{"id":24460,"date":"2026-04-26T09:39:46","date_gmt":"2026-04-26T02:39:46","guid":{"rendered":"https:\/\/vnso.vn\/?p=24460"},"modified":"2026-04-26T09:39:46","modified_gmt":"2026-04-26T02:39:46","slug":"so-sanh-nvidia-a100-vs-h100","status":"publish","type":"post","link":"https:\/\/vnso.vn\/en\/so-sanh-nvidia-a100-vs-h100\/","title":{"rendered":"So s\u00e1nh NVIDIA A100 vs H100 chuy\u00ean s\u00e2u – GPU n\u00e0o ph\u00f9 h\u1ee3p?"},"content":{"rendered":"

Vi\u1ec7c so s\u00e1nh, l\u1ef1a ch\u1ecdn gi\u1eefa NVIDIA A100 vs H100 tr\u1edf th\u00e0nh quy\u1ebft \u0111\u1ecbnh \u1ea3nh h\u01b0\u1edfng tr\u1ef1c ti\u1ebfp \u0111\u1ebfn chi ph\u00ed v\u1eadn h\u00e0nh, t\u1ed1c \u0111\u1ed9 tri\u1ec3n khai v\u00e0 kh\u1ea3 n\u0103ng m\u1edf r\u1ed9ng h\u1ec7 th\u1ed1ng AI.<\/p>\n

A100 t\u1eebng l\u00e0 ti\u00eau chu\u1ea9n v\u00e0ng trong giai \u0111o\u1ea1n 2020\u20132023. Tuy nhi\u00ean, H100 ra \u0111\u1eddi v\u1edbi ki\u1ebfn tr\u00fac Hopper \u0111\u00e3 thay \u0111\u1ed5i ho\u00e0n to\u00e0n c\u00e1ch GPU ph\u1ee5c v\u1ee5 AI, \u0111\u1eb7c bi\u1ec7t trong b\u1ed1i c\u1ea3nh m\u00f4 h\u00ecnh Transformer v\u00e0 LLM tr\u1edf th\u00e0nh trung t\u00e2m.<\/p>\n

Theo c\u00e1c benchmark th\u1ef1c t\u1ebf, H100 c\u00f3 th\u1ec3 mang l\u1ea1i hi\u1ec7u n\u0103ng cao h\u01a1n t\u1eeb 3 \u0111\u1ebfn 6 l\u1ea7n so v\u1edbi A100<\/strong>, trong khi chi ph\u00ed t\u0103ng kho\u1ea3ng 50% \u0111\u1ebfn 170%<\/strong> t\u00f9y m\u00f4i tr\u01b0\u1eddng cloud v\u00e0 c\u1ea5u h\u00ecnh. C\u00f9ng VNSO so s\u00e1nh GPU Server AI NVIDIA A100<\/a> vs H100<\/a>, v\u00e0 t\u1ea1i sao kh\u00f4ng ph\u1ea3i ch\u1ecdn GPU m\u1ea1nh h\u01a1n l\u00e0 t\u1ed1t.<\/p>\n

<\/section>\n
\n
\n
\n
\n
\n
\n
\n

NVIDIA A100 v\u00e0 H100 l\u00e0 g\u00ec?<\/h2>\n

NVIDIA A100 l\u00e0 GPU trung t\u00e2m d\u1eef li\u1ec7u thu\u1ed9c ki\u1ebfn tr\u00fac Ampere, ra m\u1eaft n\u0103m 2020, \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1ec3 t\u0103ng t\u1ed1c c\u00e1c t\u00e1c v\u1ee5 AI, Machine Learning v\u00e0 HPC. GPU n\u00e0y s\u1eed d\u1ee5ng b\u1ed9 nh\u1edb HBM2e dung l\u01b0\u1ee3ng t\u1ed1i \u0111a 80GB, b\u0103ng th\u00f4ng kho\u1ea3ng 2 TB\/s v\u00e0 nhanh ch\u00f3ng tr\u1edf th\u00e0nh ti\u00eau chu\u1ea9n cho training AI trong nhi\u1ec1u n\u0103m. A100 n\u1ed5i b\u1eadt nh\u1edd kh\u1ea3 n\u0103ng chia t\u00e0i nguy\u00ean GPU th\u00f4ng qua c\u00f4ng ngh\u1ec7 MIG, gi\u00fap t\u1ed1i \u01b0u chi ph\u00ed trong m\u00f4i tr\u01b0\u1eddng cloud v\u00e0 doanh nghi\u1ec7p.<\/p>\n

