Insert new simulation results

src/acronyms.tex

@@ -352,3 +352,7 @@
     short = HAXPY,
     long = half precision $a \cdot x + y$,
 }
+\DeclareAcronym{hpc}{
+    short = HPC,
+    long = high-performance computing,
+}

src/appendix.tex

@@ -3,8 +3,6 @@
 
 \subsection{Simulation Results}
 
- TODO !! nochmal aktualisieren!
-
 \begin{table}[H]
 \csvreader[
         head to column names,

src/chapters/conclusion.tex

@@ -13,7 +13,7 @@ For this, more detailed information is required from Samsung, as the exact inter
 To ease the currently error-prone microkernel development process, the software library could help the developer by providing building blocks that assemble the microkernel and simultaneously generate the necessary \ac{ld} and \ac{st} instructions to execute the kernel.
 In addition, the current bare-metal deployment of the software cannot realistically be used to accelerate real-world \ac{dnn} applications.
 Instead, \aca{fimdram} should be able to be used on a Linux system, which would require the integration of the software support library into a Linux device driver.
-To take into account the special alignment requirements of the \ac{pim} data structures, this device driver must also carefully consider the virtual address translation of the Linux kernel, possibly making use of so-called \acp{hugetlb}, as the alignment requirements exceed the default page size of $\qty{4}{\kilo\byte}$.
+To take into account the special alignment requirements of the \ac{pim} data structures, this device driver must also carefully consider the virtual address translation of the Linux kernel, possibly making use of so-called \acp{hugetlb}, as the alignment requirements exceed the default page size of $\qty{4}{\kibi\byte}$.
 
 For a better evaluation of the performance gains of \aca{fimdram}, it should be also compared with real-world \ac{dnn} applications.
 Effects such as the initialization overhead of \aca{fimdram} can only be evaluated in such an environment.

src/chapters/implementation/kernel.tex

@@ -53,25 +53,25 @@ In the attributes of the page table, each mapped block of address space can be a
 While most of the \ac{dram} area are should be a normal, cacheable memory region, the \ac{pim} region should be marked as a non-cacheable memory for reasons explained in \cref{sec:microkernel_execution}.
 Furthermore, special memory-mapped devices such as the \ac{uart}, which is used to print logging messages to the \ac{stdout}, must be marked as a non-cacheable device region, as otherwise the log messages may get held in the cache and not be written until the cache line is eventually flushed.
 
-In the AArch64 execution mode, the operating system can choose from three different granule sizes for the translation tables: $\qty{4}{\kilo\byte}$, $\qty{16}{\kilo\byte}$ and $\qty{64}{\kilo\byte}$.
-Each granule size has a different maximum amount of page table nesting, with up to a 4-level look-up for the $\qty{4}{\kilo\byte}$ configuration, as shown in \cref{img:pagetable_granule}.
+In the AArch64 execution mode, the operating system can choose from three different granule sizes for the translation tables: $\qty{4}{\kibi\byte}$, $\qty{16}{\kibi\byte}$ and $\qty{64}{\kibi\byte}$.
+Each granule size has a different maximum amount of page table nesting, with up to a 4-level look-up for the $\qty{4}{\kibi\byte}$ configuration, as shown in \cref{img:pagetable_granule}.
 
 \begin{figure}
 	\centering
 	\includegraphics[width=\linewidth]{images/pagetable_granule}
-	\caption[The distinct page table levels for the $\qty{4}{\kilo\byte}$ granule.]{The distinct page table levels for the $\qty{4}{\kilo\byte}$ granule \cite{arm2015}.}
+	\caption[The distinct page table levels for the $\qty{4}{\kibi\byte}$ granule.]{The distinct page table levels for the $\qty{4}{\kibi\byte}$ granule \cite{arm2015}.}
 	\label{img:pagetable_granule}
 \end{figure}
 
 As it can be seen, when using the complete 4-level page lookup process, nine bits of the virtual address are used per level to index into the corresponding page table.
 In cases where the input address is restricted to a maximum of 42 bits, the level 0 table can be omitted and translation can start with the level 1 table.
 In each table, an entry either points to the physical address of the next level page table, or alternatively can directly point to the base address of a memory block, completing the address translation prematurely.
-While regular operating systems may use the complete $\qty{4}{\kilo\byte}$ lookup procedure for maximum flexibility, in the controlled bare-metal case, where there is only one application, this is not necessary.
-For this reason, the developed kernel makes use of the first level page table and maps the complete \ac{dram} memory region using the $\qty{1}{\giga\byte}$ memory blocks.
+While regular operating systems may use the complete $\qty{4}{\kibi\byte}$ lookup procedure for maximum flexibility, in the controlled bare-metal case, where there is only one application, this is not necessary.
+For this reason, the developed kernel makes use of the first level page table and maps the complete \ac{dram} memory region using the $\qty{1}{\gibi\byte}$ memory blocks.
 In addition to the base pointer, each entry in the page table also holds certain attributes on how the memory region should be treated.
 To enable the mapping of the boot memory and \ac{io} devices such as \ac{uart}, the first memory blocks are marked with a non-cacheable attribute, followed by the normal \ac{dram} region, which is cacheable, and finally the \aca{fimdram} region, which is set to non-cacheable again.
 
-After setting up the page tables, initializing the \ac{tcr} to enable the $\qty{4}{\kilo\byte}$, and assigning the \ac{ttbr}, which holds the base pointer to the first level page table, the \ac{mmu} can be enabled, and the boot code can finally dispatch to the \texttt{main} function of the application.
+After setting up the page tables, initializing the \ac{tcr} to enable the $\qty{4}{\kibi\byte}$, and assigning the \ac{ttbr}, which holds the base pointer to the first level page table, the \ac{mmu} can be enabled, and the boot code can finally dispatch to the \texttt{main} function of the application.
 
 \subsubsection{Bare-Metal Utilities}
 % Heap Allocator (linked list allocator?...)

src/chapters/pim.tex

@@ -82,7 +82,7 @@ In the following, three \ac{pim} approaches that place the compute units at the
 
 The first publicly available real-world \ac{pim} architecture has been designed and built by the company UPMEM \cite{gomez-luna2022}.
 UPMEM combines regular DDR4 \ac{dimm} based \ac{dram} with a set of \ac{pim}-enabled UPMEM \acp{dimm} consisting of several \ac{pim} chips.
-In each \ac{pim} chip, there are of 8 \acp{dpu}, each of which has exclusive access to a $\qty{64}{\mega\byte}$ memory bank, a $\qty{24}{\kilo\byte}$ instruction memory and a $\qty{64}{\kilo\byte}$ scratchpad memory.
+In each \ac{pim} chip, there are of 8 \acp{dpu}, each of which has exclusive access to a $\qty{64}{\mebi\byte}$ memory bank, a $\qty{24}{\kibi\byte}$ instruction memory and a $\qty{64}{\kibi\byte}$ scratchpad memory.
 The host processor can access the \ac{dpu} memory banks to copy input data from main memory and retrieve results.
 While copying, the data layout must be changed to store the data words continuously in a \ac{pim} bank, in contrast to the horizontal \ac{dram} mapping used in \ac{dimm} modules, where a data word is split across multiple devices.
 UPMEM provides a \ac{sdk} that orchestrates the data movement from the main memory to the \ac{pim} banks and modifies the data layout without special attention of the developer.

