{"id":4147,"date":"2025-09-22T11:12:00","date_gmt":"2025-09-22T11:12:00","guid":{"rendered":"https:\/\/lp.szlogic.cn\/glossary\/what-is-a-gpu-graphics-processing-units\/"},"modified":"2026-06-22T06:48:30","modified_gmt":"2026-06-22T06:48:30","slug":"what-is-a-gpu-graphics-processing-units","status":"publish","type":"post","link":"https:\/\/resourceslp.szlogic.cn\/vi\/glossary\/what-is-a-gpu-graphics-processing-units","title":{"rendered":"Hi\u1ec3u V\u1ec1 GPU: C\u00e1c Th\u00f4ng S\u1ed1 Ch\u00ednh V\u00e0 \u1ee8ng D\u1ee5ng"},"content":{"rendered":"
\"Understanding<\/figure>\n\n\n\n

1\ufe0f\u20e3 GPU l\u00e0 g\u00ec?<\/h2>\n\n\n\n

\u0110\u01a1n v\u1ecb x\u1eed l\u00fd \u0111\u1ed3 h\u1ecda (GPU) l\u00e0 m\u1ed9t m\u1ea1ch \u0111i\u1ec7n t\u1eed chuy\u00ean d\u1ee5ng \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1ec3 t\u0103ng t\u1ed1c qu\u00e1 tr\u00ecnh hi\u1ec3n th\u1ecb h\u00ecnh \u1ea3nh, video v\u00e0 ho\u1ea1t \u1ea3nh. GPU l\u00e0 th\u00e0nh ph\u1ea7n kh\u00f4ng th\u1ec3 thi\u1ebfu trong \u0111i\u1ec7n to\u00e1n hi\u1ec7n \u0111\u1ea1i, n\u00e2ng cao hi\u1ec7u n\u0103ng trong tr\u00f2 ch\u01a1i, tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o (AI) v\u00e0 \u0111i\u1ec7n to\u00e1n hi\u1ec7u n\u0103ng cao (HPC<\/a>) c\u00e1c t\u00e1c v\u1ee5. Kh\u00e1c v\u1edbi \u0110\u01a1n v\u1ecb x\u1eed l\u00fd trung t\u00e2m (CPU<\/a>), v\u1ed1n \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u h\u00f3a cho x\u1eed l\u00fd tu\u1ea7n t\u1ef1, GPU v\u01b0\u1ee3t tr\u1ed9i trong x\u1eed l\u00fd song song, c\u00f3 kh\u1ea3 n\u0103ng th\u1ef1c hi\u1ec7n nhi\u1ec1u t\u00e1c v\u1ee5 \u0111\u1ed3ng th\u1eddi.<\/p>\n\n\n\n

2\ufe0f\u20e3 C\u00e1c th\u00f4ng s\u1ed1 k\u1ef9 thu\u1eadt ch\u00ednh c\u1ee7a GPU<\/h2>\n\n\n\n

S\u1ed1 l\u00f5i v\u00e0 ki\u1ebfn tr\u00fac<\/h3>\n\n\n\n

C\u00e1c GPU hi\u1ec7n \u0111\u1ea1i s\u1edf h\u1eefu h\u00e0ng ngh\u00ecn l\u00f5i nh\u1ecf h\u01a1n, ch\u1eb3ng h\u1ea1n nh\u01b0 c\u00e1c l\u00f5i CUDA tr\u00ean GPU NVIDIA ho\u1eb7c b\u1ed9 x\u1eed l\u00fd lu\u1ed3ng tr\u00ean GPU AMD. Nh\u1eefng l\u00f5i n\u00e0y cho ph\u00e9p GPU th\u1ef1c hi\u1ec7n nhi\u1ec1u ph\u00e9p t\u00ednh \u0111\u1ed3ng th\u1eddi, khi\u1ebfn ch\u00fang tr\u1edf n\u00ean l\u00fd t\u01b0\u1edfng cho c\u00e1c t\u00e1c v\u1ee5 nh\u01b0 hi\u1ec3n th\u1ecb \u0111\u1ed3 h\u1ecda v\u00e0 h\u1ecdc m\u00e1y.<\/p>\n\n\n\n

B\u1ed9 nh\u1edb (VRAM)<\/h3>\n\n\n\n

B\u1ed9 nh\u1edb truy c\u1eadp ng\u1eabu nhi\u00ean d\u00e0nh ri\u00eang cho video (VRAM) l\u00e0 b\u1ed9 nh\u1edb chuy\u00ean d\u1ee5ng m\u00e0 GPU s\u1eed d\u1ee5ng \u0111\u1ec3 l\u01b0u tr\u1eef k\u1ebft c\u1ea5u, b\u1ed9 \u0111\u1ec7m khung h\u00ecnh v\u00e0 c\u00e1c d\u1eef li\u1ec7u \u0111\u1ed3 h\u1ecda kh\u00e1c. Dung l\u01b0\u1ee3ng VRAM cao h\u01a1n cho ph\u00e9p hi\u1ec7u n\u0103ng t\u1ed1t h\u01a1n trong tr\u00f2 ch\u01a1i \u0111\u1ed9 ph\u00e2n gi\u1ea3i cao v\u00e0 c\u00e1c \u1ee9ng d\u1ee5ng chuy\u00ean nghi\u1ec7p.<\/p>\n\n\n\n

