Intel enters the laptop discrete GPU market with Xe Max

https://arstechnica.com/gadgets/2020/11/intel-enters-the-laptop-discrete-gpu-market-with-xe-max/

by bpierre • 6 years ago

112 116 6 years ago

dvdkon• 6 years ago

This is a really confusing product for me. Do they really expect laptop manufacturers to spend money on a niche GPU compute workload accelerator? I'd say it's because Intel wants to show something for their GPU-building efforts, but they already did that with Iris Xe IGPUs.

The linked Ars Technica article says it well, "who will benefit enough from an Xe Max-equipped laptop"? I don't see anyone.

gbrown_• 6 years ago

I'd say it's because Intel wants to show something for their GPU-building efforts

I think this hits the nail on the head.

pinewurst• 6 years ago

In the same sense that Intel enabled/pushed a low volume Cannon Lake laptop as a China only SKU so they could tell the analysts, "see we've shipped 10nm" - even if the onboard GPU didn't work and had to have AMD RX540 instead.

lliamander9• 6 years ago

Maybe a cheaper discrete GPU option compared to Nvidia MX GPUs, and paired with cheaper comet lake processors?

merb• 6 years ago

Do they really expect laptop manufacturers to spend money on a niche GPU compute workload accelerator

macbook pro's if they keep their intel based cpus. if you use an external monitor it will always use the discrete gpu, besides that it might not be necessary. the discrete gpu inside the macbook pros heat really fast. it really depends on how good these are and how much heat they produce. if they are slightly worse but have way better cooling than its a win-win

TheRealSteel• 6 years ago

MacBook Pro isn't keeping their Intel CPUs. Apple officially announced a change to 'Apple Silicon' for Mac a few months ago.

refulgentis• 6 years ago

IMHO you're looking at it from a software engineer lens - from a product lens as I understand it, this is huge for OEMs as it's a way to upsell while gaining the ability to target a significantly larger chunk of the market. It provides a way to get 4k dual monitor support for pros and marketable video game support for prosumers, both in ~$600 fanless ultrabooks

hedora• 6 years ago

Is the discrete accelerator required for dual 4K? The article says it only really helps (vs the integrated intel GPU) with video encoding and machine learning workloads.

Ekaros• 6 years ago

Does it do those things better or cheaper than current offerings from Nvidia or AMD?

glangdale• 6 years ago

I hate to be cynical, but I suspect there's an Intel exec who will "benefit enough" from Intel having shipped this. I share your bafflement.

It feels like, at best, a 'pipe cleaner' (i.e. a product that's put through the whole process to debug it and ensure that a later product goes smoothly).

djmips• 6 years ago

A pipe cleaner seems reason enough not to be baffled.

fogihujy2• 6 years ago

That's pretty much what they did with the i740 back in the day, isn't it? It ain't stupid if it works.

ksec2• 6 years ago

This is a really confusing product for me.

1. Discrete GPU sells, Discrete = better than Integrated. ( Forget about the actual benchmark for the time being ) Especially true in market like China.

2. I would not be surprised if this was giving out for free with or for very little cost compared to just buying TigerLake. ( Or you know, normal price but with Intel Xe Max Marketing budget rebate )

3. The sole purpose is likely a Marketing exercise to address the world / market. We ( Intel ) now have a discrete GPU too. And they are good. ( You will always need to solve the Chicken and Egg problem with continue investment of GPU, revenue generation and demand for a product new to market segment)

4. Nearly 80% of PC sold are now Laptops. You need something to generate volume, view, exposures, Xe Max seems like well positioned.

5. You also need product volume to push developers to develop for their GPU / OneAPI. Just like Apple tells you e.g 60% of Active Devices, or 600M are already on iOS 14. Developer should plan their App upgrade with new API usage accordingly. Same with OneAPI and Intel Xe Max GPU. It is easier if you have a Total Addressable Market.

Q: So this is more or a marketing move than anything?

A: Possibly, Intel are exceptionally good at sales and marketing. Despite their Technological incompetence in the past years, I have to give them credit for that.

marmaduke3• 6 years ago

I would use one. The ISPC compiler recently got support for this new generation of GPUs and in principle it’s exactly the niche they want to hit. The code I write is then used by my research group colleagues.