src/chapters/results.tex

@@ -17,6 +17,7 @@ Thus, with both the 16-wide \ac{fp} adder and the 16-wide \ac{fp} multiplier, a
 In total, the 16 processing units in a memory channel provide a throughput of $\num{16}\cdot\qty{8}{\giga FLOPS}=\qty{128}{\giga FLOPS}$.
 To compare this throughput with the vector processing unit of a real processor, a very simplified assumption can be made based on the ARM NEON architecture, which holds 8 \ac{fp16} numbers in a single $\qty{128}{\bit}$ vector register \cite{arm2020}.
 Assuming the single processor core runs at a frequency of $\qty{3}{\giga\hertz}$, the vector processing unit can achieve a maximum throughput of $\qty{8}{FLOP} \cdot \qty{3}{\giga\hertz}=\qty{24}{\giga FLOPS}$, which is about $\qty{5}{\times}$ less than the \aca{fimdram} throughput of a single memory channel.
+The simulated ARM system also contains a two-level cache hierarchy with a cache size of $\qty{16}{\kibi\byte}$ for the L1 cache and $\qty{256}{\kibi\byte}$ for the L2 cache.
 
 % some implementation details
 % hbm size, channel...
@@ -92,9 +93,11 @@ The workloads adhere to the following calculation patterns:
 \item \ac{haxpy}: $z = a \cdot x + y$
 \end{itemize}
 
-Each workload is run with different input vector dimensions to examine the effect of setup overhead and potentially identify a break-even point at which \ac{pim} becomes viable.
+Each workload is run with four different input vector dimensions to examine the effect of setup overhead and potentially identify a break-even point at which \ac{pim} becomes viable.
 \Cref{tab:dimensions_vector} lists the specific vector dimensions for the following benchmarks.
-The levels X1-X4 denote the increasing dimensions, with each successive level doubling in size, starting at 256, which is the minimum size that can be represented in a \ac{pim} data structure.
+The levels X1-X4 denote the increasing dimensions, with each successive level doubling in size, starting at 2097152.
+To accurately evaluate the performance gain of \ac{pim}, it is important that the size of the input operand is significantly larger than the cache size of the simulated system, so that the cache does not filter the memory accesses to the \ac{dram}.
+In the case of the smallest dimension level, the effective data size of the input operands is $2^{21} \cdot 2 \cdot \qty{2}{\byte}=\qty{8}{\mebi\byte}$, which is much larger than the last-level cache of $\qty{256}{\kibi\byte}$.
 
 \begin{table}
 \centering
@@ -110,10 +113,10 @@ The levels X1-X4 denote the increasing dimensions, with each successive level do
   hline{2} = {2}{-}{solid,black},
 }
 Level & Vector Dimensions \\
-X1    & (256 $\times$ 1)  \\
-X2    & (512 $\times$ 1)  \\
-X3    & (1024 $\times$ 1) \\
-X4    & (2048 $\times$ 1) 
+X1    & $(2^{21})=(2 \textrm{M})$  \\
+X2    & $(2^{22})=(4 \textrm{M})$  \\
+X3    & $(2^{23})=(8 \textrm{M})$  \\
+X4    & $(2^{24})=(16 \textrm{M})$ 
 \end{tblr}
 \caption{List of the input vector dimensions for the vector benchmarks.}
 \label{tab:dimensions_vector}
@@ -131,10 +134,11 @@ S = \frac{\textrm{\#ticks in non-\ac{pim} mode}}{\textrm{\#ticks in \ac{pim} mod
     \label{fig:vector_normal}
 \end{figure}
 
-\Cref{fig:vector_normal} shows the relative performance for the vector benchmarks, running on the generic ARM-based system at a typical clock frequency.
-The relative speedup of \ac{pim} is in the range of about $\qtyrange{12.8}{31.8}{\times}$ with limited variance for each benchmark between the different vector dimensions, since such vector operations essentially scale linearly with the length of the input operands for both the non-\ac{pim} and \ac{pim} approaches.
-The \ac{haxpy} benchmark has the highest variance with a range of $\qtyrange{19.8}{31.8}{\times}$, which is due to the fact that each value of the one input vector must first be multiplied by a scalar amount on the \ac{cpu} before the addition operation, while in the \ac{pim} case the specialized \ac{mad} instruction is used.
-As all speedup values are well above 1, it can be concluded that even the smallest representable vector size of 256 is already above the break-even point at which \ac{pim} becomes viable.
+\Cref{fig:vector_normal} shows the relative performance for the vector benchmarks, running on the generic ARM-based system at a typical clock frequency of $\qty{3}{\giga\hertz}$.
+The relative speedup of \ac{pim} is in the range of about $\qtyrange{13.6}{23.9}{\times}$ with very small variance between the different vector dimensions, because such vector operations essentially scale linearly with the length of the input operands for both the non-\ac{pim} and \ac{pim} approaches.
+The \ac{haxpy} benchmark has the highest speedup compared to the VADD and VMUL benchmarks with up to $\qty{23.9}{\times}$.
+This is due to the fact that in the non-\ac{pim} system each value of the one input vector must first be multiplied by a scalar amount on the \ac{cpu} before the addition operation can take place, while in the \ac{pim} case the specialized \ac{mad} instruction is used, where these two operations are done in a single instruction.
+As all speedup values are well above 1, it can be concluded that even the smallest benchmarked vector size is already well above the break-even point at which \ac{pim} becomes viable.
 
 \begin{figure}
     \centering
@@ -144,13 +148,11 @@ As all speedup values are well above 1, it can be concluded that even the smalle
 \end{figure}
 
 In addition to the generic ARM-based system, the same benchmarks were run on the hypothetical infinite compute system, the results of which are shown in \cref{fig:vector_infinite}.
-As it can be seen, the achievable speedup in the completely memory-bounded system is with a range of $\qtyrange{1.7}{2.4}{\times}$ lower than in the generic system.
-The variance of the speedup between the different vector dimensions are also rather small.
-For the \ac{haxpy} benchmark, the smaller variance of $\qtyrange{2.0}{2.4}{\times}$ can be interpreted as follows:
-The additional computation step of the scalar multiplication does not affect the non-\ac{pim} system as much as in the previous case, because this is insignificant compared to the memory fetch of the vector elements.
-
-% vectors: im wesentlichen skaliert beides mit der länge es vecktors, minimal weniger overhead
-% haxpy: skalarmultiplikation macht CPU bedeutend langsamer, deswegen fällt dieser unterscheid bei 100GHz auch weg
+As it can be seen, the achievable speedup in the completely memory-bounded system is with a range of $\qtyrange{10.2}{17.6}{\times}$ lower than in the generic system.
+This is expected as the system becomes completely memory-bound and no longer relies on the relatively slow ARM processor.
+The variance in speedup between different vector dimensions is also fairly low.
+% For the \ac{haxpy} benchmark, the smaller variance of $\qtyrange{2.0}{2.4}{\times}$ can be interpreted as follows:
+% The additional computation step of the scalar multiplication does not affect the non-\ac{pim} system as much as in the previous case, because this is insignificant compared to the memory fetch of the vector elements.
 
