News Windows on ARM, getting 64 bit emulation

[Richie Rich]

Another small step for ARM world domination:

Microsoft will solve Windows on ARM’s biggest flaw beginning in November

Microsoft will release code to allow Windows on ARM PCs to run 64-bit X86 code via emulation, allowing every Snapdragon PC to run the same code as one powered by AMD and Intel.

www.pcworld.com

Nuvia Phoenix CPU with Nvidia GPU running Windows on ARM will be probably best desktop gaming rig. With double IPC even in emulation it will outperform Intel gaming CPU easily.

Next MS Xbox can have Nuvia Phoenix CPU. Backward compatibility for old games is secured via emulation. Huge performance for new games too.

 

[Gideon]

Nuvia Phoenix CPU with Nvidia GPU running Windows on ARM will be probably best desktop gaming rig. With double IPC even in emulation it will outperform Intel gaming CPU easily.

Next MS Xbox can have Nuvia Phoenix CPU. Backward compatibility for old games is secured via emulation. Huge performance for new games too.

You do realize that Nuvia has specifically stated that they will focus their products for hyperscalers? Desktop and laptop products won't be coming any time soon.
It's also almost certain Phoenix isn't a 2021 product (but rather 2022) from their own blog post:

We realize the companies we have measured against in these tests are not standing still and will have new products in the market over the next 18 months.

18 months from August 2020 is in 2022. Why worry about products that will be released later that their own?

IMO it would be better if this is a 2022 product, as it would increase the chances greatly of it being built on TSMC 5NP process. No way a small startup that only just now raised $240M has enough financials to battle with Apple for 5NP wafers in 2021.

 

[JoeRambo]

Any idea if Apple BootCamp will be available for ARM stuff? That system could literally serve as a highest end ARM development machine, now that VS + .NET5 have ARM builds.

Looking forward to the bright future, not sure if i need a leather jacket or not.

 

[moinmoin]

Patents shouldn't even be involved here. We have predecent where emulation was completely legal. For example, Sony sued the developers from the Bleem PSX emulator and lost hard. And Transmeta sold x86 compatible Processors that had a VLIW core with some sort of emulator-style front end to decode x86 instructions. So, who is going to raise an eyebrown for a built-in x86-64 emulator?

BTW, QEMU TCG (Tiny Code Generator), which is its CPU emulator, does NOT support AVX. BOSCH supports even certain AVX512 sets, though.

OH WAIT. Read THIS from 2017, which is directly reelevant since it involves Microsoft and Qualcomm CPUs. So basically, Intel is now a sideshow.

That last article you link there is golden. Though it also shows that Intel does try to prevent it from happening by using its patents wherever possible, even if not successful. This may well be sufficient for preventing earlier software implementations of Intel's ISA extensions as companies are not interested in clashing with Intel. Which would explain why Microsoft waited as long as it did with an x64 emulator.

I don't think public game system emulators are really comparable in this context, those are all reverse engineered clean room implementations, mostly hobbyist and, in the case of most newer game systems, work more high level, translating several different APIs and ISAs as they are used on a specific system. As for Bleem, it won the cases but went bankrupt due to the legal fees, which likely prevented similar commercial products to go public since.

You do not need the IP (e.g. the Microarchitecture) in order to write an emulator. The ISA is just enough - which is pretty much an open description.

To extend the whole issue at play a little: Oracle seems to think differently wrt what exactly law should and does protect, see

Google LLC v. Oracle America, Inc. - Wikipedia

en.wikipedia.org

Next argument session of that case seems to be in 5 days.

 

[Hitman928]

You do realize that Nuvia has specifically stated that they will focus their products for hyperscalers? Desktop and laptop products won't be coming any time soon.
It's also almost certain Phoenix isn't a 2021 product (but rather 2022) from their own blog post:

18 months from August 2020 is in 2022. Why worry about products that will be released later that their own?

IMO it would be better if this is a 2022 product, as it would increase the chances greatly of it being built on TSMC 5NP process. No way a small startup that only just now raised $240M has enough financials to battle with Apple for 5NP wafers in 2021.

Even the 18 months estimate is probably optimistic. Anandtech said the same thing (product release in 18 months) in their article about the Nuvia slides but then updated the article and removed that sentence saying they misunderstood that information. Nuvia must have contacted them with the correction which to me says that they don't plan on having the product ready within that 18 month timeline.

NUVIA Phoenix Targets +40-50% ST Performance Over Zen 2 for Only 33% the Power

www.anandtech.com

 

[Markfw]

Even the 18 months estimate is probably optimistic. Anandtech said the same thing (product release in 18 months) in their article about the Nuvia slides but then updated the article and removed that sentence saying they misunderstood that information. Nuvia must have contacted them with the correction which to me says that they don't plan on having the product ready within that 18 month timeline.

