Chips and Cheese
NVIDIA’s Enterprise
#ChipAndCheese
Well, I guess that this will be the first non-welcome post on this website and the topic of this piece is NVIDIA Corporation, specifically the next-gen server GPU and more broadly the NVIDIA Enterprise division.
Now please remember that what is written below are leaks, rumours, and speculation, so take this with an appropriate amount of caution however what is published in this post is believed to be truthful and accurate.
Next-Generation Server GPU
As kopite7kimi has already posted on Twitter on in this tweet (https://twitter.com/kopite7kimi/status/1336900538185572357), the next-gen server GPU is an MCM design but has been delayed. Now, kopite7kimi did not go into detail as to why the next-gen server GPU has been delayed, however, which is the topic of this article.
From the information we have gathered from sources, it appears that NVIDIA is experiencing issues with making the next-gen server GPU appear as a single unified processor, instead of multiple GPUs, to the operating system because of firmware and driver issues.
NVIDIA Enterprise
We have also received reports that the Enterprise Driver team at NVIDIA has told at least one company “not to expect a response on any issues” from them; this is apparently due to people being moved to a different team at NVIDIA.
It is possible that these two events are linked, and that people were moved to work on the next-gen server GPU’s drivers, however, this is speculation by us here at Chips and Cheese.
The Future of NVIDIA in the Enterprise Space
With the rumored issues that NVIDIA is having in the Enterprise Space and with the now resurgent AMD along with Intel trying to break into the Enterprise GPU as well, NVIDIA looks like it may hit a rough spot in the near future.
source
NVIDIA’s Enterprise
#ChipAndCheese
Well, I guess that this will be the first non-welcome post on this website and the topic of this piece is NVIDIA Corporation, specifically the next-gen server GPU and more broadly the NVIDIA Enterprise division.
Now please remember that what is written below are leaks, rumours, and speculation, so take this with an appropriate amount of caution however what is published in this post is believed to be truthful and accurate.
Next-Generation Server GPU
As kopite7kimi has already posted on Twitter on in this tweet (https://twitter.com/kopite7kimi/status/1336900538185572357), the next-gen server GPU is an MCM design but has been delayed. Now, kopite7kimi did not go into detail as to why the next-gen server GPU has been delayed, however, which is the topic of this article.
From the information we have gathered from sources, it appears that NVIDIA is experiencing issues with making the next-gen server GPU appear as a single unified processor, instead of multiple GPUs, to the operating system because of firmware and driver issues.
NVIDIA Enterprise
We have also received reports that the Enterprise Driver team at NVIDIA has told at least one company “not to expect a response on any issues” from them; this is apparently due to people being moved to a different team at NVIDIA.
It is possible that these two events are linked, and that people were moved to work on the next-gen server GPU’s drivers, however, this is speculation by us here at Chips and Cheese.
The Future of NVIDIA in the Enterprise Space
With the rumored issues that NVIDIA is having in the Enterprise Space and with the now resurgent AMD along with Intel trying to break into the Enterprise GPU as well, NVIDIA looks like it may hit a rough spot in the near future.
source
开题答辩，答辩组长做NLP的，提问：分支是什么？
🤯
🤯
非常遗憾RVI最终以在RVA profile中保留C扩展的决定结束此次高通proposal的讨论。
我认为此次决定才是真正会给fragmentation埋下隐患的。
我认为此次决定才是真正会给fragmentation埋下隐患的。
#今日看了什么
https://lists.riscv.org/g/tech-profiles/topic/rvi_bod_decision_regarding/102522954
In particular, RVI made GC the defacto standard back in 2016 and reaffirmed C as mandatory in RVA profiles up through this past year. The BoD discussion was led by Dave Patterson. They felt it was important to resolve this now and not let the discussion continue. The BoD felt that not only was this the right decision but that other parties were already using this discussion to proclaim RISC-V fragmentation and lack of dependability.