 \subsubsection{Neural Network Layers}
 % GEMV
@@ -160,10 +162,10 @@ The additional computation step of the scalar multiplication does not affect the
 
 % GEMM mit stark interleavten matrizen (eher nicht)
 
-In addition to the vector operations and the level 1 \ac{blas} routine \ac{haxpy}, the performance improvement of \ac{pim} is also investigated for the level 2 \ac{blas} routine \ac{gemv}.
-Besides the regular \ac{gemv} operation, whose form is $y = A \cdot x$, several matrix-vector multiplications are chained together with the activation function \ac{relu} applied in between, modeling a simple fully connected neural network.
+In addition to the simple vector operations and the level 1 \ac{blas} routine \ac{haxpy}, the performance improvement of \ac{pim} is also investigated for the level 2 \ac{blas} routine \ac{gemv}.
+Besides the benchmark for the regular \ac{gemv} operation, whose form is $y = A \cdot x$, several matrix-vector multiplications are chained together with the activation function \ac{relu} applied in between, modeling a simple fully connected neural network in the \ac{dnn} benchmark.
 Each processing step for a \ac{dnn} layer can be described as $y = \textrm{ReLU}(A \cdot x)$, where the output of the operation is fed as input to the next layer.
-In the simplest form, quadratic matrix dimensions ensure that the output vector of each layer has the same dimensions as the input vector, which simplifies the chaining in the benchmark.
+In the simplest form, quadratic matrix dimensions ensure that the output vector of each layer has the same dimensions as the input vector, which simplifies the chaining of the outputs as inputs.
 Again, several different dimensions of the benchmark inputs are used, whose matrix dimensions for each of the two benchmarks are given in \cref{tab:dimensions_matrix}.
 
 \begin{table}
@@ -184,16 +186,17 @@ Again, several different dimensions of the benchmark inputs are used, whose matr
   hline{2} = {2}{-}{solid,black},
 }
 Level & \ac{gemv} Matrix Dimensions & \ac{dnn} Matrix Dimensions   \\
-X1    & (128 $\times$ 128)  & (128 $\times$ 128) \\
-X2    & (256 $\times$ 128)  & (256 $\times$ 256) \\
-X3    & (512 $\times$ 128)  & (512 $\times$ 512) \\
-X4    & (1024 $\times$ 128) & (1024 $\times$ 1024)
+X1    & (1024 $\times$ 4096)  & (256 $\times$ 256) \\
+X2    & (2048 $\times$ 4096)  & (512 $\times$ 512) \\
+X3    & (4096 $\times$ 8192)  & (1024 $\times$ 1024) \\
+X4    & (8192 $\times$ 8192) & (2048 $\times$ 2048)
 \end{tblr}
 \caption{List of the matrix dimensions for the neural network benchmarks.}
 \label{tab:dimensions_matrix}
 \end{table}
 
-In the \ac{gemv} benchmarks, only the number of rows is increased at each step, which means that the \ac{pim} microkernel has to perform more iterations of the \ac{mac} kernel, but does not have to load another chunk of the input vector, since it fits completely into the \ac{grf}-A registers.
+% In the \ac{gemv} benchmarks, only the number of rows is increased at each step, which means that the \ac{pim} microkernel has to perform more iterations of the \ac{mac} kernel, but does not have to load another chunk of the input vector, since it fits completely into the \ac{grf}-A registers.
+For the various \ac{gemv} benchmarks, both the number of rows and the number of columns are increased with each step, which means that the \ac{pim} microkernel has to not only perform more iterations of the \ac{mac} kernel to produce the partial sum, but also has to load different chunks of the input vector since it does not fit completely into the \acs{grf}-A registers.
 
 \begin{figure}[ht]
     \centering
@@ -202,12 +205,12 @@ In the \ac{gemv} benchmarks, only the number of rows is increased at each step,
     \label{fig:matrix_normal}
 \end{figure}
 
-\Cref{fig:matrix_normal} shows the relative performance for the \ac{gemv} benchmarks that are run on the system at a normal clock speed.
-The speedup for a single \ac{gemv} operation is in the range of $\qtyrange{3.5}{23.6}{\times}$ and for the simple \ac{dnn} layers $\qtyrange{3.0}{72.3}{\times}$.
-Unlike in the vector benchmarks, the performance gains become drastically more significant with increasing matrix dimensions, where \ac{pim} can exploit its specialized architecture for this type of operation.
+\Cref{fig:matrix_normal} shows the relative performance for the \ac{gemv} benchmarks that are run on the generic ARM system.
+The speedup for a single \ac{gemv} operation is in the range of $\qtyrange{56.0}{62.5}{\times}$ and for the simple \ac{dnn} layers in the range of $\qtyrange{10.7}{49.3}{\times}$.
+Unlike the vector benchmarks, the performance gains, especially for the \ac{dnn} benchmark, become drastically more significant with increasing matrix dimensions, where \ac{pim} can take advantage of its specialized architecture for this type of operation.
 A possible explanation is that the initial overhead of executing the microkernel in the \aca{fimdram} processing units quickly becomes insignificant with increasing operand dimensions compared to the actual execution time.
 Also, in all cases, the smallest representable operand dimensions already achieve a speedup of over one, suggesting that the break-even point of \ac{pim}'s viability for this system is below these dimensions.
-Since the speedup approaches $\qty{100}{\times}$ in the \ac{dnn} benchmark, it can be concluded that \ac{pim} offers an immense performance advantage in this system configuration.
+Since the speedup for \ac{gemv} approaches $\qty{63}{\times}$ and for \ac{dnn} $\qty{50}{\times}$, it can be concluded that \ac{pim} offers an extreme performance advantage in this system configuration.
 
 \begin{figure}
     \centering
@@ -216,21 +219,22 @@ Since the speedup approaches $\qty{100}{\times}$ in the \ac{dnn} benchmark, it c
     \label{fig:matrix_infinite}
 \end{figure}
 
-The \ac{gemv} and \ac{dnn} benchmarks, however show a more differentiated view for the infinite compute approach that models the completely memory-bounded system:
-For smaller matrix dimensions, the usage of \ac{pim} slows the execution down up to a factor of $\qty{0.21}{\times}$ for the \ac{gemv} benchmark and even $\qty{0.18}{\times}$ for the \ac{dnn} layers.
-However, the speedup quickly increases with the larger dimensions, reaches its break-even point at the third step and shows a maximum speedup of $\qty{4.7}{\times}$ and $\qty{6.1}{\times}$ for the \ac{gemv} and \ac{dnn} benchmark respectively.
+For the infinite compute approach, the \ac{gemv} and \ac{dnn} benchmarks however show a more differentiated view:
+While the \ac{gemv} benchmark plateaus at around $\qty{9}{\times}$ for all matrix sizes, the usage of \ac{pim} slows the execution down up to a factor of $\qty{0.56}{\times}$ for the \ac{dnn} benchmark.
+However, the speedup quickly increases with the matrix larger dimensions, reaches its break-even point at the second step and shows a maximum speedup of $\qty{9.2}{\times}$ and $\qty{6.0}{\times}$ for the \ac{gemv} and \ac{dnn} benchmarks respectively.
 These results provide a more realistic view of \aca{fimdram}:
-For workloads and accelerator systems that are truly memory-bound, performance improvements can be on the order of the simulated $\qty{6.1}{\times}$.
+For workloads and accelerator systems that are truly memory-bound, performance improvements can be on the order of the simulated $\qty{9}{\times}$.
 This result is largely in line with the numbers published by Samsung, which were already introduced in \cref{sec:fimdram_performance} and will be compared in more detail with the simulation results in the next section.
 