NUVIA Phoenix Targets +40-50% ST Performance Over Zen 2 for Only 33% the Power

www.anandtech.com

All I can say is "target" does not come close to "actual performance". I will believe it when I see it. Also, it states its targeting single core performance, where personal computing and servers NEED a lot of cores.

 

[dullard]

All I can say is "target" does not come close to "actual performance". I will believe it when I see it. Also, it states its targeting single core performance, where personal computing and servers NEED a lot of cores.

I'd say that the vast majority of personal computing does not NEED a lot of cores. Certain niche uses certainly needs a lot of cores. But to stretch that to all personal computing is not even remotely accurate.

For example, how did you possibly use computers in the past when they were 1 to 4 cores if you NEEDED it?

 

[Markfw]

I'd say that the vast majority of personal computing does not NEED a lot of cores. Certain niche uses certainly needs a lot of cores. But to stretch that to all personal computing is not even remotely accurate.

For example, how did you possibly use computers in the past when they were 1 to 4 cores if you NEEDED it?

Today, the smallest you can buy is a dual-core, and most of the CPUs are 2c/4t or 4c. That I think are the minimums, NOT 1c.

 

[dullard]

Today, the smallest you can buy is a dual-core, and most of the CPUs are 2c/4t or 4c. That I think are the minimums, NOT 1c.

Woudn't you agree that 2c or 4c is a far shot from your comment of "NEED a lot of cores"? To me, a lot is not 2 or even 4.

 

[Doug S]

Nuvia Phoenix CPU with Nvidia GPU running Windows on ARM will be probably best desktop gaming rig. With double IPC even in emulation it will outperform Intel gaming CPU easily.

I very much doubt Nuvia will ever sell their CPUs to consumers. Their target market are people who order by the hundreds of thousands of units, not single units. Someone would have to design a desktop board and chipset for a high end ARM gaming market that doesn't exist - Nuvia certainly won't be doing that, so who do you think will? Whoever did would have to hope the high end games that need maximum CPU performance get ported to ARM. It would cost millions to develop that board/chipset, with little reason to believe they'd ever sell enough to make their money back.

 

[Doug S]

Any idea if Apple BootCamp will be available for ARM stuff? That system could literally serve as a highest end ARM development machine, now that VS + .NET5 have ARM builds.

Looking forward to the bright future, not sure if i need a leather jacket or not.

Apple is not planning to support booting other operating systems. The EFI will be capable of being unlocked, but without drivers for stuff like the Apple designed GPU, good luck getting Windows or Linux to run on the bare metal (perhaps it might be possible to get Linux running with a serial/USB console?)

If you want to run another OS, it is likely only going to be possible in a VM.

 

[Markfw]

Woudn't you agree that 2c or 4c is a far shot from your comment of "NEED a lot of cores"? To me, a lot is not 2 or even 4.

OK, fair. I will update that to say "desktops need more than one core, and for SOME desktops,as many as 8, and servers can use all they can get., but NOBODY uses just one."

Better ?

 

[Richie Rich]

You do realize that Nuvia has specifically stated that they will focus their products for hyperscalers? Desktop and laptop products won't be coming any time soon.
It's also almost certain Phoenix isn't a 2021 product (but rather 2022) from their own blog post:

18 months from August 2020 is in 2022. Why worry about products that will be released later that their own?

IMO it would be better if this is a 2022 product, as it would increase the chances greatly of it being built on TSMC 5NP process. No way a small startup that only just now raised $240M has enough financials to battle with Apple for 5NP wafers in 2021.

Even the 18 months estimate is probably optimistic. Anandtech said the same thing (product release in 18 months) in their article about the Nuvia slides but then updated the article and removed that sentence saying they misunderstood that information. Nuvia must have contacted them with the correction which to me says that they don't plan on having the product ready within that 18 month timeline.

NUVIA Phoenix Targets +40-50% ST Performance Over Zen 2 for Only 33% the Power

www.anandtech.com

Nuvia started in 2018 so to have CPU product in 2021 is pretty optimistic, I agree. On the other hand Gerard Williams had in his mind the Apple A15 which they started development in 2018. Basically Nuvia Phoenix is probably based on uarch similar to Apple A15 so this can speed up the development IMHO. Phoenix will be kind of server version of A15 if Apple would ever decided to create one.

But two remarks:

1. Presenting performance curve in Geekbench means uarch is finished and they can run any code in SW simulation. I guess investors had an access to see real performance curve. Not only blue blob range provided publicly. Otherwise they would not received another 240 millions of foundation.
2. If uarch is finished, then how much time you need to tape it out? If they target Q1'22 or Q2'22 then Nuvia must make tape out end of this year, November or December. This means we might see Phoenix engineering samples performance leaks in March or April.

