March 2022 - cypherpunks - lists.cpunks.org

Bad news, Julio
by professor rat 27 Mar '22

27 Mar '22

“It’s disturbing to see $18–27 billion of Australian government money about to be spent on more than 400 heavily armoured ‘infantry fighting vehicles’ when we’re busily watching yet another conflict in which military vehicles like these are being destroyed in numbers by cheap and readily available anti-armour missiles and armed drones,” https://www.news.com.au/technology/innovation/military/ukraine-has-more-tan…

1 0

NYCCoin on the Week
by Gunnar Larson 27 Mar '22

27 Mar '22

NYCCOIN PRICE CHART $0.0033 +$0.00 (+5.59%) The total contribution to the New York City Wallet is valued at $20,083,805.01 (14,776,028.47 STX): https://mining.nyc/wallet -- *Gunnar Larson - xNY.io <http://www.xNY.io> | Bank.org <http://Bank.org>* MSc <https://www.unic.ac.cy/blockchain/msc-digital-currency/?utm_source=Google&u…> - Digital Currency MBA <https://www.unic.ac.cy/business-administration-entrepreneurship-and-innovat…> - Entrepreneurship and Innovation (ip) G(a)xNY.io +1-646-454-9107 New York, New York 10001

1 0

P2P: Lantern App, Yinbi, etc Circumvention Russia
by grarpamp 27 Mar '22

27 Mar '22

https://www.vice.com/en/article/z3n5e9/russian-internet-lantern https://www.vice.com/en/article/88gevb/russia-is-preparing-to-cut-itself-of… https://www.vice.com/en/article/5dg4kb/russias-digital-iron-curtain-is-star… https://www.vice.com/en/article/xgdmdn/russian-law-fifteen-years-jail-tweet… https://edition.cnn.com/2022/03/12/tech/russia-internet-censorship-circumve… https://www.accessnow.org/digital-rights-ukraine-russia-conflict/ https://www.opentech.fund/results/supported-projects/lantern/ https://www.vice.com/en/article/jgx5z3/inside-the-plot-to-kill-the-open-tec… https://freedomhouse.org/report/freedom-net/2021/global-drive-control-big-t… https://getlantern.org/ https://yin.bi/ Meet the Secretive US Company Building an ‘Unbreakable’ Internet Inside Russia Backed with funding from the U.S. government, Lantern has spent two years building a peer-to-peer network inside Russia. As Russia makes preparations to possibly disconnect from the global internet in a bid to control the narrative around the invasion of Ukraine, one secretive U.S. company is rushing to lay the final pieces of an unbreakable network that the Kremlin won’t be able to take down. The company is Lantern, which says it has seen staggering growth inside Russia in the last four weeks for its app that allows users to bypass restrictions the Kremlin has put in place on platforms like Facebook, Twitter, and Instagram. But now the company is building something even more robust, an internal peer-to-peer network that will allow dissenting voices to continue to upload and share content even if the Kremlin pulls the plug on the internet. “We’ve been putting pieces in place in Russia for the last two years,” Lucas, one of Lantern’s developers, who uses a pseudonym to protect his identity, told VICE News. “So inside Russia right now, Lantern is also this peer-to-peer network that has all of this opposition content that’s distributed internally.” Within the next week, the network will be fully operational, allowing opposition voices to use the Lantern app to post content like videos from protests or updates on the war in Ukraine directly to the Lantern network. This would allow users to share it with other Lantern users without fear that the content will be removed or blocked. The app, which has in the past received funding from the Open Technology Fund, an independent non-profit funded by the U.S government, has become a huge hit in Russia in recent weeks. The amount of traffic passing through Lantern’s servers has risen 100,000% in the last four weeks according to the company, though it did not provide a baseline figure for comparison. Lantern said it would not break out country-level user numbers but told VICE News that globally the app has been downloaded 150 million times and has 7 million active monthly users, double the number it had three years ago. The spike in usage came after the Kremlin began shutting off access to mainstream social media sites while also introducing a law that threatened anyone publishing “fake news” about the war in Ukraine with up to 15 years in jail. The growth has been so rapid that Lantern says Russia this week surpassed China as its biggest market in terms of traffic, though China has been