 \subsubsection{Comparison to Samsung's Simulation Results}
 
 To reiterate, Samsung used a real hardware accelerator platform for its analyses, which is based on a Xilinx Zynq Ultrascale+ \ac{fpga} and uses real manufactured \aca{fimdram} memory packages.
-Similarly to the above investigations, Samsung used for its microbenchmarks different input dimensions for both its \ac{gemv} and vector ADD workloads, which are listed in \cref{tab:samsung_dimensions}.
+Similarly to the previous investigations, Samsung used for its microbenchmarks different input dimensions for both its \ac{gemv} and vector ADD workloads, which are listed in \cref{tab:samsung_dimensions}.
 
 \begin{table}
 \centering
 \begin{tblr}{
+  row{1} = {c},
   cell{2}{2} = {r},
   cell{3}{2} = {r},
   cell{4}{2} = {r},
@@ -254,9 +258,12 @@ Level 4  & (8k $\times$ 8k) & (16M)
 \label{tab:samsung_dimensions}
 \end{table}
 
-Each simulation is run with different batch sizes, where a higher batch size allows for better cache utilization, as multiple operations are performed on the same data set, making the workload less memory bound and rendering \ac{pim} less effective.
-All the microbenchmarks discussed so far do not perform batching, so all comparisons are performed on the result values for the batch size of 1, which correspond with the blue bars in \cref{fig:samsung_speedup}.
-Since the Samsung \ac{fpga} platform can be assumed to be a highly optimized accelerator, the infinite compute approach would be a more viable baseline for comparison than the limited \ac{cpu} approach, as both systems should operate in the memory-bounded region.
+As can be seen, the dimensions for the \ac{gemv} benchmark and the vector add operations, which corresponds to the VADD benchmark of this thesis, match the dimensions used in the previously discussed simulations.
+Therefore, the simulations can be directly compared to gain a good understanding of how accurate they are in comparison to the real system manufactured by Samsung.
+
+Each of Samsung's benchmarks is run with different batch sizes, where a larger batch size allows for better cache utilization as multiple operations are performed on the same data set, making the workload less memory-bound and therefore \ac{pim} less effective.
+All the microbenchmarks discussed so far do not perform batching, so all comparisons are made against the results for the batch size of 1, which correspond to the blue bars in \cref{fig:samsung_speedup}.
+Since the Samsung \ac{fpga} platform can be assumed to be a highly optimized accelerator, the infinite compute approach would be a more viable baseline for comparison than the \ac{cpu} approach, as both systems should be operating in the memory-bounded region.
 
 \begin{figure}
     \centering
@@ -266,22 +273,27 @@ Since the Samsung \ac{fpga} platform can be assumed to be a highly optimized acc
 \end{figure}
 
 The performed ADD microbenchmark of Samsung show a small variance between the different input dimensions with an average speedup value of around $\qty{1.6}{\times}$.
-When compared to the simulated platform, the variance is also limited with a range of $\qtyrange{1.6}{2.4}{\times}$, which corresponds well with the findings of Samsung.
-The \ac{gemv} microbenchmark on the other hand shows a more drastic speedup with an average value of $\qty{8.3}{\times}$.
-Although the dimensions used by Samsung are different from the simulations of this thesis, the highest achieved speedup of $\qty{6.1}{\times}$ is well within the reach of the real hardware implementation.
+When compared to the simulated platform, the variance is also limited with a value of around $\qty{12.7}{\times}$, which almost an order of magnitude higher than the findings of Samsung.
+This may be a surprising result, since such vector operations are inherently memory-bound and should be a prime candidate for the use of \ac{pim}.
+Samsung explains its low value of $\qty{1.6}{\times}$ by the fact that after 8 \ac{rd} accesses, the processor has to introduce a memory barrier instruction, resulting in a severe performance hit \cite{lee2021}.
+However, this memory barrier has also been implemented in the VADD kernel of the simulations, which still show a significant performance gain.
+
+The \ac{gemv} microbenchmark on the other hand shows a more matching result with an average speedup value of $\qty{8.3}{\times}$ for Samsung's implementation, while the simulation of this thesis achieved an average speedup of $\qty{9.0}{\times}$ which is well within the reach of the real hardware implementation.
+In summary, the results for the VADD workload show some deviation from the real-world implementation of the system, while the \ac{gemv} workload shows a result that is consistent with it.
 
 \subsubsection{Comparison to Real Hardware}
-TODO: check all ranges
 
-In addition to the comparison of Samsung's real hardware implementation, the same benchmarks of the performed simulations are run on a [...] with HBM2 [...].
-As this system is using a generic \aca{hbm} \ac{dram} and not \aca{fimdram}, the measurements are only intended to serve as a vague estimation of the runtimes in a non-\ac{pim} case.
+In addition to comparing Samsung's real hardware implementation, the same benchmarks of the simulations performed are run on two real \ac{gpu} systems, here referred to as Vega and Tesla.
+The former system is the consumer \ac{gpu} \textit{Radeon RX Vega 56} from AMD, while the latter is the \textit{Tesla V100} \ac{gpu} from Nvidia, specifically tailored for \ac{hpc}.
+Both \acp{gpu} make use of \aca{hbm} and therefore are greatly suited to classify the simulation results and get an overview of the workload runtimes on a real system.
+As both systems are using generic \aca{hbm} \ac{dram} and not \aca{fimdram}, the measurements are only intended to serve as a vague estimation of the runtimes in a non-\ac{pim} case.
 
 \begin{figure}
     \centering
     \resizebox{\linewidth}{!}{%
     \input{plots/runtimes_vector}
     }
-    \caption{}
+    \caption{Runtimes of all investigated systems for the vector benchmarks.}
     \label{fig:runtimes_vector}
 \end{figure}
 