T\u1ed1c \u0111\u1ed9 xung nh\u1ecbp v\u00e0 b\u0103ng th\u00f4ng<\/h3>\n\n\n\n

T\u1ed1c \u0111\u1ed9 xung nh\u1ecbp c\u1ee7a GPU x\u00e1c \u0111\u1ecbnh t\u1ed1c \u0111\u1ed9 x\u1eed l\u00fd l\u1ec7nh c\u1ee7a n\u00f3, trong khi b\u0103ng th\u00f4ng b\u1ed9 nh\u1edb ch\u1ec9 m\u1ee9c \u0111\u1ed9 nhanh m\u00e0 d\u1eef li\u1ec7u c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c \u0111\u1ecdc t\u1eeb ho\u1eb7c ghi v\u00e0o VRAM. T\u1ed1c \u0111\u1ed9 xung nh\u1ecbp v\u00e0 b\u0103ng th\u00f4ng cao h\u01a1n g\u00f3p ph\u1ea7n c\u1ea3i thi\u1ec7n hi\u1ec7u n\u0103ng t\u1ed5ng th\u1ec3.<\/p>\n\n\n\n

C\u00f4ng su\u1ea5t thi\u1ebft k\u1ebf nhi\u1ec7t (TDP)<\/h3>\n\n\n\n

TDP \u0111\u1ec1 c\u1eadp \u0111\u1ebfn l\u01b0\u1ee3ng nhi\u1ec7t m\u00e0 GPU sinh ra d\u01b0\u1edbi t\u1ea3i t\u1ed1i \u0111a, t\u01b0\u01a1ng quan v\u1edbi m\u1ee9c ti\u00eau th\u1ee5 \u0111i\u1ec7n n\u0103ng c\u1ee7a n\u00f3. Vi\u1ec7c hi\u1ec3u r\u00f5 TDP l\u00e0 \u0111i\u1ec1u thi\u1ebft y\u1ebfu nh\u1eb1m \u0111\u1ea3m b\u1ea3o c\u00e1c gi\u1ea3i ph\u00e1p l\u00e0m m\u00e1t ph\u00f9 h\u1ee3p c\u0169ng nh\u01b0 s\u1ef1 t\u01b0\u01a1ng th\u00edch v\u1edbi ngu\u1ed3n c\u1ea5p \u0111i\u1ec7n.<\/p>\n\n\n\n

3\ufe0f\u20e3 C\u00e1c \u1ee9ng d\u1ee5ng c\u1ee7a GPU<\/h2>\n\n\n\n
\"Applications<\/figure>\n\n\n\n

Tr\u00f2 ch\u01a1i<\/h3>\n\n\n\n

Trong tr\u00f2 ch\u01a1i, GPU ch\u1ecbu tr\u00e1ch nhi\u1ec7m hi\u1ec3n th\u1ecb \u0111\u1ed3 h\u1ecda ch\u1ea5t l\u01b0\u1ee3ng cao v\u00e0 \u0111\u1ea3m b\u1ea3o tr\u1ea3i nghi\u1ec7m ch\u01a1i m\u01b0\u1ee3t m\u00e0. C\u00e1c t\u00ednh n\u0103ng ti\u00ean ti\u1ebfn nh\u01b0 theo tia (ray tracing) v\u00e0 \u0111\u1ed3 h\u1ecda \u0111\u01b0\u1ee3c t\u0103ng c\u01b0\u1eddng b\u1edfi AI \u0111\u1ec1u do GPU hi\u1ec7n \u0111\u1ea1i cung c\u1ea5p s\u1ee9c m\u1ea1nh.<\/p>\n\n\n\n

Tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o v\u00e0 h\u1ecdc m\u00e1y<\/h3>\n\n\n\n

GPU t\u0103ng t\u1ed1c qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n v\u00e0 suy lu\u1eadn trong AI v\u00e0 h\u1ecdc m\u00e1y b\u1eb1ng c\u00e1ch x\u1eed l\u00fd hi\u1ec7u qu\u1ea3 c\u00e1c t\u1eadp d\u1eef li\u1ec7u l\u1edbn v\u00e0 c\u00e1c ph\u00e9p t\u00ednh ph\u1ee9c t\u1ea1p. Kh\u1ea3 n\u0103ng x\u1eed l\u00fd song song c\u1ee7a GPU khi\u1ebfn ch\u00fang tr\u1edf n\u00ean kh\u00f4ng th\u1ec3 thi\u1ebfu trong c\u00e1c l\u0129nh v\u1ef1c n\u00e0y.<\/p>\n\n\n\n

\u0110i\u1ec7n to\u00e1n hi\u1ec7u n\u0103ng cao (HPC)<\/h3>\n\n\n\n

In HPC<\/a>, GPU \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng cho m\u00f4 ph\u1ecfng, t\u00ednh to\u00e1n khoa h\u1ecdc v\u00e0 ph\u00e2n t\u00edch d\u1eef li\u1ec7u, gi\u1ea3m \u0111\u00e1ng k\u1ec3 th\u1eddi gian t\u00ednh to\u00e1n so v\u1edbi c\u00e1c h\u1ec7 th\u1ed1ng ch\u1ec9 d\u00f9ng CPU.<\/p>\n\n\n\n

4\ufe0f\u20e3 C\u00e1ch ch\u1ecdn GPU ph\u00f9 h\u1ee3p<\/h2>\n\n\n\n

Khi l\u1ef1a ch\u1ecdn GPU, h\u00e3y c\u00e2n nh\u1eafc c\u00e1c y\u1ebfu t\u1ed1 sau:<\/p>\n\n\n\n