Intel Larrabee finally hits 1TFLOPS - 2.7x faster than nVidia GT200

During the recently held SC09 conference in Portland, Oregon - Intel finally managed to reach its original performance goal for Larrabee. Back in 2006, when the first details about Larrabee emerged, the performance goal was "1TFLOPS@ 16 cores, 2.0 GHz clock, 150W TDP". During Justin Rattner's keynote, Intel demonstrated the performance of LRB as it stands today.

At SGEMM Performance test [4K by 4K Matrix Multiply, QCD], Intel achieved 417 GFLOPS using half the cores on the prototype card, and reached 825 GFLOPS by enabling all the cores. While looking at the numbers alone, one might think that these scores are below the level of ATI Radeon 4850 and nVidia GeForce GTX 280/GTX 285. Of course, there is a "but" coming - unlike theoretical numbers that are usually disclosed by ATI and nVidia - this was an actual SGEMM benchmark calculation used in the HPC community.

[darkequitus]

Yawn. I will have another look when some non-HPC stats come out as I will not being doing weather sims at home. And until they set a re;ease date, I will stick with a Radeon 5850 as soon a they come available again in Dec.

[JsonHenry]

agreed, until they show some REAL numbers that we will actually be using I am sticking with my 5870.

[Nihilism]

Tflops mean very little in terms of performance for the average user, can this play crysis? no it cannot. If they want to market it as a cpu/gpu solution then it had better be able to hold it's own in the graphics department too, and it won't.

[meetajhu]

Tflops is nothing to games department! It handles Crysis Maximum like a 8800 Ultra. But Geforce 300 or Nvidia Fermi will be atleast 10x faster in games when compared to Larrabee!

[champ21]

current gen gpus get about 120-150gb/s memory bandwidth from the dedicated memory they got.

Larabee will be sharing the memory bandwidth between a cpu & gpu, I7 with its tripple channel memory only has 25gb/s memory bandwidth now imagine sharing that with a gpu.

IMO for larrabee to be viable memory bandwidth would have to grow 5 folds just to meet the bandwidth on current gpus.

all of us know how important this bandwidth is, i dont care how powerful a gpu is, without the memory bandwidth nothing is happening.

[Ju]

These are pretty amazing numbers. It remains to be seen what real world application (other then scientific number crunching) will produce with this baby. Anyhow, (close to) 1TF for 16 cores is a huge achievement. PowerXCell is 160GF with 8 cores, that chip is more then twice as fast. Impressive.

Graphics benchmarks would be certainly interesting, too. But the raw performance of that chip allows a pure SW shader to outperform today's top off the line graphics cards.

[ProblemSolver]

@Ju: Well, the PowerXCell 8i processor performs not 160 GFLOPS it does 202
GFLOPS on SGEMM (4k x 4k) utilising 8 SPEs. The PowerXCell's peak-performance,
counting only the SPEs (neglecting the PPU), computes as follows;

8 [[email protected]] = 8*(8 flops * 3.2GHz) = 8 * 25.6 Gflops(SP) = 204.8 Gflops(SP)
(SP := single precision)

Hence, the PowerXCell 8i as well as the Cell/B.E. processor inside the PS3
perform the SGEMM computational kernel with ~99% of its peak-performance. No
other processor in existence can match this number.

To put the PowerXCell 8i in perspective to Larrabee, with respect to the amount
of cores, we have to take two PowerXCell 8i to gain 16 SPEs. Two PowerXCell 8i
processors perform the SGEMM kernel at 406.04 GFLOPS, which amount to ~99% of
the theoretical peak performance of 409.60 GFLOPS.

[Reference]
// Fast Matrix Multiplication on Cell (SMP) Systems
http://tu-dresden.de/die_tu...

Larrabee performs the SGEMM kernel with 16 cores at 2Ghz with 825 GFLOPS, which
is only twice as fast as two PowerXCell 8i processors (16SPEs) where one has to
consider that the vector width of Larrabee is 16 and that of Cell processor
only 4.

What's the theoretical peak performance of the given Larrabee configuration?