The same goes for AVX512. It seems niche but then you use it and it’s great, but only a weird subset of laptops have it for now (but MBA 2020 and MBP 13 2020 do have it, MBP 16 doesn’t)

fluffything2• 6 years ago

How is ISPC's support for AMD MI supercomputing accelerators, or Nvidia's A100s ?

I feel like Hip, OneAPI, ISPC, etc. haven't improved anything over CUDA.

With CUDA we had one proprietary API. Now we have 3-4.

HelloNurse• 6 years ago

What poor souls use a low-power laptop GPU for content creation? This product falls squarely in the gap between inexpensive mainstream laptops and high-performance "workstation" or gaming laptops.

amelius• 6 years ago

During an extended product briefing, Intel stressed to us that the Xe Max beats Nvidia's entry-level MX 350 chipset in just about every conceivable metric. In another year, this would have been exciting—but the Xe Max is only slated to appear in systems that feature Tiger Lake processors, whose Iris Xe integrated GPUs already handily outperform the Nvidia MX 350 in both Intel's tests and our own.

Two GPUs are more powerful than one, so I don't see the problem.

robin_reala• 6 years ago

The problem is that you can’t use both at the same time:[1]

To cut right to the chase on an important question for our more technical readers, Intel has not developed any kind of multi-GPU rendering technology that allows for multiple GPUs to be used together for a single graphics task (ala NVIDIA’s SLI or AMD’s CrossFire). So there is no way to combine a Tiger Lake-U iGPU with Xe MAX and double your DOTA framerate, for example. Functionally, Xe MAX is closer to a graphics co-processor – literally a second GPU in the system.

This basically seems to make it only useful as some sort of encoding/decoding co-processor, with potential but currently undescribed / undefined ML benefits.

[1] https://www.anandtech.com/show/16210/intels-discrete-gpu-era...

mtgx• 6 years ago

They expect users to pay twice for their GPUs - once for the integrated one and once for the dedicated one.

They are probably attempting this because for a long time OEMs would include both Intel's iGPU (not like they had a choice) but also an Nvidia GPU, even if it had the same or lower performance than the iGPU, just so Intel doesn't fully monopolize their devices.

Then again, by this logic not sure why any OEM would pay twice for Intel GPUs, either.

zamadatix2• 6 years ago

SLI/CrossFire are dead technologies, the latest graphics APIs offer ways to use independent GPUs simultaneously without these.

The problem is neither the old or new method of multi-gpu has ever been really successful at actually working in the app you want it to.

numpad0• 6 years ago

You can use dGPU for rendering and mirror frame buffer content over to iGPU mostly fine these days, as NVIDIA Optimus laptops do.

That you can’t do multi-GPU means you can’t combine GPUs to make them into a NUMA GPU cluster that is better performing than any of its nodes.

baybal23• 6 years ago

It's also very odd that they did not consider that MX niche while knowing that casual laptop gaming has been crushed by mobile Ryzens, which have embedded GPUs blowing them out of water as well for "free"

trynumber9• 6 years ago

Xe HP and Xe HPG seem more interesting but I wonder if they'll be too late to be competitive.

https://www.anandtech.com/show/16018/intel-xe-hp-graphics-ea...

baybal23• 6 years ago

LDPPR4x is a very odd choice for graphics.

LPDDR4x soundly loses to HBM2E on energy/bandwidth ratio.

6-8 picojoules per bit to 1.5-2.5

wmf3• 6 years ago

Laptops generally can't afford HBM.

I have a hunch that DG1 is Tiger Lake with the cores cut out so that's where the memory controller comes from.

baybal23• 6 years ago

HBM memory is not that expensive, what is expensive is the (co-)packaging

But Intel has already committed to using active silicon interposer instead of abf for substrate, which is way more expensive.

lliamander9• 6 years ago

Given Intel's reputation stable Linux drivers, I'm excited for their discrete desktop GPUs.

DCKing3• 6 years ago

Yeah! People say AMD has the best Linux drivers, but they're wrong! Intel's are feature rich and stable, and I think they're better than AMD's [1]. They're just kind of ignored since nobody "chooses" to use an Intel GPU in their computer. You just get one bundled in your CPU, until now that is.

Intel Xe will likely be very poorly received because they'll be compared to AMD and Nvidia's best offerings and "Intel is failing at everything" is the big story now. But AMD and Nvidia are treating the low end GPU market as some sort of a backwater, and there's a proper niche for people who just want a reliable graphics card with modern video outputs and modern video codecs. If it can run desktop 3D effects smoothly on multiple high resolution screens - like a better or less expensive GT1030 and RX550 - you can be golden. Not everything GPU is about gaming or high end compute.