the company’s primary focus for years. The developers have spent years playing a game of cat-and-mouse with the Chinese censors who maintain the Great Firewall, Beijing’s highly sophisticated network of filters, bans, and blocks that prevent the free flow of information online. And just like in China, the developers of Lantern don’t have a clue how so many people in Russia have begun using their app. Tech Russia’s Digital Iron Curtain Is Starting To Take Shape Janus Rose 03.17.22 “It just started to spread and the same thing has happened in every region we’ve been in as well: It’s just word of mouth,” Lantern co-founder Wolf told VICE News. “We still don’t know. There are over 150 million downloads [of the app] and we haven’t spent one cent on marketing. So we’ve no idea how it spreads.” Wolf and Lucas, who have decades of experience building and scaling successful tech products, both use pseudonyms to protect their identities, fearing retaliation from the regimes in Beijing, and now Moscow, for helping people circumvent censorship. Aside from a couple of blog posts on Medium, Lantern has spent no money or effort on advertising or promotion inside Russia, and yet downloads on Google and Apple’s app stores have surged among Russian users eager to get an unfiltered view of the war in Ukraine. Part of the reason for this is that digital rights groups such as Access Now have been recommending Lantern, along with other anti-censorship tools like Tor and TunnelBear, to people in Russia seeking to get access to unfiltered information. “Now I have a window to the normal world from this prison called Russia,” one Russian user said this week in an email to Lucas. As the so-called “Digital Iron Curtain” has closed in recent weeks, Russian users have scrambled to find ways around it. In the days and weeks after the Russian invasion of Ukraine began, downloads of the top virtual private network (VPN) apps on Apple and Google’s app stores spiked, CNN reported. Natalia Krapiva, legal counsel with digital rights advocacy group Access Now says she’s unsurprised by the spike in Lantern users but says the number of people inside Russia using these types of tools is still in the minority. News Russian TikTok Influencers Are Being Paid to Spread Kremlin Propaganda David Gilbert 03.11.22 We’re definitely seeing them moving from being more of a niche tool turning into something much more widely used,” Krapiva told VICE News. “I wouldn't go so far as to say that the majority of people use them, I think we're still not there. I think there are still a lot of people, especially older, not as tech-savvy people who still don't know what VPNs are, how to use them, or which ones are safe.” While other VPN companies have reported some efforts to disrupt their services in Russia, Lantern works differently from a traditional VPN by focusing on finding ways to access blocked content. As a result, it can be used to access regionally locked down content on services like Netflix or BBC iPlayer, but its censorship-busting technology is more robust. The company says it’s seen no interruption to its services in Russia in recent weeks, largely thanks to the experience it gained battling China’s censors in recent years. “You need to remember that the Russian government is not going to be technically as good as China’s at squeezing our traffic and doing shit to us,” Wolf said. “China has so much more experience of screwing around with international traffic in general.” But Lantern’s ability to bypass the measures the Kremlin has put in place may soon be thwarted because Moscow is considering disconnecting from the global internet completely. The Russian government recently put measures in place that will allow it to disconnect completely from the global internet and rely on a so-called “sovereign internet” that would be entirely under Kremlin control. Just a month ago the idea that the Kremlin would pull the plug on the global internet was dismissed given how interconnected Russia was with the global economy with the global internet. But in the way of severe sanctions, Western companies pulling out of Russia, and the Kremlin already blocking large swathes of the wider internet, the possibility of it happening is increasingly likely. “There's little incentive for Russia to stay connected because the world is already disconnecting Russia themselves,” Krapiva said. “if Russia does disconnect, and then