@@ -290,7 +302,7 @@ As this system is using a generic \aca{hbm} \ac{dram} and not \aca{fimdram}, the
     % \resizebox{\linewidth}{!}{%
     \input{plots/runtimes_matrix}
     % }
-    \caption{}
+    \caption{Runtimes of all investigated systems for the matrix benchmarks.}
     \label{fig:runtimes_matrix}
 \end{figure}
 

src/doc.bib

@@ -12,8 +12,8 @@
   editor = {Cuesta, Carlos E. and Garlan, David and Pérez, Jennifer},
   date = {2018},
   pages = {115--130},
-  publisher = {{Springer International Publishing}},
-  location = {{Cham}},
+  publisher = {Springer International Publishing},
+  location = {Cham},
   abstract = {Continuous software engineering aims at orchestrating engineering knowledge from various disciplines in order to deal with the rapid changes within the ecosystems of which software-based systems are part of. The literature claims that one means to ensure these prompt responses is to incorporate virtual prototypes of the system as early as possible in the development process, such that requirements and architecture decisions are verified early and continuously by means of simulations. Despite the maturity of practices for designing and assessing architectures, as well as for virtual prototyping, it is still not clear how to jointly consider the practices from these disciplines within development processes, in order to address the dynamics imposed by continuous software engineering. In this regard, we discuss in this paper how to orchestrate architecture drivers and design specification techniques with virtual prototypes, to address the demands of continuous software engineering in development processes. Our proposals are based on experiences from research and industry projects in various domains such as automotive, agriculture, construction, and medical devices.},
   isbn = {978-3-030-00761-4},
   file = {/home/derek/Nextcloud/Verschiedenes/Zotero/storage/KGD8N29E/Antonino et al. - 2018 - Enabling Continuous Software Engineering for Embed.pdf}
@@ -53,7 +53,7 @@
   date = {2023-11-10},
   url = {https://dvcon-europe.org/wp-content/uploads/sites/14/2023/12/Keynote-Pervasive-and-Sustainable-AI-with-Adaptive.pdf},
   urldate = {2024-01-23},
-  venue = {{DVCon 2023}}
+  venue = {DVCon 2023}
 }
 