[1]: Okay, AMD's Linux drivers are also really good, but they can still have more hairy edge cases and lack of development than Intel's.

pedrocr• 6 years ago

But why would any of those people want this part? The only way I see it is if you want an Intel GPU for the compatibility in an otherwise AMD system. If not the Intel CPU you'd build already includes a more than capable GPU this generation.

lliamander9• 6 years ago

Exactly. The market for GPUs using 75W should get interesting soon.

echlebek• 6 years ago

You're absolutely correct. Especially in the low-power segment, which AMD and nvidia have really ignored lately.

If you want a <=75 watt GPU, which happens to be the limit for the PCIe slot without an external power connector, you have very limited options. For nVidia, the 1650 is the only modern part that will fit the bill. For AMD, you have to reach back to the RX 500 series, but good luck finding one. They're extremely uncommon even among the lower-end 500 series, as many of them require an external connector anyways.

AMD do have lower-powered "workstation" offerings, but they're so expensive that it's hard to imagine putting one in a workstation voluntarily. They're also extremely poor performers, to the extent that you might not really see much of a benefit over a Vega iGPU.

The Xe Max is only 25W TDP, which is very attractive if you're hoping to build a passively cooled Linux machine with a low-wattage power supply.

tw042• 6 years ago

But AMD and Nvidia are treating the low end GPU market as some sort of a backwater

Are they though? The GPU embedded in my ryzen 4900 seems to be just as good, if not better than this xe. The desktop version I would expect to show even better.

pabs3• 6 years ago

More recently Intel's GPUs have started to require proprietary firmware, which they did not need before, although probably it was embedded in the GPU before.

xbmcuser2• 6 years ago

Yeah AMD drivers on linux are not very good. I fell for the nvidia hate because of its propritory drivers so upgraded with an AMD Ryzen Apu instead of just getting a new nvidia graphics card. It's not even a recent APU it's almost 1+ years old. Now I am still using the my nvidia card as the Ryzen Apu can't output 4k and need to upgrade my graphics card anyway.

pjmlp3• 6 years ago

Intel also has an history of bad OpenGL implementations, where their drivers would state to implement features, which were actually done in software, making meaningless the feature query to decide what to call.

ksec2• 6 years ago

They're just kind of ignored since nobody "chooses" to use an Intel GPU in their computer.

That is not entirely accurate though.

From a total GPU market presumptive, Low End / Low Power equals integrated GPU. And both AMD and Nvidia simply cant compete with that. The market of a Low End GPU is tiny and unprofitable. When most consumers or customers are happy with their Intel iGPU.

I dont blame them for lack of SKUs in this segment. And AMD is competing with their own APU as well.

fogihujy2• 6 years ago

They're just kind of ignored since nobody "chooses" to use an Intel GPU in their computer. You just get one bundled in your CPU, until now that is.

People looking for Linux laptops certainly do. I've tried the AMD route twice the last ten years and I've been disappointed by crappy driver support and battery life in both cases.

gpderetta• 6 years ago

But for your scenario, an integrated GPU is just fine, and Intel already cover that market well.

To justify it existence, a discrete GPU need to be better than that.

zucker422• 6 years ago

But AMD and Nvidia are treating the low end GPU market as some sort of a backwater

I have to guess that's because the low end GPU market is not very profitable.

Const-me• 6 years ago

nobody "chooses" to use an Intel GPU in their computer.

I did. My laptop has i3-6157u CPU, I chose it due to the Intel Iris GPU there.

If I would be looking for a new one today, for similar reasons I would have picked an AMD APU.

bayindirh• 6 years ago

Yeah! People say AMD has the best Linux drivers, but they're wrong! Intel's are feature rich and stable, and I think they're better than AMD's [1].

I'd like to counter this. Intel's driver quality is going down for a year or so. My old EliteBook 850G2 is running with Intel Graphics but, it started to show some flickering and corruptions for no reason. Nobody bothered to fix it yet.

Intel's e1000e drivers started to behave awry on some older cards and, Intel didn't fix it. On the contrary, they continued to pile patches over this buggy driver. Kernel guys and Debian developers had to revert some patches back to stabilize the driver again. It still loses connectivity sometimes.