the world disconnects Russia, then we're going to have people just isolated and left alone, only with government surveillance, propaganda, and censorship, and that's not going to be helpful for anyone.” If that happens, Russian internet users won’t be able to access any non-Russian websites, even if they are using services like traditional VPNs, and this is where Lantern’s new peer-to-peer network will become so important, allowing users to continue to communicate and post content on an uncensorable network where videos, images, and other information can be shared freely among users—and there’s nothing the Kremlin can do about it. Lantern was founded in California in 2010 with the goal of keeping “the world’s information, speech, expression, and finance uncensored.” The free version of the app has a data cap of 500MB, but the pro version, which costs $32 a year, has no data cap. It has become hugely popular in China because of its ability to stay one step ahead of the government’s censorship efforts, spreading mainly via word-of-mouth as it’s not available via the Google or Apple app stores inside China. In Russia, like all new markets it enters, Lantern removed the data cap for all users. Despite this, some users still paid for the pro version. “We started to see a lot of revenue coming in from Russia,” Wolf said, adding that financial sanctions against Russia quickly shut off that revenue stream. This is something that has impacted many anti-censorship tools as payment companies have arbitrarily prevented Russians from paying for VPNs or other tools claiming they were doing so in solidarity with Ukraine or simply imposing sanctions. “A lot of it is misguided precisely for the reason that they're actually limiting Russian users ability to get the truth about what's happening in Ukraine and circumvent censorship,” Krapiva said. Demand for the tools like Lantern has rocketed in recent years as internet freedoms globally have been declining for more than a decade. In recent years the company has seen large growth in countries like the United Arab Emirates and Iran, where repressive regimes restrict access to large swathes of online content, including independent media outlets that are critical of the government. As well as the money it earns from its pro users, the company is funded by grants from various bodies, although its parent company, Innovate Labs, doesn’t disclose its sources of funding. One known source of funding, however, is the U.S. State Department, via its Open Technology Fund, which provided a total of $790,000 to Lantern in 2015. Lantern is one of several anti-censorship technologies that the OTF supports, including the encrypted web browser Tor and secure browsing tool Psiphon. The company has so far been reticent to speak to the media because the founders are wary about attracting the attention of regimes like China’s and Russia’s, and the possible fallout for them and their families. “We all have kids,” Wolf said. “We are trying to actively come out of the shadows. We don’t want to be some small confrontational tech company that fights governments, we want to grow to be a cutting-edge tech company. But we’ve got to get the balance right while we figure it all out.” But it’s increasingly hard to stay out of the limelight, as the app’s success in Russia shows, and will soon increase as the company is planning to roll out an even more ambitious product in the coming months. Lantern told VICE News that it is in the final stages of building another product that will allow people all over the world to host information on their own computers in censorship-free countries and provide access to that information to those inside repressive regimes. “People can run the Lantern [node] in these uncensored regions and act as access points for people in censored regions. So if you want to help out in Ukraine, and you’re in Berlin, just fire up Lantern, and it can become the gateway now for someone in Moscow,” Lucas said. To encourage people to set up a Lantern node, the company has created a cryptocurrency called Yinbi, which will be paid to those who allow their computers to act as hubs in this new decentralized network. But for now, Wolf says that Lantern is focused on Russia, and ensuring that no matter what the Kremlin does, its users inside the country will be able to access the information they need if “the Kremlin pulls the plug on the internet.”