\"GPU<\/a>

H\u00ecnh \u1ea3nh th\u1ef1c t\u1ebf c\u1ee7a NVIDIA A100<\/p><\/div>\n

NVIDIA H100 l\u00e0 th\u1ebf h\u1ec7 GPU k\u1ebf nhi\u1ec7m, d\u1ef1a tr\u00ean ki\u1ebfn tr\u00fac Hopper, ra m\u1eaft t\u1eeb 2022\u20132023, \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u tr\u1ef1c ti\u1ebfp cho AI hi\u1ec7n \u0111\u1ea1i, \u0111\u1eb7c bi\u1ec7t l\u00e0 c\u00e1c m\u00f4 h\u00ecnh Transformer v\u00e0 LLM. H100 s\u1eed d\u1ee5ng b\u1ed9 nh\u1edb HBM3 v\u1edbi b\u0103ng th\u00f4ng l\u00ean t\u1edbi kho\u1ea3ng 3.35 TB\/s, \u0111\u1ed3ng th\u1eddi t\u00edch h\u1ee3p Transformer Engine v\u00e0 h\u1ed7 tr\u1ee3 FP8, gi\u00fap t\u0103ng t\u1ed1c \u0111\u00e1ng k\u1ec3 c\u1ea3 training v\u00e0 inference. \u0110\u00e2y l\u00e0 GPU \u0111\u01b0\u1ee3c \u0111\u1ecbnh h\u01b0\u1edbng cho c\u00e1c h\u1ec7 th\u1ed1ng AI quy m\u00f4 l\u1edbn v\u00e0 m\u00f4i tr\u01b0\u1eddng production.<\/p>\n

\"GPU<\/a>

GPU NVIDIA H100<\/p><\/div>\n

>>> Xem th\u00eam Thu\u00ea NVIDIA A100 Cloud GPU gi\u00e1 ch\u1ec9 t\u1eeb 59k\/gi\u1edd<\/a><\/strong><\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/section>\n

So s\u00e1nh \u0111i\u1ec3m kh\u00e1c bi\u1ec7t c\u1ee7a th\u00f4ng s\u1ed1 k\u1ef9 thu\u1eadt NVIDIA A100 vs H100<\/h2>\n

S\u1ef1 kh\u00e1c bi\u1ec7t gi\u1eefa hai GPU kh\u00f4ng n\u1eb1m \u1edf m\u1ed9t ch\u1ec9 s\u1ed1 ri\u00eang l\u1ebb m\u00e0 l\u00e0 t\u1ed5ng h\u00f2a c\u1ee7a compute, b\u1ed9 nh\u1edb v\u00e0 k\u1ebft n\u1ed1i.<\/p>\n

A100 s\u1eed d\u1ee5ng b\u1ed9 nh\u1edb HBM2e v\u1edbi bandwidth kho\u1ea3ng 1.9\u20132.0 TB\/s, trong khi H100 n\u00e2ng c\u1ea5p l\u00ean HBM3 \u0111\u1ea1t kho\u1ea3ng 3.35 TB\/s, t\u1ee9c t\u0103ng h\u01a1n 70%. \u0110i\u1ec1u n\u00e0y \u0111\u1eb7c bi\u1ec7t quan tr\u1ecdng v\u1edbi c\u00e1c m\u00f4 h\u00ecnh l\u1edbn v\u1ed1n ph\u1ee5 thu\u1ed9c m\u1ea1nh v\u00e0o t\u1ed1c \u0111\u1ed9 truy xu\u1ea5t d\u1eef li\u1ec7u.<\/p>\n