By older cards, I mean some of them are barely two years old.

IF, a big IF, Intel gets its thing together and restores the quality, it's great. Otherwise, I don't put them on a pedestal for driver quality.

On the Windows side, I have an Intel wireless card which is officially supported by current drivers but, the new drivers bork the card after standby and system has to be powered off to restore card connectivity. The computer is a HP Spectre X2 convertible and, only older, out of the box Windows drivers (which are stock Intel drivers) can utilize the card correctly.

klodolph• 6 years ago

I think that this is because Intel's drivers are just Mesa drivers, and part of why that's possible in the first place is because Intel's not really competing with AMD or Nvidia and don't make efforts to keep their designs secret... which means that Intel isn't duplicating work already done in Mesa, but building on top of it and making Mesa better.

zamadatix2• 6 years ago

AMD uses Mesa as well.

matheusmoreira• 6 years ago

Same here. My laptops have lots of Intel hardware and they all work out of the box on Linux. Hopefully Intel GPUs will also be an option for desktops.

wegs• 6 years ago

I'd pay a pretty penny for a decent, stable, well-supported open GPU.

It'd need to cover the corner cases well. Multimonitor setups. High res. Etc.

toolslive• 6 years ago

they tried this (!?) a decade ago too.

https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

wmf3• 6 years ago

As meh as Xe is, it's a working GPU which is dramatically more mature than Larrabee ever was.

wtallis3• 6 years ago

Larrabee was a weird attempt to force x86 into the GPU space. It was more hubris than silicon. Xe is Intel actually making a serious attempt to build decent discrete GPUs, without bizarre political constraints on the technological designs the team can pursue.

skissane• 6 years ago

I view it as an attempt to explore the question "can a competitive GPU architecture be built as an extension of a mainstream CPU architecture?". And they tried it and it didn't work out, which suggests the answer is "No". But I don't fault them for trying – a willingness to try out different areas of the design space, even test them in the market, is a positive thing. Of course, you can see why a vendor whose major product is a mainstream CPU architecture might have a special interest in exploring that particular design question, but it still was a valid question to explore, even if they didn't get the result they were hoping for. We can learn as much from failures as from successes.

badsectoracula• 6 years ago

IIRC Larrabee actually worked but it failed largely due to internal politics, not because of technical issues.

AceJohnny2• 6 years ago

It's funny to consider RISC-V in the light of Larrabee. RISC-V was conceived as a multi-purpose extensible ISA to unify the toolspace for MCUs, DSPs, CPUs, GPUs, etc.

Larrabee wanted to leverage the toolspace of the x86 ISA, but that same ISA complicated their architecture and made it nonviable.

jcranmer• 6 years ago

Larrabee kind of lived on in the Xeon Phi line, although the Xeon Phis were purely targeted at compute with no graphics capability like Larrabee was.

Intel Xe is closer to taking the existing integrated GPUs and sticking them on a discrete card instead of the same socket as the main CPU.

mrec2• 6 years ago

And the i740 two decades ago. Intel does not have a great track record with graphics.

https://en.wikipedia.org/wiki/Intel740

arthurcolle2• 6 years ago

Sorry for the aside, but is it possible to create a GPU cluster for, say, machine learning or rendering that uses both NVDA and AMD GPUs, together? I have 4 AMD GPUs and 2 NVDA 1080 TIs and was curious if its possible to wire up a computer that uses all 6 simultaneously

projct• 6 years ago

Sort of.

Mixed brand has a lot of bugs, and the different ways you'd hook up more than 4 cards have a lot of drawbacks so you would have to plan this carefully. It would be easier and more flexible to get a 4 slot board for the AMD GPUs and use the NVIDIA ones in a different machine.

EDIT: note also that the 4 slot board does not need to populate all the PCIE lanes necessarily depending on your workload.

arthurcolle2• 6 years ago

Are there issues with using GPU drivers from several different GPU vendors on the same machine? I heard that from some people that were into the crypto-mining scene, and was curious if that carries over to all potential GPU cluster applications... Thanks for answering!

nl2• 6 years ago

For ML maybe?

You can get the AMD builds of TF and PyTorch and then build a cluster out.