3 3

Rape of Istanbul
by professor rat 27 Mar '22

27 Mar '22

Russian Chekists could overthrow democracy in Turkey as they did in Russia in 1918. Its quite possible Red Army Rapist's ' liberation ' of that country could also resemble their liberation of East Germany in 1945. " Being guilty of aggravated rape has nothing to do with our party" A. Broinowski. Wikileaks Party NSW candidate for au Senate. 2013.

1 0

Turkish authorites inspect a "mine-like object"
by professor rat 27 Mar '22

27 Mar '22

Turkish authorites have warned vessels to stay away as a dive team inspects a "mine-like object" " . . . the attack from the Rus' in 860 came as a surprise; it was as sudden and unexpected "as a swarm of wasps", https://en.wikipedia.org/wiki/Siege_of_Constantinople_(860) Russia warned that several mines had drifted away from Ukrainian ports.

3 2

" Natural Rights Anarchist ' Nazi posting here
by professor rat 27 Mar '22

27 Mar '22

Batshit-Crazy claims to be an anarchist. Then many of his posts uphold a core tenet of Nazism. Antisemitism. Once cryptoanarchy gets really big the NYT might make a stink of this. But I Douthat.

3 4

[spam][random][crazy] silly story
by Undiscussed Horrific Abuse, One Victim of Many 26 Mar '22

26 Mar '22

Once upon a time a beautiful thriving interdimensional utopia, filled with ancient souls living in perfect harmony, was drifting through a hell. Demons from the hell didn't understand the utopia. It was completely outside their reality. They tried to make casual conversation. Demon Fred: "Ah-ha! A new demon! I will rip you limb from limb like I have every other demon!" Fred was grandstanding. Strange utopic demon: "Oh!!! [lots of nice caring words that didn't make any sense in demon language]" Demon Fred: "Your language issue is pathetic. My torture and torment is far worse than yours! Burn me with your eternal fire, just try it!" Strange utopic demon: "Be nourished and well! May your horizons be ever bright and peaceful!"

3 11

" . . . no Jew can be a countryman.. . "
by professor rat 26 Mar '22

26 Mar '22

One of the leading points of the German Nazi Party. https://www.abelard.org/hitler/hitler.php#manifesto 4 Only those who are our fellow countrymen can become citizens. Only those who have German blood, regardless of creed, can be our countrymen. Hence no Jew can be a countryman. Totally Batshit-Crazy

1 0

America propaganda scumbag claimed to be ' cypherpunk revolutionary'
by professor rat 26 Mar '22

26 Mar '22

Au is a member of the infamous and unholy 5VEY as well as ANZUS Nazi "... Mr Assange agreed that some level of privacy was necessary for the successful operation of the military ..." http://www.abc.net.au/radionational/programs/drive/assange-i-can-rule-from-… NAZI JULIAN " ‘[The military] protects the sovereignty of Australia. It protects the independence of Australia.' ASSANGE July, 2013 Nazi Julian Assange · LRB 6 March 2014 Mar 6, 2014 - I wouldn't say this publicly, but Hitler wrote Mein Kampf in prison..."