V\u1ec1 s\u1ee9c m\u1ea1nh t\u00ednh to\u00e1n, A100 \u0111\u1ea1t kho\u1ea3ng 19.5 TFLOPS FP32, c\u00f2n H100 \u0111\u1ea1t t\u1edbi 60\u201367 TFLOPS, t\u01b0\u01a1ng \u0111\u01b0\u01a1ng h\u01a1n 3 l\u1ea7n. Khi x\u00e9t \u0111\u1ebfn AI workload v\u1edbi FP16\/BF16, kho\u1ea3ng c\u00e1ch c\u00f2n l\u1edbn h\u01a1n khi H100 \u0111\u1ea1t t\u1edbi g\u1ea7n 1000\u20132000 TFLOPS, so v\u1edbi kho\u1ea3ng 312 TFLOPS c\u1ee7a A100 .<\/p>\n

M\u1ed9t \u0111i\u1ec3m mang t\u00ednh b\u01b0\u1edbc ngo\u1eb7t l\u00e0 H100 h\u1ed7 tr\u1ee3 FP8 \u2013 \u0111\u1ecbnh d\u1ea1ng s\u1ed1 m\u1edbi gi\u00fap t\u0103ng t\u1ed1c AI m\u00e0 v\u1eabn \u0111\u1ea3m b\u1ea3o \u0111\u1ed9 ch\u00ednh x\u00e1c. \u0110\u00e2y l\u00e0 y\u1ebfu t\u1ed1 A100 ho\u00e0n to\u00e0n kh\u00f4ng c\u00f3.<\/p>\n

Hi\u1ec7u n\u0103ng th\u1ef1c t\u1ebf c\u1ee7a NVIDIA A100 vs H100<\/h2>\n

V\u00ec sao H100 l\u00e0 GPU \u201cAI-native\u201d \u1edf m\u1eb7t ki\u1ebfn tr\u00fac<\/h3>\n

A100 d\u1ef1a tr\u00ean ki\u1ebfn tr\u00fac Ampere v\u1edbi Tensor Core th\u1ebf h\u1ec7 3, h\u1ed7 tr\u1ee3 t\u1ed1t Deep Learning truy\u1ec1n th\u1ed1ng. N\u00f3 ho\u1ea1t \u0111\u1ed9ng hi\u1ec7u qu\u1ea3 v\u1edbi CNN, NLP c\u01a1 b\u1ea3n v\u00e0 nhi\u1ec1u workload AI ph\u1ed5 bi\u1ebfn.<\/p>\n

Tuy nhi\u00ean, H100 v\u1edbi ki\u1ebfn tr\u00fac Hopper \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf xoay quanh Transformer \u2013 n\u1ec1n t\u1ea3ng c\u1ee7a LLM hi\u1ec7n \u0111\u1ea1i. \u0110i\u1ec3m c\u1ed1t l\u00f5i n\u1eb1m \u1edf Transformer Engine, cho ph\u00e9p GPU t\u1ef1 \u0111\u1ed9ng chuy\u1ec3n \u0111\u1ed5i gi\u1eefa FP8 v\u00e0 FP16 nh\u1eb1m t\u1ed1i \u01b0u hi\u1ec7u n\u0103ng v\u00e0 b\u0103ng th\u00f4ng.<\/p>\n

Theo NVIDIA, vi\u1ec7c s\u1eed d\u1ee5ng Transformer Engine c\u00f3 th\u1ec3 gi\u00fap t\u0103ng t\u1ed1c training l\u00ean t\u1edbi 9 l\u1ea7n trong m\u1ed9t s\u1ed1 tr\u01b0\u1eddng h\u1ee3p<\/strong> v\u00e0 c\u1ea3i thi\u1ec7n \u0111\u00e1ng k\u1ec3 hi\u1ec7u su\u1ea5t inference.<\/p>\n

Training m\u00f4 h\u00ecnh AI v\u00e0 LLM<\/h3>\n

Trong c\u00e1c b\u00e0i test training LLM c\u1ee1 l\u1edbn, h\u1ec7 th\u1ed1ng s\u1eed d\u1ee5ng H100 cho throughput cao h\u01a1n kho\u1ea3ng 3 l\u1ea7n so v\u1edbi A100<\/strong>. V\u00ed d\u1ee5, v\u1edbi c\u00f9ng c\u1ea5u h\u00ecnh 8 GPU, H100 \u0111\u1ea1t kho\u1ea3ng 1350 samples\/gi\u00e2y<\/strong>, trong khi A100 ch\u1ec9 kho\u1ea3ng 450 samples\/gi\u00e2y<\/strong>. \u0110i\u1ec1u n\u00e0y gi\u00fap r\u00fat ng\u1eafn th\u1eddi gian training t\u1eeb v\u00e0i tu\u1ea7n xu\u1ed1ng c\u00f2n v\u00e0i ng\u00e0y, m\u1ed9t y\u1ebfu t\u1ed1 c\u00f3 \u00fd ngh\u0129a l\u1edbn trong m\u00f4i tr\u01b0\u1eddng AI c\u1ea1nh tranh.<\/p>\n