We have,
16 [[email protected]] = 16*(32 flops * 2.0GHz) = 16 * 64 Gflops(SP) = 1024 Gflops(SP)

Now we can compute the efficiency of the SGEMM kernel for Larrabee;
(825 GFLOPS * 100) / 1024 = ~81%

Hence, we have
2 PowerXCell 8i @ SGEMM (4k x 4k) = 406.04 GFLOPS; efficiency = ~99%
Larrabee @ SGEMM (4k x 4k) = 825 GFLOPS; efficiency = ~81%

This shows that Larrabee's computational units get starved for date -- a possible
weak spot of Larrabee's architecture. It seems like that Larrabee's memory model
(an implicit coherent shared memory model) can't deliver the data fast enough to
compete with the explicit memory model (direct DMA'ing) of the Cell processor.
The explicit memory model of the Cell processor is what makes this processor
so efficient and unique, but also difficult to program for some developers.

[TABSF]

Great read ProblemSolver.

Although the Cell my be efficient it does not mean its the best.

IBM, Intel and AMD are very scared at the moment imho with what Nvidia is doing and that's the reason these chip makers are going down the hybrid route

Nvidia plan is to have CPU free Desktop, Laptops, Net-tops, Netbooks and Servers. Fermi is the big step to Nvidia overall plan.

[Ju]

@ProblemSover. Interesting thoughts. As always. Just wanted to note that down here.

[ProblemSolver]

@TABSF: Best depends on a metric.

If I wanna have a processor that computes scientific workloads at near
peak-performance and that offers the most GFLOPS / $1 / Watt, then the Cell
processor is the best within this regard. As such it's no surprise that the
Cell processor was used in Roadrunner. And given the latest order of 2000 PS3s
from the military proves this fact as well. In HPC (High Performance Computing)
the metric GFLOPS / $1 isn't even the most important one, its GLOPS / Watt that
counts. The electricity bill of a cluster can easily exceed the initial cost of
a cluster over its lifetime.

From an scientific view-point I haven't seen any better architecture as the
one of the Cell processor. Unfortunately, this processor wasn't backed up
sufficiently with strong libraries from the get go, making developing software
difficult for many game-developers. Like I've said before, this architecture is
ahead of its time. I currently don't know whether IBM is going to make another
Cell processor, but I'm pretty sure that we will see a similar architecture
again in the future. Why? If you wanna know this, then you have to understand
at what point a coherent (cache-) memory / multi-threaded architecture breaks.
The only advantage of such systems is that it's easier to program. But their
efficiency diminishes when they start to scale. Larrabee will be an indication
of this. Standard programming on Larrabee won't give any performance due to
many cache misses. Hiding those misses via hyper-threading is very limited. The
assumption is that the three other threads (out of four) have to work on the
current data within the cache since otherwise they would produce another cache
miss leading to trash the cache completely. To gain efficiency on Larrabee one
has to program as follows;
Given that each core has only 256KB cache at its disposal, the cache must
implicitly be splitted. One thread is the main thread that works out of about
200KB cache. Another thread (a helper thread) loads data (implicitly by
touching a memory address) into the remaining part of the cache that is worked
on by thread three and four once thread one undergoes a cache miss. If you
look at it, this is a much more complicated way to gain efficiency compared
to the Cell processor, because the memory hierarchy within Larrabee isn't
programmable. However, as I said, in general it will be easier for the
programmer to throw code at Larrabee, but this code won't run very efficiently.
Cache miss elimination will be key. And if one goes as far trying to eliminate
most of the cache misses by code-restructuring and intelligent threading, then
that person might come to the conclusion that programming on Cell would be much
easier. So why Larrabee at all? Well, Intel has spent many years in optimizing
their compilers for coherent (cache-) memory / multi-threaded architectures.
They have no other choice. If they could, they would start anew, for sure.
IBM got the chance to do so, due to high-volume productions of the PS3. But even
than, it requires too much time to build strong compilers and backends to
support the new processor, which is actually Cell's weakest point. To bring the
Cell architecture back on stage, IBM has to spend some more year in optimizing
compilers for such an architecture. Possibly that is what they are doing now.
If true, we may likely see something similar to Cell, if not Cell2, in the
future.

