensor/AI Cores | 512 (Gen 4) | 320 (Gen 4) | 264 (Gen 4) | N/A (AI Acc) |
| VRAM | 24GB GDDR6X | 16GB GDDR6X | 16GB GDDR6X | 24GB GDDR6 |
| Bus VRAM | 384-bit | 256-bit | 256-bit | 384-bit |
| TDP | 450W | 320W | 285W | 355W |
| Gia VN (8/2026) | ~45 trieu | ~25 trieu | ~18 trieu | ~20 trieu |
3. RT Core Gen 3: Kien Truc BVH Traversal Chuyen Biet
Ray Tracing tren GPU truoc Ada Lovelace phai dung CUDA Core de tinh toan — cham. RT Core Gen 3 la don vi phan cung chuyen biet cho BVH (Bounding Volume Hierarchy) traversal va triangle intersection — tinh toan xem tia sang co cham vao hinh tam giac hay khong. Ada Lovelace RT Core Gen 3 nhanh gap doi so voi Ampere Gen 2, cho phep DLSS 3 Frame Generation.
4. Tensor Core Gen 4: AI Sieu Mau (DLSS) va Compute
Tensor Core la don vi phan cung cho phep tinh toan ma tran (Matrix Multiply–Accumulate — MMA) toc do cao — can cho AI inference. Tren RTX 4090, Tensor Core Gen 4 dat 1.321 TFLOPS FP8 — la nen tang cua DLSS 3, RTX Video Super Resolution va cac ung dung AI nhu Stable Diffusion, ChatGPT local inference.
5. RDNA 3 (RX 7000 Series): Compute Unit, Dual Issue va Infinity Cache
AMD RDNA 3 dung kien truc chiplet: GPU Die (GCD) san xuat tren TSMC 5nm, Memory Cache Die (MCD) tren 6nm. Moi Compute Unit (CU) cua RDNA 3 co 64 Stream Processors nhung ho tro dual-issue (2 FP32 lenh/cycle) → hieu qua nhu 128 SP the he truoc. Infinity Cache (IC) la cache L3 lon (96MB tren RX 7900 XTX) tren MCD de bu bang thong bo nho.
6. Kien Truc Thong Nhat vs ROCm vs CUDA: He Sinh Thai Phan Mem
CUDA vs ROCm: He Sinh Thai Quyet Dinh Chon GPU
NVIDIA CUDA co he sinh thai phan mem AI/ML vuot troi — PyTorch, TensorFlow, Stable Diffusion, ComfyUI deu uu tien CUDA. AMD ROCm da cai thien nhung van thieu mot so thu vien. Neu lam viec voi AI/ML hoac Content Creation co CUDA acceleration, NVIDIA la lua chon duy nhat hien tai (2026).
? CUDA Core la gi va nhieu CUDA Core co nghia la GPU manh hon khong?
Khong nhat thiet. RTX 4070 co 5.888 CUDA Core nhung nhanh hon RX 7800 XT co 3.840 SP (trong game). Kien truc, xung nhip, bang thong bo nho va thuat toan moi quyet dinh hieu nang thuc te.
? GPU co can Hyper-Threading nhu CPU khong?
Khong — GPU da song song thuan tuy. Moi CUDA Core la mot don vi tinh toan doc lap. GPU xu ly cac lenh theo warp (nhom 32 thread chay cung instruction). Neu co thread phan khi (divergent), hieu nang giam.
? Tai sao RX 7900 XTX re hon RTX 4080 nhung VRAM nhieu hon?
AMD chon GDDR6 (re hon GDDR6X) va bus 384-bit rong hon, tang VRAM len 24GB. NVIDIA chon GDDR6X toc do cao hon, bus hep hon. VRAM nhieu giup cho workload >16GB nhung GDDR6X cho bang thong cao hon trong thuc te gaming.
? GPU tich hop (iGPU) va GPU roi khac nhau nhu the nao?
iGPU (trong CPU: Intel Iris Xe, AMD Radeon 890M) chia se RAM he thong, khong co VRAM rieng, kem hon nhieu. GPU roi co VRAM rieng (GDDR6/GDDR6X), CUDA/CU doc lap, manh hon 5-20 lan tuy model.
? Khi nao nen chon GPU AMD thay NVIDIA?
GPU AMD la lua chon tot khi: Ngan sach han hep (RX 7700 vs RTX 4060), gaming thuan tuy khong can CUDA, man hinh FreeSync (khong can G-Sync). Chon NVIDIA khi: Lam viec voi AI/ML, CUDA workload, co man hinh G-Sync.
? Sau GPU doi moi cua NVIDIA va AMD khi nao?
NVIDIA: Blackwell Architecture (RTX 50 series) da ra mat Q1/2025, RTX 5090/5080 co san tai VN 2025. AMD: RDNA 4 (RX 9000 series) ra mat 2025. Neu can mua 2026, PCIe 5.0 bus tren RTX 50 series la loi the duong dai.
Can Tu Van Chon GPU / VGA Phu Hop?
Doi ngu ky su VietITPro tu van chon GPU NVIDIA / AMD phu hop nhu cau — Gaming 1080p/1440p/4K, AI Render, CUDA Workstation. Bao hanh chinh hang, lap dat tai nha TP.HCM.