Nuvia has no product on market right now so by releasing performance numbers like ARM does, they can only gain publicity and traction against Intel and AMD. Or hurt their selling in servers. So I expect they will release tons of ES performance numbers.

 

[Gideon]

All I can say is "target" does not come close to "actual performance". I will believe it when I see it. Also, it states its targeting single core performance, where personal computing and servers NEED a lot of cores.

I do agree that "target" and "actual perfomance" are different things and it's good to retain a healthy amount of skepticism. Just bear in mind these guys have people that designed most of Apple's ARM cores (even some not released yet), so they are far from being charlatans.

By the way, Nuvia is excluseivly targeting servers only, hyperscalers more specifically. Their people have mentioned that in interviews time-and-time again. (and among other people they even have some working for them that come from Google's custom hardware projects).

I recommend reading the cited Nuvia blog-post in it's entirety, you'll see that they mention servers multiple times.

They also state quite clearly why they showed ST performance per watt estimates between 1W - 4.5W and why it matters for servers:

What does all this data mean for the server market? It means a lot. All current and future flagship server SoCs are power constrained, very much like mobile SoCs. This trend is only going to continue as there is a push to integrate even more cores. In addition, the AMD, Intel and ARM client computing cores tested are comparable to their current and future data center products. As core count increases, what is not increasing is the TDP. TDPs are likely going to remain in the 250W - 300W range, which is the maximum power that can be dissipated in an air-cooled environment in a typical datacenter. Hyperscalers and other enterprise data centers still must operate their servers within these TDP limits to optimize the TCO for their data centers. As more cores are added, the power allocation per core shrinks significantly.

A rough calculation can be made to determine the high and low bounds of the per-core power allocation in servers. We can assume that future flagship SoCs will have a minimum of 64 cores and a maximum of 128 cores. The TDP range is 250W - 300W, and the power outside of the CPU can range between 10W - 120W depending upon the workload. Taking into consideration these factors, the amount of power that each CPU can be allocated ranges between 1W - 4.5W when heavily utilized, as is the case in a datacenter environment. Drawing a set of vertical bars denoting this power range, it becomes evident why the NUVIA Phoenix CPU core has the potential to reset the bar for the market. No matter which scenario is considered, either unconstrained peak performance or power/thermally constrained performance, Phoenix should have a significant lead. Below is a preview into the planned capabilities of Phoenix, however we have left the upper part of the curve out to fully disclose at a later date. When measured against current products available in-market in the 1W-4.5W power envelope (per core), the Phoenix CPU core performs up to 2X faster than the competition. NUVIA’s Phoenix CPU performance is projected using architectural performance modeling techniques consistent with industry-standard practices on future CPU cores.

TL;DR That's because that is what upcoming 64-128 core servers will have to spend per-core being fully loaded (when being air-cooled).

That quote should also make it quite clear what their target is (250-300W TDP 128 core server).

And what they are targeting is 1.5x Zen 2 performance at 1/3 Zen 2 power draw. While it might sound hard to believe It isn't that unheard of if they are targeting 5 nm. After all AMD slashed power-draw in half going from 14nm -> 7nm.

Anyway, this wouldn't be that far off from a improved and die-shrinked Apple A13 (which is something these guys were part of making) if you take a look at the charts:

This isn't as bad as it sounds. If AMD manages to double perf/watt again with 5nm (as they have with 7nm) and improve perfomance by another 20-25% (with IPC + clocks) They'd be in the ballpark only losing slightly in power draw.

I'm far from taking Nuvias claims for granted, these are target's after all. But they have many "rockstars" of the industry. What I'm trying to say is that what they claim is something, which based on their past record , is at least theoretically achievable (not some outlandish snake oil claim it would seem at first glance).

 

[Markfw]

I do agree that "target" and "actual perfomance" are different things and it's good to retain a healthy amount of skepticism. Just bear in mind these guys have people that designed most of Apple's ARM cores (even some not released yet), so they are far from being charlatans.

By the way, Nuvia is excluseivly targeting servers only, hyperscalers more specifically. Their people have mentioned that in interviews time-and-time again. (and among other people they even have some working for them that come from Google's custom hardware projects).

I recommend reading the cited Nuvia blog-post in it's entirety, you'll see that they mention servers multiple times.

They also state quite clearly why they showed ST performance per watt estimates between 1W - 4.5W and why it matters for servers:


TL;DR That's because that is what upcoming 64-128 core servers will have to spend per-core being fully loaded (when being air-cooled).

That quote should also make it quite clear what their target is (250-300W TDP 128 core server).