[TABSF]

Intel propaganda

GTX 285 (GT 200) has a single precision floating point of 1.1Tflop/s
GT 300 (Fermi) were looking at around 630Gflop/s Double Precision

[Perkel]

yep

1. Intel Larrabee [LRB, 45nm] - 1006 GFLOPS
2. EVGA GeForce GTX 285 FTW - 425 GFLOPS
3. nVidia Tesla C1060 [GT200, 65nm] - 370 GFLOPS
4. AMD FireStream 9270 [RV770, 55nm] - 300 GFLOPS
5. IBM PowerXCell 8i [Cell, 65nm] - 164 GFLOPS

test is fail itself how can gtx285 be better than firestream, tesla, powerxcell ?

[TABSF]

@ Perkel

Sorry but your claims are wrong

the Intel Larrabee 1006 GFLOP/S is correct but I'm guessing it will be 32nm like the Gulftown.
Nvidia GeForce GTX 285 is 1.1TFlop/s
NVidia Tesla C1060u is 4.1TFlop/s with 4 GT 200 chips

"test is fail itself how can gtx285 be better than firestream, tesla, powerxcell ?"

the GTX 285 is better the PowerXcell 8i. the Tesla S1070u consists of 4 GTX 280 chip and is 4.1TFlop/s Single Precision, not sure about the Firestream

PowerXcell 8i is 25.6Gflop/s Double precision
GTX 285 is 77.8Gflop/s Double precision
GT 300 (Fermi) is 630Gflop/s Double precision

Parallel processing the S1070u blades are the Best in the world soon to be S2070u which will be 2.5TFlops Double Precision on one blade and thats why IBM has abandoned the Cell and roadrunner for future HPC.

IBM, Intel and AMD are all going down the CPU/GPU chips route because the benefits of using unified core (or as Nvidia has branded them Cuda core)are amazing. These unified core are High speed, Low power consuming and produces less heat than conventional processors and yet are almost 25 fold the performance.

Whats better 8 cores clocked at 3.2GHz (Cell) or 960 cores clock at 1.7GHz (s1070u)

S1070u Blade is the best Power to Performance blade in the world
S1070u Blade is the best Price to Performance blade in the world

Tell me a Blade that will do 2.5Tflop/s Double precision, because the S2070u Blade being release in January or February will do that.

[GIJeff]

my pen15 does 129 thousand hundred tiggaflops at quad-presision X3.0

[pixelsword]

The PS3 is more powerful than I thought.

[champ21]

ps3's got a 7800 gtx on it.

thats all what matters pretty much.

its reason why pretty much any game thats a bit cutting edge runs @ 600p on the ps3 or the 360.

[Ju]

err, games like UC2, right ? Because they offload huge computational task away from the RSX. Isn't that so ?

Extending the shader power by its vector cores. Exactly what Larrabee is designed for (+having a HW texture unit on chip, basically an integrated RSX).

Larrabee is a super PS3 - well, without PPU (but one can always add a Core duo or Quad or Power7 or whatever to it).

A Cell 2.x could look like a Larrabee chip. Add the missing function from the RSX into the CELL, add more SPUs, add maybe another PPU core (add OOE and better hyperthreading) and optimize the SPU ISA to be more efficient on pure shader operations. Well, and make it twice as fast. Petty IBM/Sony doesn't want to go that route.

[champ21]

uc2 is still 720 and med res textures.

let me know when they make most of their games in 1080p or beyond.

no point being a fan boy. fact is fact.

most console games 720p or below. if the all might cell was so powerful there would be no reason not to have games in 1080p right? and why not atleast have most games running @ 60fps even in 720.. reason because the hardware isnt strong enough.

out of 300 titles released on the ps3 only about 8% do 1080p, rsx just cant handle it.

also if cell was that good.. then why not have 2 cells working together.. instead of having a gpu.

[Ju]

That's quite a change in your original statement from "bit cutting edge runs @ 600p" (looks like CoD is your bar for cutting edge). At least you can admit it can handle 720p. That's great. And it's not just the resolution. Go count some polys in that game.

Even your "belittle" statement "med res textures" tries to lowball that achievement that UC2 has the best textures in any console game so far (med res to what exactly ? - well, and surpasses most PC games as well. I'm not even impressed by Crysis texture work, TBH. Walk up to a wall for example, and don't show me those 2560pixel vegetation shots - because that's where Crysis is good at, but that's about it).

