On November 8, AMD Instinct MI200 and the new Epycy will be introduced at the Accelerated Data Center Premiere

AMD CEO Lisa Su, Forrest Norrod chief of data center and Dan McNamara, head of servers, will speak at the Accelerated Data Center Premiere on Monday, November 8, to introduce some news. It has not been officially announced what the products will be, but according to leakers, AMD is going to introduce the CDNA 2 architecture and the accelerators built on it from the Instinct MI200 family (MI250 and MI250X), as well as some of the processor innovations based on Zen 3 – Either Milan-X (ie Zen 3 with V-cache for servers) or Trento (server Zen 3 s podporou Unified Memory Architecture).

Additional information on Instinct accelerators has emerged in connection with these plans. The first is more of a marketing one: AMD is abandoning the use of the term GPU in computing GPUs. It goes to accelerators. However, this change is not justified: CDNA 2 no longer contains any fixed units used to accelerate 3D graphics. Missing rasterizer, missing texturing units, missing ROP, missing accelerators for ray-tracing. Only the multimedia circuit for video acceleration remains. Some sources claimed that fixed units for 3D graphics were already missing from CDNA (1), while others said at least some were present. Anyway, with CDNA 2 they are a thing of the past.

AMD Radeon
Instinct MI60
Instinct
MI100
Instinct
MI250X
Instinct
MI300
Nvidia A100
GPUVega 20ArcturusAldebaranRigelGA100
architectureGCN 4CDNACDNA 2CDNA 3Ampere
CPU
formatPCIePCIeOAMOAMSXM4 / PCIe
CU / SM60120220
(256)
(384-512?)108
FP32 jader3840768014080
(16384)
(24k-33k?)6912
FP64 jader3456
INT32 jader6912
You have. Colors???432
rate1800 MHz1502 MHz≤1700 MHz?1410 MHz
↓↓↓ T(FL)OPS ↓↓↓
FP16
29,5184,6383?78
BF16
92,3383?39
FP32
14,723,595,7?19,5
FP64
7,411,547,9?9,7
INT4
118184,6???
INT859,0
184,6???
INT1629,5????
INT32????19,5
FP16 tensor184,6383??312/624*
BF16 tensor92,3383??312/624*
FP32 tensor46,195,7
?19,5
TF32 tensor
?156/312*
FP64 tensor
47,9??19,5
INT8 tensor
184,6383??624/1248*
INT4 tensor
?1248/2496*
↑↑↑ T(FL)OPS ↑↑↑
TMU240480??432
bus4096bit4096bit8192bit?5120bit
capacity
memoirs
32 GB32 GB128 GB?40 GB
80 GB
HBM22,0 GHz2,4 GHz3,2 GHzHBM3?2,43 GHz
3,20 GHz
memory.
permeable
1024 GB/s1229 GB/s3277 GB/s?1555 GB/s
2048 GB/s
TDP300 W300 W500W~600W?400 / 250 W
transistorů13.2 billion50.0 billion
>100 mld.??54.2 billion
GPU area331 mm²750 mm²
??826 mm²
process7 nm7 nm7nm??7 nm
date2018202020212022-20232020

Newly (though still unofficially) “confirmed” values ​​are highlighted in bold, more significant changes in red

The situation with support for the FP32 format is further clarified. CDNA is the first GPU-based architecture to natively support the full-speed FP64 format. However, the sources were inconsistent in terms of the speed of FP32 support. CDNA 2 supports packed-FP32, which means FP32 processing: FP64 2: 1, in other words the Instinct MI250X will reach up to 95.7 TFLOPS in FP32. The performance in FP64, FP32 and FP16 is therefore five times higher than that of the Nvidia A100, the performance in BF16 format is probably ten times.

Paradoxically, although all the essentials are known about the MI200 series accelerators, practically nothing has been leaked about the processors so far, so either AMD manages to keep these innovations under wraps better, or the emphasis of the action will be on accelerators.


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!