The BoD has sent a strong message that unless there is something catastrophic (e.g. a security issue -- which we have none), that they want RVI members to honor the bond of RVI's word for items like this that are published as mandatory in profiles.
https://lists.riscv.org/g/tech-profiles/topic/rvi_bod_decision_regarding/102522954
In particular, RVI made GC the defacto standard back in 2016 and reaffirmed C as mandatory in RVA profiles up through this past year. The BoD discussion was led by Dave Patterson. They felt it was important to resolve this now and not let the discussion continue. The BoD felt that not only was this the right decision but that other parties were already using this discussion to proclaim RISC-V fragmentation and lack of dependability.
The BoD has sent a strong message that unless there is something catastrophic (e.g. a security issue -- which we have none), that they want RVI members to honor the bond of RVI's word for items like this that are published as mandatory in profiles.
Daniel Lemire's blog
Science and Technology links (November 12 2023)
1.
1. Vitamin K2 supplements might reduce the risk of myocardial infarction (heart attacks) and of all-cause death (Hasific et al. 2022). You find vitamin K2 in some Gouda cheeses and in eggs.
2. Most of the water on Earth is salinated (in the oceans) and cannot be consumed. Fresh water is often scarce. Yet Israel is desalinating water for less than a dollar per cubic meter.
3. People living in South America engaged in warfare for 10,000 years before the arrival of the Europeans (Standen et al. 2023).
4. The last glacial period ended about 12,000 years ago and last about 100,000 years. About 26,000 years ago, all of Canada was covered by a permanent ice sheet. Thus many of us were taught in school that human beings first colonized America about 12,000 years ago by the Bering land bridge, that existed back then between modern-day Russia and modern-day Alaska. The evidence accumulates that there were human beings in America much earlier than initially thougth. They would have been present 21,000 to 23,000 years ago in New Mexico. We even have their footprints.
5. As recently as 20,000 years ago—not long in geological terms—Britain was not, in fact, an island. Instead, the terrain that became the British Isles was linked to mainland Europe by Doggerland, a tract of now-submerged territory where early Mesolithic hunter-gatherers lived, settled and traveled (McGreevy, 2020). Correspondly, there were human beings in Ireland 31,000 years ago.
6. Gray et al. (2023) argue that the limited freedom that children enjoy in our modern societies is leading to a rise in mental disorders.
7. Most people cannot understand the bat and ball problem, even after the solution is given. The problem can be stated as follows: “A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?”
8. When hiring, we find a slight bias in favour of females in male-dominated fields, and a strong bias in favour of females in female-dominated fields (Schaerer et al., 2023). Overall, people greatly overestimate gender biases in hiring.
9. Retinol, a common cosmetic product, keeps one’s skin younger.
10. Unwarranted financial optimism might be the result of low cognitive abilities.
source
Science and Technology links (November 12 2023)
1.
1. Vitamin K2 supplements might reduce the risk of myocardial infarction (heart attacks) and of all-cause death (Hasific et al. 2022). You find vitamin K2 in some Gouda cheeses and in eggs.
2. Most of the water on Earth is salinated (in the oceans) and cannot be consumed. Fresh water is often scarce. Yet Israel is desalinating water for less than a dollar per cubic meter.
3. People living in South America engaged in warfare for 10,000 years before the arrival of the Europeans (Standen et al. 2023).
4. The last glacial period ended about 12,000 years ago and last about 100,000 years. About 26,000 years ago, all of Canada was covered by a permanent ice sheet. Thus many of us were taught in school that human beings first colonized America about 12,000 years ago by the Bering land bridge, that existed back then between modern-day Russia and modern-day Alaska. The evidence accumulates that there were human beings in America much earlier than initially thougth. They would have been present 21,000 to 23,000 years ago in New Mexico. We even have their footprints.
5. As recently as 20,000 years ago—not long in geological terms—Britain was not, in fact, an island. Instead, the terrain that became the British Isles was linked to mainland Europe by Doggerland, a tract of now-submerged territory where early Mesolithic hunter-gatherers lived, settled and traveled (McGreevy, 2020). Correspondly, there were human beings in Ireland 31,000 years ago.
6. Gray et al. (2023) argue that the limited freedom that children enjoy in our modern societies is leading to a rise in mental disorders.
7. Most people cannot understand the bat and ball problem, even after the solution is given. The problem can be stated as follows: “A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?”
8. When hiring, we find a slight bias in favour of females in male-dominated fields, and a strong bias in favour of females in female-dominated fields (Schaerer et al., 2023). Overall, people greatly overestimate gender biases in hiring.