What you totally fail to realize is, that the vector cores (in UC2) take over graphics tasks from the GPU, thus freeing it up to what it's supposed to do best: Rasterization. BTW: most games (certainly 100% (?) of 3rd party games) do not render with the SPU/CELL, but RSX only. No wonder why you can't see the difference.

Well, yeah, I can agree, that 1080p is a stretch to the current generation of consoles (which design goes back to 2001! BTW). But it is the base for pretty much every concept modern graphics HW is coming up today (incl. Larrabee).

BTW: The cell is good in vector ops, what it is not good at is fixed function set pixel ops (like HW texture and lightning, etc). And the vector performance is not big enough to run those in SW.

Larrabee solves this issue by adding dedicated functions on chip for exactly that purpose, hence, a HW texture unit on the chip.

It uses the best of the CELL (and tries to "solve" SW managed data transfer by adding caches and X bar instead of on die memory and ring bus) and adds other parts from (old school) GPUs. But it totaly replaces the gp (general purpose) part of next generation gpGPU (like Tesla) by its own little vector cores.

Oh, and I got one more. The X bar in Larrabee makes it scaleable. The problem with the ring in the cell is, that it pretty much gets saturated with about 8 SPUs. To scale it to 16 or 32 (which could compete against Larrabee), they either would need to switch to X bar too, or add additional rings.

[champ21]

i dont see any mention of cod in my post.. but since you bring it up, its based on an old engine yet still runs only @ 600p on consoles just like most of the releases these days.

So why isnt uncharted 2 running at 60fps @ even 720p if it was soo well optimised? on larger screens i see so many jaggies on uncharted.

clearly they should have offloaded more tasks to the cell freeing up the RSX :P i wonder why they didnt though.

by the way any 3rd party game on the pc out does uncharted.. i dont know what you are smoking. Ever seen Assassins creed 1 run on pc, that blows uncharted away.

[TABSF]

@ champ21

"also if cell was that good.. then why not have 2 cells working together.. instead of having a gpu."

That was Sony's original plan but the Cell can't produce as much triangles as RSX

@ Ju

"err, games like UC2, right ? Because they offload huge computational task away from the RSX. Isn't that so ?"

This is correct the RSX is not good at producing these Graphics features that are present in UC2 or KZ 2 so they only used the RSX for its rending performance, by off loading the complicated task the RSX can use 100% of its performance for millions of triangles.

"Extending the shader power by its vector cores. Exactly what Larrabee is designed for (+having a HW texture unit on chip, basically an integrated RSX)."

Don't understand this one, the RSX uses old pipeline, pixel and vertex cores were as the Larrabee will probably use Unified Shaders.

"A Cell 2.x could look like a Larrabee chip."

The Cell has been abandoned By IBM for future use and by the looks of it the PS4 will using Power7

Sony only owns little rights to the Cell B.E and have almost 0% right in RSX and that's the reason the RSX will remain 65nm because Nvidia Drop RSX almost as soon as the PS3 was released (90nm to 65nm was for financial reasons), Sony investment for the RSX is probably the only reason why Nvidia has not told TSMC to stop production because the chip is expansive to make.

[Ju]

Power 7 (or Power 5, 6 etc) is just a definition from IBM. Despite most people's believe, the CELL's PPU is neither a Power5 nor 6 core, but a based on IBMs mainstream architecture, the G3.

I would rather say, a lot from the CELL went into Power 7 (Power 6 was developed at the same time as CELL). Energy efficiency has not been found in any Power series CPU from IBM but has been vastly improved with the CELL. Those things will find its way into the Power7.

Power7 in combination with a CELL chip is also no contradiction, since a new CELL could replace the old PPU core, now based on a Power7 core (since G3 and Power series seam to have moved closer).

The Cell has not been abandoned. The PowerX8iCell has, which was only targeted to the HPC market. It does not mean, we won't see a consumer Cell in the future (but who knows. Maybe not - this is just to far away).

[the RSX uses old pipeline]

Forget this for a moment and see this from an abstract point of view.

"Shaders" are nothing but vector programs targeting large data sets.