Realistically it'd likely to be more trouble than its worth. The AMD versions are always behind the core versions.

ehvatum• 6 years ago

Double precision FP, cl_khr_fp64, could not be supported by intel’s GPU silicon until Xeon phi / larabee was inevitably cancelled. That finally did happen. I would love to add more double precision compute capacity without being robbed; it’s getting tough to find 1st gen nvidia titan cards. Here’s hoping this discrete unit will have strong fp64 performance.

Edit: It had been a while since I checked into current-model GPU fp64 throughput. The Titan V has outstanding DP performance!!! Order placed.

inasio• 6 years ago

Of note here I believe TSMC is fabbing the bigger version of this chip for intel on their 7/6 nm node.

geogra4• 6 years ago

Would be nice to have a pcie version of this

SomeoneFromCA• 6 years ago

If the standalone GPU (I am sure they will produce one soon) comes out cheap and power efficient, ironically AMD could benefit from that more than Intel. Most of the top AMD processors have no integrated GPU. I would certainly prefer 3600+Intel GPU over more expensive i7-9700.

Traster2• 6 years ago

Well this is just outright confusing. The scale that Intel operates at means this product is going to be a drop in the bucket in terms of revenue which makes their enormous investments in dGPUs look a little silly. It fits in this really weird place where you want a little bit of extra GPU power for co-processing, but really very little extra power. I literally can't understand how this product justified the engineering effort. Which leaves us with the fact that this must be a first release of something much more significant, which means that they've been forced to rush it out the door because their main effort is behind (well 'technically' we did ship the first series of chips in this project in Q42020). However, if you watch what happens a lot in Intel - they release a compromised first version of something, that ties them to a concrete failure (this isn't going to make any money) which then allows the executive team to can the project.

purplecats• 6 years ago

how often are you correct about your chain of reasonings as such? if it's non-negligible, would you mind sharing where I could read more of your ponderings? thanks.

Traster2• 6 years ago

Well, on this particular topic, I used to work for Intel so a lot of it comes from knowledge of how the company works institutionally. It's difficult to explain to an outsider but Intel has this weird combination of making enormous bets but with an incredibly ossified corporate culture. So some project will get kicked off - they'll buy a company like Nervana or they'll hire in someone like Jim Keller and give them a mandate to go off and do something huge.

The problem is how this intersects with the dynamics with the Intel organisation. Firstly, there's already five teams that already do whatever you want to do, so you go up the chain saying "I need 5 engineers to do X" and the answer comes back down the chain "Team Y already has 20 people, they do X, you should get them to do it". But so now rather than doing X, you're paying Team Y to do X, oh and Team Y doesn't give a shit about X, they care about Z, so about half your budget for X is now being subtly shifted to work on Z. So the cost of your project massively inflates.

Oh, and because when your project was approved they canned something else, some senior executive is saying that half your headcount should be "internal transfers" so suddenly your project has to find work for 50 software engineers at Folsom who just finished working on Intel's Modem team (before that they worked on Intel's previous GPU team too before that got canned).

Now comes the real issues, because you've got all these different teams that you rely on for your project and they don't answer to you, they answer to random different org structures, it's practically impossible to hold them accountable. Which leads to hilarious dynamics where team A will fall behind on their tasks, but they already know team B has fallen behind, so team A will lie about being able to deliver, because they know they'll never have to. So the critical path on your project isn't the team that's fallen behind, it's the team that's fallen behind and all the other teams that have made the calculated judgement your work can be deprioritized because team B will sink your project anyway.

Finally comes the release. All of these internal fights don't change the fact that you're responsible for delivering a product to market, and almost certainly if it's an important project it'll be part of some Group's goals (like the Datacentre group or networking group). Now, the General Manager of the group isn't going to miss her/his goals. So you'll "ship" a product, by which we mean you'll send a handful of sample products to a partner. They probably wont work, they'll be a tiny subset of the functionality you originally promised. So anyway, that's the point at which someone realises we've just spent 10,000 years of engineering effort to deliver a discrete Graphics chip that is marginally less powerful than the integrated one.

But the product is out there, so you've got sales data now and revenue. So you need to agree your roadmap for the future! But the first product was shit, and getting it into any sort of state to be competitive would be an enormous amount of work because it was compromised at every point, so no one will sign off on throwing good money after bad and so what happens? Nervana? Fuck you! We're betting on Habana Labs now!

srg0• 6 years ago

If it's a content creation tool, shouldn't it be compared to the Quadro series?

Intel enters the laptop discrete GPU market with Xe Max

Related Stories