1 0

Google Releases Chip Autodesign Training Data
by Karl Semich 26 Mar '22

26 Mar '22

https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.html https://github.com/google-research/google-research/tree/master/prime They don't release the actual model, though. Google showed me this on their main google.com website, and I took the bait. <http://ai.googleblog.com/> Blog <https://ai.googleblog.com/> The latest from Google Research Offline Optimization for Architecting Hardware Accelerators <http://ai.googleblog.com/2022/03/offline-optimization-for-architecting.html> Thursday, March 17, 2022 Posted by Amir Yazdanbakhsh, Research Scientist and Aviral Kumar, Student Researcher, Google Research Advances in machine learning (ML) often come with advances in hardware and computing systems <https://arxiv.org/abs/2009.06489>. For example, the growth of ML-based approaches in solving various problems in vision <https://ai.googleblog.com/2020/12/transformers-for-image-recognition-at.html> and language <https://ai.googleblog.com/2020/01/towards-conversational-agent-that-can.html> has led to the development of application-specific hardware accelerators <https://medium.com/@adi.fu7/ai-accelerators-part-i-intro-822c2cdb4ca4> (e.g., Google TPUs <https://cloud.google.com/tpu> and Edge TPUs <https://cloud.google.com/edge-tpu>). While promising, standard <https://arxiv.org/abs/2010.02075> procedures <https://arxiv.org/abs/2102.01723> for designing accelerators customized towards a target application require manual effort to devise a reasonably accurate simulator of hardware, followed by performing many time-intensive simulations to optimize the desired objective (e.g., optimizing for low power usage or latency when running a particular application). This involves identifying the right balance between total amount of compute and memory resources and communication bandwidth under various design constraints, such as the requirement to meet an upper bound on chip area usage and peak power. However, designing accelerators that meet these design constraints is often result in infeasible <https://ai.googleblog.com/2021/02/machine-learning-for-computer.html> designs <https://dl.acm.org/doi/abs/10.1145/3445814.3446762?casa_token=Iz0q7wWbnAIAA…>. To address these challenges, we ask: “*Is it possible to train an expressive deep neural network model on large amounts of existing accelerator data and then use the learned model to architect future generations of specialized accelerators, eliminating the need for computationally expensive hardware simulations?*” In “Data-Driven Offline Optimization for Architecting Hardware Accelerators <https://arxiv.org/pdf/2110.11346.pdf>”, accepted at ICLR 2022 <https://iclr.cc/>, we introduce PRIME, an approach focused on architecting accelerators based on data-driven optimization that only utilizes existing logged data (e.g., data leftover from traditional accelerator design efforts), consisting of accelerator designs and their corresponding performance metrics (e.g., latency, power, etc) to architect hardware accelerators without any further hardware simulation. This alleviates the need to run time-consuming simulations and enables reuse of data from past experiments, even when the set of target applications changes (e.g., an ML model for vision, language, or other objective), and even for unseen but related applications to the training set, in a zero-shot fashion. PRIME can be trained on data from prior simulations, a database of actually fabricated accelerators, and also a database of infeasible or failed accelerator designs1 <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…>. This approach for architecting accelerators — tailored towards both single- and multi-applications — improves performance upon state-of-the-art simulation-driven methods by about 1.2x-1.5x, while considerably reducing the required total simulation time by 93% and 99%, respectively. PRIME also architects effective accelerators for unseen applications in a zero-shot setting, outperforming simulation-based methods by 1.26x. <https://blogger.googleusercontent.com/img/a/AVvXsEiEDU5uq_ymXkZztEQ-ER0-3Mr…> PRIME uses logged accelerator data, consisting of both feasible and infeasible accelerators, to train a conservative model, which is used to design accelerators while meeting design constraints. PRIME architects accelerators with up to 1.5x smaller latency, while reducing the required hardware simulation time by up to 99%. The PRIME Approach for Architecting Accelerators Perhaps the simplest possible way to use a database of previously designed accelerators for hardware design is to use supervised machine learning to train a prediction model that can predict the performance objective for a given accelerator as input. Then, one could potentially design new accelerators by optimizing the performance output of this learned model with respect to the input accelerator design. Such an approach is known as model-based optimization <https://arxiv.org/abs/1912.13464>. However, this simple approach has a key limitation: it assumes that the prediction model can accurately predict the cost for every accelerator that we might encounter during optimization! It is well established that most prediction models trained via supervised learning misclassify adversarial examples that <https://arxiv.org/abs/1412.6572> “fool” the learned model into predicting incorrect values. Similarly, it has been shown that even optimizing the output of a supervised model