Inference v\u00e0 tri\u1ec3n khai AI c\u1ee7a NVIDIA A100 vs H100<\/h3>\n

Trong inference, \u0111\u1eb7c bi\u1ec7t v\u1edbi c\u00e1c m\u00f4 h\u00ecnh LLM 13B\u201370B, H100 ti\u1ebfp t\u1ee5c th\u1ec3 hi\u1ec7n \u01b0u th\u1ebf khi \u0111\u1ea1t t\u1ed1c \u0111\u1ed9 250\u2013300 tokens\/gi\u00e2y<\/strong>, trong khi A100 th\u01b0\u1eddng ch\u1ec9 \u0111\u1ea1t kho\u1ea3ng 130 tokens\/gi\u00e2y<\/strong>.<\/p>\n

M\u1ed9t s\u1ed1 benchmark MLPerf cho th\u1ea5y H100 c\u00f3 th\u1ec3 nhanh h\u01a1n t\u1edbi 4.5 l\u1ea7n trong inference<\/strong>, \u0111\u1eb7c bi\u1ec7t khi s\u1eed d\u1ee5ng FP8. N\u1ebfu h\u1ec7 th\u1ed1ng AI ph\u1ee5c v\u1ee5 ng\u01b0\u1eddi d\u00f9ng cu\u1ed1i, H100 mang l\u1ea1i tr\u1ea3i nghi\u1ec7m t\u1ed1t h\u01a1n nh\u1edd \u0111\u1ed9 tr\u1ec5 th\u1ea5p h\u01a1n.<\/p>\n

Kh\u1ea3 n\u0103ng m\u1edf r\u1ed9ng v\u00e0 multi-GPU<\/h3>\n

Trong m\u00f4i tr\u01b0\u1eddng data center, kh\u1ea3 n\u0103ng scale quy\u1ebft \u0111\u1ecbnh hi\u1ec7u qu\u1ea3 t\u1ed5ng th\u1ec3. H100 s\u1edf h\u1eefu NVLink b\u0103ng th\u00f4ng 900 GB\/s<\/strong>, cao h\u01a1n \u0111\u00e1ng k\u1ec3 so v\u1edbi 600 GB\/s<\/strong> c\u1ee7a A100.<\/p>\n

Khi tri\u1ec3n khai cluster l\u1edbn, \u0111i\u1ec1u n\u00e0y gi\u00fap gi\u1ea3m bottleneck truy\u1ec1n d\u1eef li\u1ec7u gi\u1eefa GPU, t\u1eeb \u0111\u00f3 duy tr\u00ec hi\u1ec7u su\u1ea5t cao h\u01a1n khi m\u1edf r\u1ed9ng.<\/p>\n

Chi ph\u00ed v\u1eadn h\u00e0nh NVIDIA A100 vs H100: B\u00e0i to\u00e1n th\u1ef1c t\u1ebf c\u1ee7a doanh nghi\u1ec7p<\/h2>\n

V\u1ec1 gi\u00e1 cloud, A100 th\u01b0\u1eddng c\u00f3 m\u1ee9c kho\u1ea3ng 0.6 USD\/gi\u1edd<\/strong>, trong khi H100 dao \u0111\u1ed9ng kho\u1ea3ng 2 USD\/gi\u1edd<\/strong>. Tho\u1ea1t nh\u00ecn, H100 \u0111\u1eaft h\u01a1n g\u1ea5p 2\u20133 l\u1ea7n. Tuy nhi\u00ean, n\u1ebfu x\u00e9t tr\u00ean hi\u1ec7u n\u0103ng:<\/p>\n

