-
Notifications
You must be signed in to change notification settings - Fork 1
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
1 parent
adaf1cb
commit b8d1550
Showing
1 changed file
with
50 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,50 @@ | ||
Intel® Xeon® Gold 6254 Processor | ||
Total Cores - 18 | ||
Total Threads - 36 | ||
Processor Base Frequency - 3.10 GHz | ||
Memory Types - DDR4-2933 | ||
Maximum Memory Speed - 2933 MHz | ||
Max # of Memory Channels - 6 | ||
|
||
https://www.intel.com/content/www/us/en/products/sku/192451/intel-xeon-gold-6254-processor-24-75m-cache-3-10-ghz/specifications.html | ||
|
||
Dandelion Machine specs | ||
Commodity Server Two-socket 18-core Intel Xeon Gold 6254 | ||
72 h/w threads in total (incl. hyper-threads) | ||
System Memory 8× 32GB DDR4-2933 (256GB in total) | ||
|
||
Bandwidth of 1 dimm | ||
> 2933 * 8 = 23.464 Gb/s | ||
Bandwidth of using 4 Channels | ||
> 23.464 * 4 = 93.856 Gb/s | ||
--- | ||
Cache Lines (64 bytes) | ||
1 dim | ||
> 366_625_000 cache line reads or writes per second | ||
To saturate bandwidth need to request a cache line every 1/366_625_000 seconds. | ||
|
||
4 dims | ||
> 1_466_500_000 cache line reads or writes per second | ||
To saturate bandwidth need to request a cache line every 1/1_466_500_000 seconds | ||
> 1/1_466_500_000 seconds = ~.68 ns | ||
Using all 18 cores for memory operations would mean a request every ~12 ns | ||
> 18/1_466_500_000 = ~12 ns | ||
Or assuming the base frequency of the processor, 3.1 Ghz | ||
> (18/1_466_500_000)* 3100000000 = ~38 cycles | ||
Gives each core a cycle budget of 38 cycles (76 cycles with 2 threads) | ||
|
||
Using both sockets, 8 dim theoretical bandwidth | ||
> 366_625_000*8 = 2_933_000_000 cache line reads or writes per second | ||
Estimating for latency of random reads | ||
> 2_933_000_000 *(85.4/127.8) = ~1_959_923_318 | ||
DRAMHiT Random Memory Reads Latency (Intel Xeon Gold 6142) | ||
> 85.4/127.8 = ~0.67 |