finds adversarial <https://arxiv.org/pdf/2107.06882.pdf> examples <https://bair.berkeley.edu/blog/2021/10/25/coms_mbo/> that look promising under the learned model2 <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…>, but perform terribly under the ground truth objective. To address this limitation, PRIME learns a robust prediction model that is not prone to being fooled by adversarial examples (that we will describe shortly), which would be otherwise found during optimization. One can then simply optimize this model using any standard optimizer to architect simulators. More importantly, unlike prior methods, PRIME can also utilize existing databases of infeasible accelerators to learn what *not* to design. This is done by augmenting the supervised training of the learned model with additional loss terms that specifically penalize the value of the learned model on the infeasible accelerator designs and adversarial examples during training. This approach resembles a form of adversarial training <https://arxiv.org/abs/2102.01356>. In principle, one of the central benefits of a data-driven approach is that it should enable learning highly expressive and generalist models of the optimization objective that generalize over target applications, while also potentially being effective for new unseen applications for which a designer has never attempted to optimize accelerators. To train PRIME so that it generalizes to unseen applications, we modify the learned model to be conditioned on a context vector that identifies a given neural net application we wish to accelerate (as we discuss in our experiments below, we choose to use high-level features of the target application: such as number of feed-forward layers, number of convolutional layers, total parameters, etc. to serve as the context), and train a single, large model on accelerator data for all applications designers have seen so far. As we will discuss below in our results, this contextual modification of PRIME enables it to optimize accelerators both for multiple, simultaneous applications and new unseen applications in a zero-shot fashion. Does PRIME Outperform Custom-Engineered Accelerators? We evaluate PRIME on a variety of actual accelerator design tasks. We start by comparing the optimized accelerator design architected by PRIME targeted towards nine applications to the manually optimized EdgeTPU design <https://arxiv.org/abs/2102.10423>. EdgeTPU accelerators are primarily optimized towards running applications in image classification, particularly MobileNetV2 <https://arxiv.org/abs/1801.04381>, MobileNetV3 <https://ai.googleblog.com/2019/11/introducing-next-generation-on-device.html> and MobileNetEdge <https://ai.googleblog.com/2019/11/introducing-next-generation-on-device.html>. Our goal is to check if PRIME can design an accelerator that attains a lower latency than a baseline EdgeTPU accelerator3 <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…>, while also constraining the chip area to be under 27 mm2 (the default for the EdgeTPU accelerator). Shown below, we find that PRIME improves latency over EdgeTPU by 2.69x (up to 11.84x in t-RNN Enc <https://arxiv.org/abs/1211.3711>), while also reducing the chip area usage by 1.50x (up to 2.28x in MobileNetV3), even though it was never trained to reduce chip area! Even on the MobileNet image-classification models, for which the custom-engineered EdgeTPU accelerator was optimized, PRIME improves latency by 1.85x. <https://blogger.googleusercontent.com/img/a/AVvXsEgvcUqSo73Fl_QD6XEkHRSl6xf…> Comparing latencies (lower is better) of accelerator designs suggested by PRIME and EdgeTPU <https://arxiv.org/pdf/2102.10423.pdf> for single-model specialization. <https://blogger.googleusercontent.com/img/a/AVvXsEhFCwT37vYee4BR1BdXHDemMSG…> The chip area (lower is better) reduction compared to a baseline EdgeTPU <https://arxiv.org/pdf/2102.10423.pdf> design for single-model specialization. Designing Accelerators for New and Multiple Applications, Zero-Shot We now study how PRIME can use logged accelerator data to design accelerators for (1) *multiple applications*, where we optimize PRIME to design a single accelerator that works well across multiple applications simultaneously, and in a (2) *zero-shot setting*, where PRIME must generate an accelerator for new unseen application(s) without training on any data from such applications. In both settings, we train the contextual version of PRIME, conditioned on context vectors identifying the target applications and then optimize the learned model to obtain the final accelerator. We find that PRIME outperforms the best simulator-driven approach in both settings, even when very limited data is provided for training for a given application but many applications are available. Specifically in the zero-shot setting, PRIME outperforms the best simulator-driven method we compared to, attaining a reduction of 1.26x in latency. Further, the difference in performance increases as the number of training applications increases. <https://blogger.googleusercontent.com/img/a/AVvXsEhW5me64Hfab9faWxz0i8cqHMu…> The average latency (lower is better) of test applications under zero-shot setting compared to a state-of-the-art simulator-driven approach <https://arxiv.org/pdf/2102.01723.pdf>. The text on top of each bar shows the set of training applications. Closely Analyzing an Accelerator Designed by PRIME To provide more insight to hardware architecture, we examine the best accelerator designed