Both announced variants of the Instinct MI300 series are the first AMD products to feature more than 100 billion transistors. They will be among the largest mass-produced chips at the time of release.
product transistorsAMD Instinct MI300X (Rigel-X?)153 billionAMD Instinct MI300A (Rigel-A?)146 billionApple M2 Ultra (dual-die)134 billiondApple M1 Ultra (dual-die)114 billiondIntel )80 billiondNvidia GeForce RTX 4090 / AD10276.4 billiondAMD Instinct MI250X (Aldebaraan)58.2 billiondAMD Radeon RX 7900 billiondAMD Phoenix (APU)>25 billiond
The Instinct MI300A is a combination CDNA 3 a Zen 4 chiplets, so it is a processor / APU / SoC, a self-contained unit that integrates both processor and computing cores. The advantage of the clock solution is significantly lower requirements for space and coolers (the space normally used by processors and their coolers can be used for other SoCs). This configuration has been supplied by AMD since June for the El Capitan supercomputer.
The Instinct MI300X forms purely CDNA 3 chiplets, so it is a classic independent accelerator with which (through the OAM interface) the system manufacturer adds a processor to the board.
Instinct
MI100Instinct
MI210Instinct
MI250XInstinct
MI300AInstinct
MI300XdesignationArcturusAldebaranRigelarchitectureCDNACDNA 2CDNA 3CPU24× Zen 4formatPCIePCIeOAMsocket SH5OAMCU/MS120104
(128)220
(256)384+??FP32 jarder76806656
(8192)14080
(16384)24k+??FP64 jader—–INT32 jarder—–Tens. Cores440?416880??rate1502 MHz ≤1700 MHz?? ↓↓↓ T(FL)OPS ↓↓↓FP16
184,6181383??BF16
92,3181383??FP32
23,545,3
22,695,7
47,9??FP64
11,522,647,9??INT4
184,6181383??INT8184,6181383??INT16?????INT32?????FP8 tensor1500
3000*?FP16 tensor184,6181383750??BF16 tensor92,3181383750??FP32 tensor46,145,395,7
190??TF32 tensor
??FP64 tensor
45,395,7190??INT4 tensor??INT8 tensor
184,6181383?? ↑↑↑ T(FL)OPS ↑↑↑TMU480?—bus4096bit4096bit8192bit8192bitcapacity
memoirs32 GB64 GB128 GB128 GB192 GBHBM2,4 GHz3,2 GHz3,2 GHzHBM3 >5 GHzmemory.
permeable1229 GB/s1639 GB/s3277 GB/s5,2 TB/s?5,2 TB/sTDP300 W300 W500W
560W? 750Wtransistors50 mld.
25.6 billion 29.1 billion 58.2 billion 146 billion 153 billionGPU area750 mm²
362 mm²724 mm²660 mm²?trial7 nm6nm6nm5nm+6nmdatum20202022202120232023
AMD plans to publish detailed specifications and launch both models at the Advancing AI event announced for Wednesday, December 6. According to unofficial sources, this date will also be the release, in other words, the start of distribution for other customers as well.
AMD has not yet officially mentioned the Instinct MI300C model, which according to some sources it has in the labs. This variant lacks CDNA 3 cores and stands purely on x86 cores Zen 4 supplemented with HBM3 memory like other models. It is possible that this version will be released later or not at all, as it targets a more specific segment (processor acceleration of language models), a smaller part of the market. It is possible that its possible release will also depend on the total production capacities and their use by the main models Instinct MI300A and MI300X.
2023-11-09 23:08:59
#AMD #launch #Instinct #MI300 #sales #December #Advancing #event
