Nvidia Hopper accelerator for the first time in photos

On March 22, Nvidia only released renders of the 4nm GPU GH100 / H100 accelerator, so the NDA has only taken pictures of real hardware until now. The editors of ServeTheHome and CNet were given the opportunity to take a picture of a real accelerator:

Nvidia H100, rear; source: ServeTheHome

Patrick J Kennedy from ServeTheHome (whose pictures you see above) had the opportunity to take a picture of the card (module) on April 7; however, Nvidia did not give permission for the publication until now in May.

source: CNet

There are no peculiarities of the SXM module, perhaps it is worth noting that compared to the Ampere, the GPU has been rotated by 90 °, so that the three HBM memories on each side of the core are oriented “longitudinally” to the longer side of the module.

AMD
MI250X
AMD
MI300
Nvidia
Tesla V100
Nvidia A100Nvidia
H100
GPUAldebaranRigelGV100GA100GH100
architectureCDNA 2CDNA 3LapAmpereHopper
formatOAMOAMSXM2SXM4SXM5
CU / SM220?80108132
FP32 jader14080?5120691215872
16896
FP64 jader256034568448
INT32 jader512069128448
Tensor Cores880?640432528
rate1700 MHz?1530 MHz1410 MHz?
↓↓↓ T(FL)OPS ↓↓↓
FP16
383?31,478120
BF16
383?39120
FP32
95,7
47,8
?15,719,560
FP64
47,8?7,89,730
INT4
383????
INT8383????
INT16?????
INT32??15,719,530
FP8 tensor?2000/4000*
FP16 tensor383?125312/624*
1000/2000*
BF16 tensor383?312/624*
1000/2000*
FP32 tensor95,7?19,560?
TF32 tensor
?156/312*
500/1000*
FP64 tensor
95,7?19,560
INT8 tensor
383?624/1248*
2000/4000*
INT4 tensor
??1248/2496*
?
↑↑↑ T(FL)OPS ↑↑↑
TMU??320432528
LLC16 MB?6 MB40 MB50 MB
bus8192bit?4096bit5120bit5120bit
memory128 GB128 GB32 GB / 16 GB40 GB80 GB80 GB
HBM3,2 GHz?1,755 GHz2,43 GHz3,2 GHz4,8 GHz
pam. prop.3277 GB / s?900 GB / s1555 GB / s2048 GB / s3072 GB / s
TDP500 W
560 W
?300 W400 W700 W
transistors58,2 mld.
?21,1 mld.54,2 mld.80 mld.
GPU area2× ?4× ?815 mm²826 mm²814 mm²
trial6 nm5nm12 nm7 nm4nm
datum11. 20212022?20175. 202011. 20202022?

According to information from the time of the March performance, the novelty should not be available in the first half of this year. In Japan, there was a possibility of pre-ordering the PCIe variant (20% lower number of function blocks, HBM2E instead of HBM3) for $ 33,000.


Source: Diit.cz by diit.cz.

*The article has been translated based on the content of Diit.cz by diit.cz. If there is any problem regarding the content, copyright, please leave a report below the article. We will try to process as quickly as possible to protect the rights of the author. Thank you very much!

*We just want readers to access information more quickly and easily with other multilingual content, instead of information only available in a certain language.

*We always respect the copyright of the content of the author and always include the original link of the source article.If the author disagrees, just leave the report below the article, the article will be edited or deleted at the request of the author. Thanks very much! Best regards!