AMD has a 64 core 200 watt chip, so twice that would be 400 watts. 300 watts (or even 250) is not 33% of 400. Thats what I saw as their "target". And 50% more performance than Milan ? (will be its competitor). I don't see it.

 

[moinmoin]

That quote should also make it quite clear what their target is (250-300W TDP 128 core server).

That's what I expect AMD to target with Epyc 4 (Zen 4, Genoa) on 5nm.

 

[NTMBK]

All I can say is "target" does not come close to "actual performance". I will believe it when I see it. Also, it states its targeting single core performance, where personal computing and servers NEED a lot of cores.

Nuvia's blog post says that they are targeting server CPUs with 64-128 cores- definitely not single core!

 

[dullard]

OK, fair. I will update that to say "desktops need more than one core, and for SOME desktops,as many as 8, and servers can use all they can get., but NOBODY uses just one."

Better ?

Seems better. If you just said HEDT computers or workstation computers or some other similar phrase representing computers that need a lot of cores, then I would have been happy too.

 

[Gideon]

AMD has a 64 core 200 watt chip, so twice that would be 400 watts. 300 watts (or even 250) is not 33% of 400. Thats what I saw as their "target". And 50% more performance than Milan ? (will be its competitor). I don't see it.

At 225W TDP Rome has about 3.5W to spend per-core (probably a bit less, as I/O and other stuff will draw power as well besides cores). 32-48 core models have a bit more to spend.

What Nuvia is claiming is that compared to current competition using 1.5-4W per core (e.g. where Epyc lands) they'll still manage to provide 50-100% more performance while using 1/3 of the current (not future) x86 designs power budget. E.g. score 1.5-2x better in geekbench @ 1.5W of what Zen 2 can currently do @ 4.5W.

This might mean for example a 128 core chip using 1.5W per core.

This is somewhat believable as the clock speeds of Epyc are in the ballpark of current ARM cores (A13) that do show up to 2x perf/watt in that region.

Now considering Epyc has about 2x performance/watt of 14nm Xeons and that Milan should provide ~20% more perfomance while Genoa hopefully does the same while also (hopefully) doubling perf/watt, this is pretty much the target they need to have to still be distruptive (e.g. offer a tangible advantage) in 2022.

 

[A///]

Does Nuvia have a physical product up for testing or are they basing their figures from where the sun don't shine?

 

[Hitman928]

Does Nuvia have a physical product up for testing or are they basing their figures from where the sun don't shine?

They don't have a product yet, their performance estimates will be based upon simulated performance.

 

[Markfw]

At 225W TDP Rome has about 3.5W to spend per-core (probably a bit less, as I/O and other stuff will draw power as well besides cores). 32-48 core models have a bit more to spend.

What Nuvia is claiming is that compared to current competition using 1.5-4W per core (e.g. where Epyc lands) they'll still manage to provide 50-100% more performance while using 1/3 of the current (not future) x86 designs power budget. E.g. score 1.5-2x better in geekbench @ 1.5W of what Zen 2 can currently do @ 4.5W.

This might mean for example a 128 core chip using 1.5W per core.

This is somewhat believable as the clock speeds of Epyc are in the ballpark of current ARM cores (A13) that do show up to 2x perf/watt in that region.

Now considering Epyc has about 2x performance/watt of 14nm Xeons and that Milan should provide ~20% more perfomance while Genoa hopefully does the same while also (hopefully) doubling perf/watt, this is pretty much the target they need to have to still be distruptive (e.g. offer a tangible advantage) in 2022.

The 7702 is a 200 watt 64 core EPYC, that is the one I was referencing. Thats 3.125 watts per CPU, and thats not even Milan, which will be competing with Nuvia.

 

[A///]

They don't have a product yet, their performance estimates will be based upon simulated performance.

I'll wait for a physical product.

 

[Thunder 57]

I very much doubt Nuvia will ever sell their CPUs to consumers. Their target market are people who order by the hundreds of thousands of units, not single units. Someone would have to design a desktop board and chipset for a high end ARM gaming market that doesn't exist - Nuvia certainly won't be doing that, so who do you think will? Whoever did would have to hope the high end games that need maximum CPU performance get ported to ARM. It would cost millions to develop that board/chipset, with little reason to believe they'd ever sell enough to make their money back.

How dare you bring logic and reason into all of this!

 

[TheGiant]

Nuvia Phoenix CPU with Nvidia GPU running Windows on ARM will be probably best desktop gaming rig. With double IPC even in emulation it will outperform Intel gaming CPU easily.

Next MS Xbox can have Nuvia Phoenix CPU. Backward compatibility for old games is secured via emulation. Huge performance for new games too.

you did have a smoke with Nosta ?
WTB that material