Proceedings of the Second International Symposium on Memory Systems 2016
DOI: 10.1145/2989081.2989099
|View full text |Cite
|
Sign up to set email alerts
|

Exploring Time and Energy for Complex Accesses to a Hybrid Memory Cube

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
1
1
1

Citation Types

2
11
0

Year Published

2017
2017
2022
2022

Publication Types

Select...
4
2

Relationship

0
6

Authors

Journals

citations
Cited by 12 publications
(13 citation statements)
references
References 9 publications
2
11
0
Order By: Relevance
“…AC-510 accelerator with (a) and without the heatsink (b), and two images of the HMC with various temperature, taken by the thermal camera showing heatsink surface temperature (c,d) 9. …”
mentioning
confidence: 99%
See 1 more Smart Citation
“…AC-510 accelerator with (a) and without the heatsink (b), and two images of the HMC with various temperature, taken by the thermal camera showing heatsink surface temperature (c,d) 9. …”
mentioning
confidence: 99%
“…Part#: xcku060-ffva1156-2-e 8 For all experiments, after 200 seconds, temperature is stable 9. To display the heat island created by HMC, we modify the heat scale, also shown at the right of each image 10.…”
mentioning
confidence: 99%
“…To tune the number of accesses and the size of request packets, we use the multi-port stream implementation. Figure 7 depicts that as the number of requests in a stream (stream in this context means a limited number of requests) increases from one to 55, the average latency increases from 0.7 to 1.1 µs for the request size of 16 Fig. 7: Average latency of low-load accesses for various request sizes for the number of requests in the range of one to 55.…”
Section: B Low-contention Access Latencymentioning
confidence: 99%
“…We design the PIM core to be able to function in two operating modes: regular and high memory load (HML). Previous work in [102] [103] demonstrate that the HMC DRAM read/write delay increases as the percentage of pending read/write requests increases too. To this end a read/write request may take from some ns up to several µs [103] depending on the amount of pending DRAM operations.…”
Section: Pim Core Architecturementioning
confidence: 98%
“…We also implement 6 PIM designs that operate in various supply voltages (0.72V, 0.81V and 0.99V) and facilitate pipelined (PE) and non-pipelined (NPE) functional units, as described in the previous section. We set the PCU HML threshold, to 53% due to the fact that beyond this limit the HMC DRAM latency dramatically increases [102].In order to maximize the RISC-V BOOM core pipeline performance,we set its clock frequency to 800 MHz. On the other hand, the PIM cores operate under an adaptive clock frequency of range 200MHz -1.5GHz, depending on tRCD=8ns tCAS=4ns Serial links 160 GBps,6-cycle latency the executing instruction timing requirements and the supply voltage of the PIM implementation.…”
Section: Design Space Exploration and Parameter Considerationsmentioning
confidence: 99%