9. Retinol, a common cosmetic product, keeps one’s skin younger.
10. Unwarranted financial optimism might be the result of low cognitive abilities.
source
Chips and Cheese
ARM’s Cortex A72: aarch64 for the Masses
#ChipAndCheese
Telegraph | source
(author: clamchowder)
ARM’s Cortex A72: aarch64 for the Masses
#ChipAndCheese
Telegraph | source
(author: clamchowder)
想想你日常生活中用的是狗屁论文，还是能跑的代码
BIRD Linux LLVM GCC Rocket真的能跑
ASPLOS ISCA MICRO 废纸一张
这不比什么狗屁论文好多了
内容：Soha为第一作者、我为第二作者向BIRD（BIRD Internet Routing Daemon）提交的关于OSPF路由协议的patch在昨天被社区接收了。这是我们团队首次在路由协议软件中做出贡献。此前，我们团队曾多次向Linux内核、LLVM编译器、GCC编译器、gem5模拟器、Rocket处理器核等社区提交过多份patch并获得接收。
Chips and Cheese
Core to Core Latency Data on Large Systems
#ChipAndCheese
Telegraph | source
(author: clamchowder)
Core to Core Latency Data on Large Systems
#ChipAndCheese
Telegraph | source
(author: clamchowder)
Daniel Lemire's blog
Generating arrays at compile-time in C++ with lambdas
Suppose that you want to check whether a character in C++ belongs to a fixed set, such as ‘\0’, ‘\x09’, ‘\x0a’,’\x0d’, ‘ ‘, ‘#’, ‘/’, ‘:’, ‘<‘, ‘>’, ‘?’, ‘@’, ‘[‘, ‘\\’, ‘]’, ‘^’, ‘|’. A simple way is to generate a 256-byte array of Boolean values and lookup the value. This approach is sometimes called memoization (and not memorization!!!). You might do it as follows:
It is reasonably efficient in practice. Some people might object to how the table is generated. Can you have the C++ compiler generate the array at compile-time from a function?
Using C++17, you might do it with an std::array as follows:
These two approaches should be equivalent in practice.
source
Generating arrays at compile-time in C++ with lambdas
Suppose that you want to check whether a character in C++ belongs to a fixed set, such as ‘\0’, ‘\x09’, ‘\x0a’,’\x0d’, ‘ ‘, ‘#’, ‘/’, ‘:’, ‘<‘, ‘>’, ‘?’, ‘@’, ‘[‘, ‘\\’, ‘]’, ‘^’, ‘|’. A simple way is to generate a 256-byte array of Boolean values and lookup the value. This approach is sometimes called memoization (and not memorization!!!). You might do it as follows:
constexpr static bool is_forbidden_host_code_point_table[] = {
1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
bool is_forbidden_host_code_point(char c) {
return is_forbidden_host_code_point_table[uint8_t(c)];
}
It is reasonably efficient in practice. Some people might object to how the table is generated. Can you have the C++ compiler generate the array at compile-time from a function?
Using C++17, you might do it with an std::array as follows:
constexpr static std::array<uint8_t, 256> is_forbidden_array = []() {
std::array<uint8_t, 256> result{};
for (uint8_t c : {'\0', '\x09', '\x0a','\x0d', ' ', '#', '/', ':',
'<', '>', '?', '@', '[', '\\', ']', '^', '|'}) {
result[c] = true;
}
return result;
}();
bool is_forbidden_host_code_point_array(char c) {
return is_forbidden_array[uint8_t(c)];
}
These two approaches should be equivalent in practice.
source
最近新u好多，我其实比较好奇他们到底用了多少面积。
总给我一种投了很多面积换来的感觉。
总给我一种投了很多面积换来的感觉。
#今日看了什么
Facile: Fast, Accurate, and Interpretable Basic-Block Throughput Prediction
https://arxiv.org/abs/2310.13212
Facile: Fast, Accurate, and Interpretable Basic-Block Throughput Prediction
https://arxiv.org/abs/2310.13212
感觉可以当综述论文来看，前端覆盖得比较全了。比较遗憾的是访存和缓存很少（也许这就是为什么国内没有M3）