 @online{chen2023,
@@ -168,8 +168,8 @@
   author = {He, Mingxuan and Song, Choungki and Kim, Ilkon and Jeong, Chunseok and Kim, Seho and Park, Il and Thottethodi, Mithuna and Vijaykumar, T. N.},
   date = {2020-10},
   pages = {372--385},
-  publisher = {{IEEE}},
-  location = {{Athens, Greece}},
+  publisher = {IEEE},
+  location = {Athens, Greece},
   doi = {10.1109/MICRO50266.2020.00040},
   url = {https://ieeexplore.ieee.org/document/9251855/},
   urldate = {2024-01-09},
@@ -191,8 +191,8 @@
   shorttitle = {Memory Systems},
   author = {Jacob, Bruce and Ng, Spencer W. and Wang, David T. and Wang, David and Rodriguez, Samuel},
   date = {2008},
-  publisher = {{Elsevier/Morgan Kaufmann}},
-  location = {{Amsterdam Heidelberg}},
+  publisher = {Elsevier/Morgan Kaufmann},
+  location = {Amsterdam Heidelberg},
   isbn = {978-0-12-379751-3},
   langid = {english},
   pagetotal = {982},
@@ -219,8 +219,8 @@
   author = {Jouppi, Norman P. and Young, Cliff and Patil, Nishant and Patterson, David and Agrawal, Gaurav and Bajwa, Raminder and Bates, Sarah and Bhatia, Suresh and Boden, Nan and Borchers, Al and Boyle, Rick and Cantin, Pierre-luc and Chao, Clifford and Clark, Chris and Coriell, Jeremy and Daley, Mike and Dau, Matt and Dean, Jeffrey and Gelb, Ben and Ghaemmaghami, Tara Vazir and Gottipati, Rajendra and Gulland, William and Hagmann, Robert and Ho, C. Richard and Hogberg, Doug and Hu, John and Hundt, Robert and Hurt, Dan and Ibarz, Julian and Jaffey, Aaron and Jaworski, Alek and Kaplan, Alexander and Khaitan, Harshit and Killebrew, Daniel and Koch, Andy and Kumar, Naveen and Lacy, Steve and Laudon, James and Law, James and Le, Diemthu and Leary, Chris and Liu, Zhuyuan and Lucke, Kyle and Lundin, Alan and MacKean, Gordon and Maggiore, Adriana and Mahony, Maire and Miller, Kieran and Nagarajan, Rahul and Narayanaswami, Ravi and Ni, Ray and Nix, Kathy and Norrie, Thomas and Omernick, Mark and Penukonda, Narayana and Phelps, Andy and Ross, Jonathan and Ross, Matt and Salek, Amir and Samadiani, Emad and Severn, Chris and Sizikov, Gregory and Snelham, Matthew and Souter, Jed and Steinberg, Dan and Swing, Andy and Tan, Mercedes and Thorson, Gregory and Tian, Bo and Toma, Horia and Tuttle, Erick and Vasudevan, Vijay and Walter, Richard and Wang, Walter and Wilcox, Eric and Yoon, Doe Hyun},
   date = {2017-06-24},
   pages = {1--12},
-  publisher = {{ACM}},
-  location = {{Toronto ON Canada}},
+  publisher = {ACM},
+  location = {Toronto ON Canada},
   doi = {10.1145/3079856.3080246},
   url = {https://dl.acm.org/doi/10.1145/3079856.3080246},
   urldate = {2024-01-22},
@@ -235,7 +235,7 @@
   author = {Jung, Matthias},
   date = {2017},
   series = {Forschungsberichte {{Mikroelektronik}}},
-  publisher = {{Technische Universität Kaiserslautern}},
+  publisher = {Technische Universität Kaiserslautern},
   isbn = {978-3-95974-051-7},
   file = {/home/derek/Nextcloud/Verschiedenes/Zotero/storage/Y9YSTV6C/Jung - 2017 - System-level Modeling, Analysis and Optimization o.pdf}
 }
@@ -263,8 +263,8 @@
   author = {Kal, Hongju and Yoo, Chanyoung and Ro, Won Woo},
   date = {2023-10-28},
   pages = {815--827},
-  publisher = {{ACM}},
-  location = {{Toronto ON Canada}},
+  publisher = {ACM},
+  location = {Toronto ON Canada},
   doi = {10.1145/3613424.3614314},
   url = {https://dl.acm.org/doi/10.1145/3613424.3614314},
   urldate = {2024-01-08},
@@ -282,8 +282,8 @@
   author = {Kang, Shinhaeng and Lee, Sukhan and Kim, Byeongho and Kim, Hweesoo and Sohn, Kyomin and Kim, Nam Sung and Lee, Eojin},
   date = {2022-02-13},
   pages = {146--152},
-  publisher = {{ACM}},
-  location = {{Virtual Event USA}},
+  publisher = {ACM},
+  location = {Virtual Event USA},
   doi = {10.1145/3490422.3502355},
   url = {https://dl.acm.org/doi/10.1145/3490422.3502355},
   urldate = {2024-01-08},
@@ -301,8 +301,8 @@
   author = {Kwon, Young-Cheon and Lee, Suk Han and Lee, Jaehoon and Kwon, Sang-Hyuk and Ryu, Je Min and Son, Jong-Pil and Seongil, O and Yu, Hak-Soo and Lee, Haesuk and Kim, Soo Young and Cho, Youngmin and Kim, Jin Guk and Choi, Jongyoon and Shin, Hyun-Sung and Kim, Jin and Phuah, BengSeng and Kim, HyoungMin and Song, Myeong Jun and Choi, Ahn and Kim, Daeho and Kim, SooYoung and Kim, Eun-Bong and Wang, David and Kang, Shinhaeng and Ro, Yuhwan and Seo, Seungwoo and Song, JoonHo and Youn, Jaeyoun and Sohn, Kyomin and Kim, Nam Sung},
   date = {2021-02-13},
   pages = {350--352},
-  publisher = {{IEEE}},
-  location = {{San Francisco, CA, USA}},
+  publisher = {IEEE},
+  location = {San Francisco, CA, USA},
   doi = {10.1109/ISSCC42613.2021.9365862},
   url = {https://ieeexplore.ieee.org/document/9365862/},
   urldate = {2024-01-08},
@@ -319,8 +319,8 @@
   author = {Kwon, Yongkee and Vladimir, Kornijcuk and Kim, Nahsung and Shin, Woojae and Won, Jongsoon and Lee, Minkyu and Joo, Hyunha and Choi, Haerang and Kim, Guhyun and An, Byeongju and Kim, Jeongbin and Lee, Jaewook and Kim, Ilkon and Park, Jaehan and Park, Chanwook and Song, Yosub and Yang, Byeongsu and Lee, Hyungdeok and Kim, Seho and Kwon, Daehan and Lee, Seongju and Kim, Kyuyoung and Oh, Sanghoon and Park, Joonhong and Hong, Gimoon and Ka, Dongyoon and Hwang, Kyudong and Park, Jeongje and Kang, Kyeongpil and Kim, Jungyeon and Jeon, Junyeol and Lee, Myeongjun and Shin, Minyoung and Shin, Minhwan and Cha, Jaekyung and Jung, Changson and Chang, Kijoon and Jeong, Chunseok and Lim, Euicheol and Park, Il and Chun, Junhyun and Hynix, Sk},
   date = {2022-08-21},
   pages = {1--25},
-  publisher = {{IEEE}},
-  location = {{Cupertino, CA, USA}},
+  publisher = {IEEE},
+  location = {Cupertino, CA, USA},
   doi = {10.1109/HCS55958.2022.9895629},
   url = {https://ieeexplore.ieee.org/document/9895629/},
   urldate = {2024-01-22},
@@ -336,8 +336,8 @@
   author = {Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph and Zhang, Hao and Stoica, Ion},
   date = {2023-10-23},
   pages = {611--626},
-  publisher = {{ACM}},
-  location = {{Koblenz Germany}},
+  publisher = {ACM},
+  location = {Koblenz Germany},
   doi = {10.1145/3600006.3613165},
   url = {https://dl.acm.org/doi/10.1145/3600006.3613165},
   urldate = {2024-01-12},
@@ -354,8 +354,8 @@
   author = {Lee, Sukhan and Kang, Shin-haeng and Lee, Jaehoon and Kim, Hyeonsu and Lee, Eojin and Seo, Seungwoo and Yoon, Hosang and Lee, Seungwon and Lim, Kyounghwan and Shin, Hyunsung and Kim, Jinhyun and Seongil, O and Iyer, Anand and Wang, David and Sohn, Kyomin and Kim, Nam Sung},
   date = {2021-06},
   pages = {43--56},
-  publisher = {{IEEE}},
-  location = {{Valencia, Spain}},
+  publisher = {IEEE},
+  location = {Valencia, Spain},
   doi = {10.1109/ISCA52012.2021.00013},
   url = {https://ieeexplore.ieee.org/document/9499894/},
   urldate = {2024-01-08},
@@ -426,7 +426,7 @@
   title = {Neural Networks and Deep Learning},
   author = {Nielsen, Michael A.},
   date = {2015},
-  publisher = {{Determination Press}},
+  publisher = {Determination Press},
   url = {http://neuralnetworksanddeeplearning.com/},
   file = {/home/derek/Nextcloud/Verschiedenes/Zotero/storage/E6FRVMZ3/Nielsen - 2015 - Neural networks and deep learning.pdf}
 }
@@ -460,7 +460,7 @@
   shorttitle = {Processing in {{Memory}}},
   author = {Radojković, Petar and Carpenter, Paul and Esmaili-Dokht, Pouya and Cimadomo, Rémy and Charles, Henri-Pierre and Sebastian, Abu and Amato, Paolo},
   date = {2021-07-29},
-  institution = {{Zenodo}},
+  institution = {Zenodo},
   doi = {10.5281/ZENODO.4767489},
   url = {https://zenodo.org/record/4767489},
   urldate = {2024-02-06},
@@ -493,8 +493,8 @@
   author = {Samajdar, Ananda and Joseph, Jan Moritz and Zhu, Yuhao and Whatmough, Paul and Mattina, Matthew and Krishna, Tushar},
   date = {2020-08},
   pages = {58--68},
-  publisher = {{IEEE}},
-  location = {{Boston, MA, USA}},
+  publisher = {IEEE},
+  location = {Boston, MA, USA},
   doi = {10.1109/ISPASS48437.2020.00016},
   url = {https://ieeexplore.ieee.org/document/9238602/},
   urldate = {2024-02-14},
@@ -512,8 +512,8 @@
   author = {Seshadri, Vivek and Kim, Yoongu and Fallin, Chris and Lee, Donghyuk and Ausavarungnirun, Rachata and Pekhimenko, Gennady and Luo, Yixin and Mutlu, Onur and Gibbons, Phillip B. and Kozuch, Michael A. and Mowry, Todd C.},
   date = {2013-12-07},
   pages = {185--197},
-  publisher = {{ACM}},
-  location = {{Davis California}},
+  publisher = {ACM},
+  location = {Davis California},
   doi = {10.1145/2540708.2540725},
   url = {https://dl.acm.org/doi/10.1145/2540708.2540725},
   urldate = {2024-02-05},
@@ -584,8 +584,8 @@
   date = {2022},
   volume = {13511},
   pages = {362--379},
-  publisher = {{Springer International Publishing}},
-  location = {{Cham}},
+  publisher = {Springer International Publishing},
+  location = {Cham},
   doi = {10.1007/978-3-031-15074-6_23},
   url = {https://link.springer.com/10.1007/978-3-031-15074-6_23},
   urldate = {2024-01-21},
@@ -615,8 +615,8 @@
 @book{systemc2023,
   title = {1666-2023 - {{IEEE Standard}} for {{Standard SystemC Language Reference Manual}}},
   date = {2023},
-  publisher = {{IEEE}},
-  location = {{New York}},
+  publisher = {IEEE},
+  location = {New York},
   abstract = {SystemC® is defined in this standard. SystemC is an ISO standard C++ class library for system and hardware design for use by designers and architects who need to address complex systems that are a hybrid between hardware and software. This standard provides a precise and complete definition of the SystemC class library so that a SystemC implementation can be developed with reference to this standard alone. The primary audiences for this standard are the implementors of the SystemC class library, the implementors of tools supporting the class library, and the users of the class library},
   isbn = {978-1-5044-9867-8},
   langid = {english},
@@ -624,6 +624,14 @@
   file = {/home/derek/Nextcloud/Verschiedenes/Zotero/storage/46IIZIMH/2023 - 1666-2023 - IEEE Standard for Standard SystemC Lan.pdf}
 }
 