– H100 nhanh h\u01a1n 3-6 l\u1ea7n<\/p>\n

– Th\u1eddi gian x\u1eed l\u00fd gi\u1ea3m m\u1ea1nh<\/p>\n

– S\u1ed1 l\u01b0\u1ee3ng GPU c\u1ea7n thi\u1ebft \u00edt h\u01a1n<\/p>\n

T\u1ed5ng chi ph\u00ed s\u1edf h\u1eefu (TCO) c\u1ee7a H100 trong nhi\u1ec1u tr\u01b0\u1eddng h\u1ee3p c\u00f3 th\u1ec3 th\u1ea5p h\u01a1n khi tri\u1ec3n khai \u1edf quy m\u00f4 l\u1edbn.<\/p>\n

\u0110i\u1ec7n n\u0103ng v\u00e0 h\u1ea1 t\u1ea7ng: y\u1ebfu t\u1ed1 \u00edt \u0111\u01b0\u1ee3c nh\u1eafc \u0111\u1ebfn<\/h3>\n

H100 c\u00f3 m\u1ee9c ti\u00eau th\u1ee5 \u0111i\u1ec7n kho\u1ea3ng 700W<\/strong>, cao h\u01a1n \u0111\u00e1ng k\u1ec3 so v\u1edbi 400W c\u1ee7a A100.<\/strong><\/p>\n

Tuy nhi\u00ean, do ho\u00e0n th\u00e0nh c\u00f4ng vi\u1ec7c nhanh h\u01a1n, t\u1ed5ng \u0111i\u1ec7n n\u0103ng ti\u00eau th\u1ee5 tr\u00ean m\u1ed7i workload c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u. M\u1ed9t s\u1ed1 nghi\u00ean c\u1ee9u cho th\u1ea5y vi\u1ec7c s\u1eed d\u1ee5ng H100 c\u00f3 th\u1ec3 gi\u00fap gi\u1ea3m \u0111\u00e1ng k\u1ec3 chi ph\u00ed v\u1eadn h\u00e0nh data center n\u1ebfu \u0111\u01b0\u1ee3c tri\u1ec3n khai \u0111\u00fang c\u00e1ch.<\/p>\n

GPU n\u00e0o ph\u00f9 h\u1ee3p? G\u00f3c nh\u00ecn th\u1ef1c t\u1ebf cho doanh nghi\u1ec7p<\/h2>\n

Vi\u1ec7c l\u1ef1a ch\u1ecdn kh\u00f4ng n\u00ean d\u1ef1a tr\u00ean GPU m\u1ea1nh nh\u1ea5t, m\u00e0 d\u1ef1a tr\u00ean b\u00e0i to\u00e1n c\u1ee5 th\u1ec3.<\/p>\n

Trong c\u00e1c d\u1ef1 \u00e1n AI quy m\u00f4 v\u1eeba, startup ho\u1eb7c doanh nghi\u1ec7p SMEs, A100 v\u1eabn l\u00e0 l\u1ef1a ch\u1ecdn h\u1ee3p l\u00fd nh\u1edd chi ph\u00ed th\u1ea5p, d\u1ec5 tri\u1ec3n khai v\u00e0 h\u1ec7 sinh th\u00e1i \u1ed5n \u0111\u1ecbnh. V\u1edbi c\u00e1c m\u00f4 h\u00ecnh d\u01b0\u1edbi 30B parameters ho\u1eb7c c\u00e1c workload AI truy\u1ec1n th\u1ed1ng, A100 \u0111\u00e1p \u1ee9ng t\u1ed1t c\u1ea3 v\u1ec1 hi\u1ec7u n\u0103ng l\u1eabn chi ph\u00ed.<\/p>\n

Ng\u01b0\u1ee3c l\u1ea1i, v\u1edbi c\u00e1c h\u1ec7 th\u1ed1ng AI quy m\u00f4 l\u1edbn, \u0111\u1eb7c bi\u1ec7t l\u00e0 LLM, inference production ho\u1eb7c c\u00e1c n\u1ec1n t\u1ea3ng AI ph\u1ee5c v\u1ee5 h\u00e0ng ngh\u00ecn ng\u01b0\u1eddi d\u00f9ng \u0111\u1ed3ng th\u1eddi, H100 tr\u1edf th\u00e0nh l\u1ef1a ch\u1ecdn g\u1ea7n nh\u01b0 b\u1eaft bu\u1ed9c. L\u1ee3i th\u1ebf v\u1ec1 t\u1ed1c \u0111\u1ed9, \u0111\u1ed9 tr\u1ec5 v\u00e0 kh\u1ea3 n\u0103ng m\u1edf r\u1ed9ng gi\u00fap t\u1ed1i \u01b0u hi\u1ec7u qu\u1ea3 v\u1eadn h\u00e0nh d\u00e0i h\u1ea1n.<\/p>\n