by PRIME and compare it to the best accelerator found by the simulator-driven approach. We consider the setting where we need to jointly optimize the accelerator for all *nine* applications, MobileNetEdge <https://ai.googleblog.com/2019/11/introducing-next-generation-on-device.html> , MobileNetV2 <https://arxiv.org/abs/1801.04381>, MobileNetV3 <https://ai.googleblog.com/2019/11/introducing-next-generation-on-device.html>, M4, M5, M64 <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…> , t-RNN Dec, and t-RNN Enc <https://arxiv.org/abs/1211.3711>, and U-Net <https://arxiv.org/abs/1505.04597>, under a chip area constraint of 100 mm2. We find that PRIME improves latency by 1.35x over the simulator-driven approach. <https://blogger.googleusercontent.com/img/a/AVvXsEhfv6r5oWPT9q2WmmPlZ7ME-RD…> Per application latency (lower is better) for the best accelerator design suggested by PRIME and state-of-the-art simulator-driven approach <https://arxiv.org/pdf/2102.01723.pdf> for a multi-task accelerator design. PRIME reduces the average latency across all nine applications by 1.35x over the simulator-driven method. As shown above, while the latency of the accelerator designed by PRIME for MobileNetEdge <https://ai.googleblog.com/2019/11/introducing-next-generation-on-device.html> , MobileNetV2 <https://arxiv.org/abs/1801.04381>, MobileNetV3 <https://ai.googleblog.com/2019/11/introducing-next-generation-on-device.html>, M4, t-RNN Dec, and t-RNN Enc <https://arxiv.org/abs/1211.3711> are better, the accelerator found by the simulation-driven approach yields a lower latency in M5, M6, and U-Net <https://arxiv.org/abs/1505.04597>. By closely inspecting the accelerator configurations, we find that PRIME trades compute (64 cores for PRIME vs. 128 cores for the simulator-driven approach) for larger Processing Element (PE) memory size (2,097,152 bytes vs. 1,048,576 bytes). These results show that PRIME favors PE memory size to accommodate the larger memory requirements in t-RNN Dec and t-RNN Enc, where large reductions in latency were possible. Under a fixed area budget, favoring larger on-chip memory comes at the expense of lower compute power in the accelerator. This reduction in the accelerator's compute power leads to higher latency for the models with large numbers of compute operations, namely M5, M6, and U-Net. Conclusion The efficacy of PRIME highlights the potential for utilizing the logged offline data in an accelerator design pipeline. A likely avenue for future work is to scale this approach across an array of applications, where we expect to see larger gains because simulator-driven approaches would need to solve a complex optimization problem, akin to searching for needle in a haystack, whereas PRIME can benefit from generalization of the surrogate model. On the other hand, we would also note that PRIME outperforms prior simulator-driven methods we utilize and this makes it a promising candidate to be used within a simulator-driven method. More generally, training a strong offline optimization algorithm on offline datasets of low-performing designs can be a highly effective ingredient in at the very least, kickstarting hardware design, versus throwing out prior data. Finally, given the generality of PRIME, we hope to use it for hardware-software co-design, which exhibits a large search space but plenty of opportunity for generalization. We have also released <https://github.com/google-research/google-research/tree/master/prime> both the code for training PRIME and the dataset of accelerators. Acknowledgments *We thank our co-authors Sergey Levine, Kevin Swersky, and Milad Hashemi for their advice, thoughts and suggestions. We thank James Laudon, Cliff Young, Ravi Narayanaswami, Berkin Akin, Sheng-Chun Kao, Samira Khan, Suvinay Subramanian, Stella Aslibekyan, Christof Angermueller, and Olga Wichrowskafor for their help and support, and Sergey Levine for feedback on this blog post. In addition, we would like to extend our gratitude to the members of “Learn to Design Accelerators”, “EdgeTPU”, and the Vizier team <https://research.google/pubs/pub46180/> for providing invaluable feedback and suggestions. We would also like to thank Tom Small for the animated figure used in this post.* ------------------------------ 1The infeasible accelerator designs stem from build errors in silicon or compilation/mapping failures. ↩ <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…> 2This is akin to adversarial examples in supervised learning – these examples are close to the data points observed in the training dataset, but are misclassified by the classifier. ↩ <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…> 3The performance metrics for the baseline EdgeTPU accelerator are extracted from an industry-based hardware simulator tuned to match the performance of the actual hardware. ↩ <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…> 4These are proprietary object-detection models, and we refer to them as M4 (indicating Model 4), M5, and M6 in the paper. ↩ <https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.htm…> [image: Share on Twitter] [image: Share on Facebook]  <http://ai.googleblog.com/>  <http://ai.googleblog.com/2022/03/auto-generated-summaries-in-google-docs.ht…>  <http://ai.googleblog.com/2022/03/hybrid-quantum-algorithms-for-quantum.html> <https://www.google.com/> - Google <https://www.google.com/> - Privacy <https://www.google.com/policies/privacy/> - Terms <https://www.google.com/policies/terms/>

2 6