+@online{tesla2018,
+  title = {{{NVIDIA Tesla V100 PCIe}} 32 {{GB Specs}}},
+  author = {{techpowerup.com}},
+  date = {2018},
+  url = {https://www.techpowerup.com/gpu-specs/tesla-v100-pcie-32-gb.c3184},
+  urldate = {2024-03-07}
+}
+
 @online{touvron2023,
   title = {{{LLaMA}}: {{Open}} and {{Efficient Foundation Language Models}}},
   shorttitle = {{{LLaMA}}},
@@ -639,6 +647,14 @@
   file = {/home/derek/Nextcloud/Verschiedenes/Zotero/storage/MGQYNDPQ/Touvron et al. - 2023 - LLaMA Open and Efficient Foundation Language Mode.pdf;/home/derek/Nextcloud/Verschiedenes/Zotero/storage/YDAT8K7L/2302.html}
 }
 
+@online{vega2017,
+  title = {{{AMD Radeon RX Vega}} 56 {{Specs}}},
+  author = {{techpowerup.com}},
+  date = {2017},
+  url = {https://www.techpowerup.com/gpu-specs/radeon-rx-vega-56.c2993},
+  urldate = {2024-03-07}
+}
+
 @article{zou2021,
   title = {Breaking the von {{Neumann}} Bottleneck: Architecture-Level Processing-in-Memory Technology},
   shorttitle = {Breaking the von {{Neumann}} Bottleneck},

src/index.tex

@@ -85,13 +85,13 @@
 \setcounter{page}{1}
 
 % Chapters
-% \include{chapters/introduction}
-% \include{chapters/dram}
-% \include{chapters/pim}
-% \include{chapters/vp}
-% \include{chapters/implementation}
+\include{chapters/introduction}
+\include{chapters/dram}
+\include{chapters/pim}
+\include{chapters/vp}
+\include{chapters/implementation}
 \include{chapters/results}
-% \include{chapters/conclusion}
+\include{chapters/conclusion}
 
 % Appendix
 \appendix

src/plots/matrix_infinite.tex

@@ -4,7 +4,7 @@
             width=0.9\textwidth,
             ybar=1pt,
             bar width = 15pt,
-            ymin=0.1,
+            ymin=0,
             ymax=10,
             ymajorgrids,
             ylabel={Relative Performance},

src/plots/matrix_normal.tex

@@ -4,8 +4,8 @@
             width=0.9\textwidth,
             ybar=1pt,
             bar width = 15pt,
-            ymin=0.1,
-            ymax=75,
+            ymin=0,
+            ymax=80,
             ymajorgrids,
             ylabel={Relative Performance},
             tick pos=left,

src/plots/runtime_tables/pim_100GHz.csv

@@ -1,5 +1,5 @@
 level,vadd,vmul,haxpy,gemv,dnn
-X1,911446480,911416480,954454480,951904860,536177760
-X2,1822806480,1822776480,1908822480,1814530860,738329760
-X3,3645526480,3645496480,3817558480,6990944860,1547139760
-X4,7290966480,7290936480,7635030480,13892610860,4782339760
+X1,911403240,911388240,954411240,907312430,301697880
+X2,1822763240,1822748240,1908779240,1769985430,504205880
+X3,3645483240,3645468240,3817515240,6946352430,1312770880
+X4,7290923240,7290908240,7634987240,13848065430,4547969880

src/plots/runtime_tables/pim_3GHz.csv

@@ -1,5 +1,5 @@
 level,vadd,vmul,haxpy,gemv,dnn
-X1,1475510346,1475512344,1543044078,1377734886,933823908
-X2,2950962084,2950964082,3085992252,2601142920,1220409702
-X3,5901852240,5901848244,6171893928,9942655392,2367353610
-X4,11803639878,11803641876,12343693950,19731271644,6955629408
+X1,1475478045,1475481042,1542998124,1300650714,514715103
+X2,2950928118,2950925121,3085945965,2524123683,801373158
+X3,5901817275,5901817941,6171846975,9865572552,1948383000
+X4,11803603914,11803603914,12343646997,19654106886,6536710746

src/plots/runtimes_matrix.tex

@@ -49,9 +49,9 @@
     \addplot[fill=_orange!90]   table [x expr=\coordindex, y={gemv}]{\hbmpim};
     \addlegendentry{PIM ARM}
     \addplot[fill=_yellow!90]   table [x expr=\coordindex, y={gemv}]{\hbminf};
-    \addlegendentry{Non-PIM Inf}
+    \addlegendentry{Non-PIM Inf.}
     \addplot[fill=_green!90]    table [x expr=\coordindex, y={gemv}]{\piminf};
-    \addlegendentry{PIM Inf}
+    \addlegendentry{PIM Inf.}
     \addplot[fill=_darkblue!90] table [x expr=\coordindex, y={gemv}]{\vega};
     \addlegendentry{Vega}
     \addplot[fill=violet!90]    table [x expr=\coordindex, y={gemv}]{\tesla};

src/plots/runtimes_vector.tex

@@ -62,9 +62,9 @@
     \addplot[fill=_orange!90]   table [x expr=\coordindex, y={vmul}]{\hbmpim};
     \addlegendentry{PIM ARM}
     \addplot[fill=_yellow!90]   table [x expr=\coordindex, y={vmul}]{\hbminf};
-    \addlegendentry{Non-PIM Inf}
+    \addlegendentry{Non-PIM Inf.}
     \addplot[fill=_green!90]    table [x expr=\coordindex, y={vmul}]{\piminf};
-    \addlegendentry{PIM Inf}
+    \addlegendentry{PIM Inf.}
     \addplot[fill=_darkblue!90] table [x expr=\coordindex, y={vmul}]{\vega};
     \addlegendentry{Vega}
     \addplot[fill=violet!90]    table [x expr=\coordindex, y={vmul}]{\tesla};

src/plots/speedup_tables/matrix_100GHz.csv

@@ -1,5 +1,5 @@
 level,gemv,dnn
-X1,8.316378361593825,0.3293391169376365
-X2,8.707496426927674,2.520540889480061
-X3,8.952627753954278,4.565606038073768
-X4,9.178586247394538,5.717860064798073
+X1,8.725110246753701,0.5853017926410354
+X2,8.926639006288317,3.6909334536122427
+X3,9.010099560986427,5.380703318160134
+X4,9.208111243015697,6.012517728019782

src/plots/speedup_tables/matrix_3GHz.csv

@@ -1,5 +1,5 @@
 level,gemv,dnn
-X1,52.86562483836241,5.894128466670185
-X2,58.14176507187079,17.599027693570402
-X3,61.79055586904123,36.290342473171975
-X4,62.23855088820042,46.32048031346181
+X1,55.99875110667106,10.693445843962344
+X2,59.9158597463229,26.801526765137798
+X3,62.27334503373079,44.09403760041019
+X4,62.48290809788771,49.2890243396736