K\u1ebft lu\u1eadn<\/h2>\n

So s\u00e1nh NVIDIA A100 v\u00e0 NVIDIA H100 cho th\u1ea5y hai vai tr\u00f2 r\u1ea5t r\u00f5 r\u00e0ng trong h\u1ec7 sinh th\u00e1i AI hi\u1ec7n nay. A100 ph\u00f9 h\u1ee3p v\u1edbi c\u00e1c h\u1ec7 th\u1ed1ng c\u1ea7n t\u1ed1i \u01b0u chi ph\u00ed v\u00e0 \u0111\u1ed9 \u1ed5n \u0111\u1ecbnh. H100 \u0111\u1ea1i di\u1ec7n cho th\u1ebf h\u1ec7 h\u1ea1 t\u1ea7ng AI m\u1edbi, n\u01a1i hi\u1ec7u n\u0103ng v\u00e0 kh\u1ea3 n\u0103ng scale l\u00e0 y\u1ebfu t\u1ed1 quy\u1ebft \u0111\u1ecbnh.<\/p>\n

Trong b\u1ed1i c\u1ea3nh AI n\u0103m 2026, xu h\u01b0\u1edbng \u0111ang nghi\u00eang v\u1ec1 H100 cho c\u00e1c h\u1ec7 th\u1ed1ng l\u1edbn, trong khi A100 v\u1eabn gi\u1eef vai tr\u00f2 n\u1ec1n t\u1ea3ng cho ph\u1ea7n l\u1edbn doanh nghi\u1ec7p. L\u1ef1a ch\u1ecdn \u0111\u00fang GPU kh\u00f4ng ch\u1ec9 gi\u00fap ti\u1ebft ki\u1ec7m chi ph\u00ed m\u00e0 c\u00f2n quy\u1ebft \u0111\u1ecbnh t\u1ed1c \u0111\u1ed9 tri\u1ec3n khai v\u00e0 kh\u1ea3 n\u0103ng c\u1ea1nh tranh c\u1ee7a to\u00e0n b\u1ed9 h\u1ec7 th\u1ed1ng AI.<\/p>\n

Tri\u1ec3n khai GPU AI hi\u1ec7u n\u0103ng cao, \u1ed5n \u0111\u1ecbnh cho doanh nghi\u1ec7p Vi\u1ec7t<\/h3>\n

VNSO cung c\u1ea5p h\u1ea1 t\u1ea7ng Cloud GPU s\u1eed d\u1ee5ng NVIDIA A100 ch\u1ec9 t\u1eeb 59k\/gi\u1edd<\/a> v\u00e0 Server GPU NVIDIA H100<\/a>, s\u1eb5n s\u00e0ng cho m\u1ecdi nhu c\u1ea7u t\u1eeb training AI \u0111\u1ebfn tri\u1ec3n khai LLM production. H\u1ea1 t\u1ea7ng t\u1ed1i \u01b0u hi\u1ec7u n\u0103ng, tri\u1ec3n khai nhanh, linh ho\u1ea1t m\u1edf r\u1ed9ng theo th\u1ef1c t\u1ebf s\u1eed d\u1ee5ng.<\/p>\n

>>> H\u00e3y \u0111\u1ec3 l\u1ea1i th\u00f4ng tin \u0111\u1ec3 VNSO li\u00ean h\u1ec7 h\u1ed7 tr\u1ee3 t\u01b0 v\u1ea5n tri\u1ec3n khai t\u1eeb A-Z tr\u00ean h\u1ec7 th\u1ed1ng c\u1ee7a b\u1ea1n!<\/p>\n

\n

\n

<\/p>

    <\/ul><\/div>\n
    \n
    \n\n\n\n\n\n\n\n<\/div>\n