IEEE Circuits and Systems Magazine - Q3 2021 - 51

to 1, according to the algorithm weight range), or random
drift, with a rate of conductance drift by 10% over 10 years.
The results show that, in scenarios with fixed drift directions,
drifting to maximum or minimal states degrades the
accuracy faster than drifting to the middle states, while
the random drift is the most robust for maintaining inference
accuracy. More experimental characterizations are
still needed to fully capture the retention behaviors of
multilevel states of the eNVMs.
Next, we briefly discuss the properties that matter
for training.
Endurance: On-chip training requires frequent write
operations to the memory cells, however, it is often
thought that endurance requirement is prohibitive for
today's eNVMs. It turns out that endurance ~106 cycles
is sufficient for most of the training workloads if using
the batch-mode training. For example, the real training
traces of activations, errors and gradients are tracked
during VGG-8 training with 150 epochs, the average
number of write for each cell is around 412,500 times
thanks to the sparsity [119]. It is also worthwhile to note
that the weight update is incremental, and the switching
of eNVMs is thus not fully between the highest conductance
and the lowest conductance. The intermediate
switching could relax the endurance requirement.
For example, RRAM could switch ~105 cycles in the full
switching while could switch ~1011 cycles in the intermediate
switching [120].
Variation: Device-to-device variation is typically absorbed
in the iterative training [49], and could be tightened
with write-verify for the inference [90]. Cycle-tocycle
variation is more problematic for training if the
randomness exceeds the deterministic weight update
direction [49]. However, a small amount of noise injected
into the training actually could improve the robustness
of the DNN model against variations in the subsequent
inference [91], as the system could be converged to a local
minima where its energy landscape of loss function
is shallow (thus insensitive to variations).
Asymmetry/nonlinearity: It is known that asymmetry
and nonlinearity in the weight update curve (conductance
vs. # identical programming pulses) generally
degrade in-situ training accuracy [49] [50] [121]. Two algorithmic
tricks (Tiki-Taka in weight update [122], and
momentum in weight update [123]) gave hopes of relaxed
requirements on linearity and symmetry. However,
realizing software-comparable in-situ training accuracy
is still quite challenging for complex dataset. Some alternative
hybrid-precision synapse designs [124] [125]
were recently proposed, which leveraged a volatile capacitor
for symmetric and linear weight fine tuning and
a non-volatile memory component for weight coarse
tuning, thereby approaching software baseline training
THIRD QUARTER 2021
accuracy. It should be pointed out that nonlinearity and
asymmetry is not a problem for inference, as write-verify
could be used to enforce a linear mapping between
weights in the algorithm and conductance of the devices.
6. Outlook
In conclusion, CIM paradigm saves the intermediate
data movement and parallelizes the computation,
thereby significantly improving energy efficiency and
throughput, as demonstrated in the recent prototype
chips at macro-level. We believe that the first tangible
application of CIM based accelerators would be the
edge inference. On-chip training acceleration with
SRAM or hybrid precision synapses is feasible [126]
[127], but the bottleneck is still DRAM access as the intermediate
data during training could easily goes up to
GB or tens of GB level even for a moderate size of DNN
model with 10MB-level weights. In-situ training from
Drift to
Min G
Drift to
Max G
Log (t)
Drift to
Mid G
Random
Drift
Log (t)
(a) Drift Scenarios
100
Baseline: 92%
80
60
40
101
103
105
Time (s)
(b) 10% G Drift at 10 Years
Figure 12. (a) Various scenarios of conductance drift. (b)
CIFAR-10 inference accuracy vs. time considering conductance
drift.
IEEE CIRCUITS AND SYSTEMS MAGAZINE
51
107
109
Drift Target
1
0.75
0.5
0.25
-0.25
-0.5
-0.75
-1
Random
Log (t)
Log (t)
Inference Accuracy (%)
Log (G)
Log (G)
Log (G)
Log (G)
at 10 Years
IEEE Circuits and Systems Magazine - Q3 2021

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2021