src/plots/speedup_tables/vector_100GHz.csv

@@ -1,5 +1,5 @@
 level,vadd,vmul,haxpy
-X1,12.912332482758615,10.706896577073085,17.57261802574388
-X2,12.656964545133722,10.410011429377231,17.530374532261376
-X3,12.857948841452387,10.179649930700249,17.28682620992881
-X4,12.517518527941442,10.158700762676236,17.568276160961705
+X1,12.912945086743383,10.707228337727948,17.57341416054572
+X2,12.657264796496554,10.41017271260676,17.530771651728568
+X3,12.858101352840125,10.179728788420332,17.287022013303083
+X4,12.5175927651105,10.158740110546228,17.568375657167437

src/plots/speedup_tables/vector_3GHz.csv

@@ -1,5 +1,5 @@
 level,vadd,vmul,haxpy
-X1,14.459220593631812,13.63052633194372,23.727849658355645
-X2,14.66245011706494,13.551652528382078,23.822458866951166
-X3,14.484534634672588,13.652770583844921,23.665607976080565
-X4,14.143299569834605,13.81101374775393,23.87690729103017
+X1,14.45953713326856,13.630815500508477,23.72855632778462
+X2,14.662618885927062,13.55183145055479,23.822816186932165
+X3,14.484620447521396,13.652840684263325,23.665788014778187
+X4,14.143342662573861,13.811058165857734,23.876998114465763

src/tables/simulations_100GHz.csv

@@ -1,21 +1,21 @@
 workload,level,hbm,pim
-VADD,X1,11768899990,455723240
-VADD,X2,23071196990,911403240
-VADD,X3,46873992980,1822763240
-VADD,X4,91264808000,3645483240
-VMUL,X1,9758441990,455708240
-VMUL,X2,18975123990,911388240
-VMUL,X3,37109877990,1822748240
-VMUL,X4,74066441980,3645468240
-HAXPY,X1,16772264000,477227240
-HAXPY,X2,33462372990,954411240
-HAXPY,X3,65993469990,1908779240
-HAXPY,X4,134134323970,3817515240
-GEMV,X1,7916400980,475952430
-GEMV,X2,15800020980,907265430
-GEMV,X3,62587326980,3495472430
-GEMV,X4,127514526980,6946305430
-DNN,X1,176584310,268088880
-DNN,X2,1860990350,369164880
-DNN,X3,7063630630,773569880
-DNN,X4,27344749530,2391169880
+VADD,X1,11768899990,911403240
+VADD,X2,23071196990,1822763240
+VADD,X3,46873992980,3645483240
+VADD,X4,91264808000,7290923240
+VMUL,X1,9758441990,911388240
+VMUL,X2,18975123990,1822748240
+VMUL,X3,37109877990,3645468240
+VMUL,X4,74066441980,7290908240
+HAXPY,X1,16772264000,954411240
+HAXPY,X2,33462372990,1908779240
+HAXPY,X3,65993469990,3817515240
+HAXPY,X4,134134323970,7634987240
+GEMV,X1,7916400980,907312430
+GEMV,X2,15800020980,1769985430
+GEMV,X3,62587326980,6946352430
+GEMV,X4,127514526980,13848065430
+DNN,X1,176584310,301697880
+DNN,X2,1860990350,504205880
+DNN,X3,7063630630,1312770880
+DNN,X4,27344749530,4547969880

src/tables/simulations_3GHz.csv

@@ -1,21 +1,21 @@
 workload,level,hbm,pim
-VADD,X1,21334729581,737755173
-VADD,X2,43268334354,1475481042
-VADD,X3,85485583179,2950926120
-VADD,X4,166942414809,5901819939
-VMUL,X1,20112009858,737756172
-VMUL,X2,39990439863,1475482041
-VMUL,X3,80576580096,2950924122
-VMUL,X4,163020260223,5901820938
-HAXPY,X1,36613117899,771522039
-HAXPY,X2,73515923487,1542996126
-HAXPY,X3,146061622170,3085946964
-HAXPY,X4,294729236073,6171846975
-GEMV,X1,72834815610,688867443
-GEMV,X2,151235040573,1300571460
-GEMV,X3,614362203486,4971327696
-GEMV,X4,1228045754304,9865635822
-DNN,X1,5504078079,466911954
-DNN,X2,21478024143,610204851
-DNN,X3,85912073262,1183676805
-DNN,X4,322188095061,3477814704
+VADD,X1,21334729581,1475478045
+VADD,X2,43268334354,2950928118
+VADD,X3,85485583179,5901817275
+VADD,X4,166942414809,11803603914
+VMUL,X1,20112009858,1475481042
+VMUL,X2,39990439863,2950925121
+VMUL,X3,80576580096,5901817941
+VMUL,X4,163020260223,11803603914
+HAXPY,X1,36613117899,1542998124
+HAXPY,X2,73515923487,3085945965
+HAXPY,X3,146061622170,6171846975
+HAXPY,X4,294729236073,12343646997
+GEMV,X1,72834815610,1300650714
+GEMV,X2,151235040573,2524123683
+GEMV,X3,614362203486,9865572552
+GEMV,X4,1228045754304,19654106886
+DNN,X1,5504078079,514715103
+DNN,X2,21478024143,801373158
+DNN,X3,85912073262,1948383000
+DNN,X4,322188095061,6536710746

src/tables/torch.csv

@@ -1,21 +1,21 @@
 workload,level,vega,tesla
-VADD,X1,69572650,TODO
-VADD,X2,123217536,TODO
-VADD,X3,207693503,TODO
-VADD,X4,378089165,TODO
-VMUL,X1,67408281,TODO
-VMUL,X2,103994272,TODO
-VMUL,X3,182162140,TODO
-VMUL,X4,350280326,TODO
-HAXPY,X1,69791189,TODO
-HAXPY,X2,123543145,TODO
-HAXPY,X3,207947543,TODO
-HAXPY,X4,377434890,TODO
-GEMV,X1,750246152,TODO
-GEMV,X2,648714601,TODO
-GEMV,X3,2454455479,TODO
-GEMV,X4,4968984949,TODO
-DNN,X1,231093065,TODO
-DNN,X2,431703456,TODO
-DNN,X3,877622611,TODO
-DNN,X4,2175751385,TODO
+VADD,X1,69572650,69572650
+VADD,X2,123217536,123217536
+VADD,X3,207693503,207693503
+VADD,X4,378089165,378089165
+VMUL,X1,67408281,67408281
+VMUL,X2,103994272,103994272
+VMUL,X3,182162140,182162140
+VMUL,X4,350280326,350280326
+HAXPY,X1,69791189,69791189
+HAXPY,X2,123543145,123543145
+HAXPY,X3,207947543,207947543
+HAXPY,X4,377434890,377434890
+GEMV,X1,750246152,750246152
+GEMV,X2,648714601,648714601
+GEMV,X3,2454455479,2454455479
+GEMV,X4,4968984949,4968984949
+DNN,X1,231093065,231093065
+DNN,X2,431703456,431703456
+DNN,X3,877622611,877622611
+DNN,X4,2175751385,2175751385