In that case either vertices or pixels (texels), hence "Vertex-" or "Pixelshader". If you take those entities apart, you will find, either one is based on components. One (in the traditional sense out of e.g.) RGBA (or YUV, or whatever) and the other in XYZ0 components. Essentially, both are just vectors.

Before, Vertex shaders took care of coords (XYZ0), while Pixelshaders worked on RGBA values. Unified shaders don't care. They work on vectors. Now, there you have it. SPUs work on vectors. A PS3 has 3 type of vector processors: SPU, RSX-Vertex-Shaders and RSX-Pixel-Shaders (and if you want to be very exact, another VMX unit inside the PPU).

If you can define your dataset properly, you can redirect those vectors to the proper execution unit, (the SPU being a "Unified Shader" in that sense). That's why I said "extending shader power" with the CELL. You would still want to use high bandwidth pixel/vertex shading in the proper units (in the RSX), but support them with some pre/post-processing in the SPU.

[N4g_null]

Ju how much bandwith does a shader take, why does the newer gpus have such faster gddr ram and even SRAM in some cases? Well shaders gometry to pixel to programmable shaders all need what?

Bandwith..... Period

Now tell me how to write code that can get around this problem so I cam get paid. There is no way in he'll the rsx is using it's full polygon setup spec because that would mean that texture work is being done on the spu ok so what is the write speed from and to the cell and the rsx.

I don't care how you cut it you run into the same problem every one runs into howuch ram and how fast is your bottle neck. Not form of off loading fixes this.

Now in krazy kens imaginary cell patient it does not really call for a system on a chip design and franklywhat you are talking about would be crazy expensive to make let along design an coding interface. So far the most I've seen people get out of the cell is low level after effect like vfxs. Most physics are canned other wise you would see cell cgi cards for major 3d modeling programmers and render man over hauled to use it.

Face it man it was a great idea yet poorly executed and then propted up with fanboys raves trying desperatly to keep hope alive or either stick it to ms. Both are valant but in vien enjoy your games ok cell is not the future or even close to it hardware wise and compiler wise.

Another thing is the larrabee clearly needs faster ram access to reach it's max gflop per dollar. Intel is serious about this and it seems they are on track and they do have all of the tech in place to make it happen from design to foundry tweaking so they will not have to ship chips with dead cores in it standard.

Ut2 is nice I guess but crysis amd quake wars says hey. Mega textures and local mega textures are coming what do you think tessalation was for? Go download any of those games and open up a editor we all know you are lucky to get a 512 texture in a ps3 game and 256 is the norm since most games are in sub hd resolutions. The details are possibly fractals.

[Ju]

Point is, because of bandwidth limitation one would try to limit the amount of data you transfer across any bus in the system, hence optimize data before you deliver them to the execution units. In that case RSX bandwith is limited, reduce polys (to avoid overdrawing), etc. Pre-processing in you render pipeline to optimize data transfer the RSX needs to process. You are right, bandwidth is the problem. The RSX ideally only renders polys which it needs to! No transparency would be ideal, because you can really cull all surfaces behind - before they even hit the RSX. If those are fully textured, you save a lot of bandwidth.

Tessellation is nothing but a vertex (!) compressor done in HW. You send less polys to the GPU and the GPU _expands_ those in realtime to give you a higher virtual polycount (per mesh). Why ? To, again, reduce the load on the bus, and reduce the amount of polys the GPU needs to process. That's all it does. The idea is basically the same. Keeping polycount low is essential to keep processing speed up.

Well, of course chips which are released today - almost 5 years after the CELL (and RSX) - are supposed to perform better. Who would doubt that? But it is very surprising, that we even have this discussion today. One would expect, if the CELL idea would have been so wrong, that it wouldn't even be worth talking about here.

Larrabee is just prove, that somebody else thought it's worth following that path.

[ProblemSolver]

@scissor_runner:
"Another thing is the larrabee clearly needs faster ram access to reach it's
max gflop per dollar. Intel is serious about this and it seems they are on
track and they do have all of the tech in place to make it happen from design
to foundry tweaking so they will not have to ship chips with dead cores in it
standard."

Up to the current point Larrabee's memory sub-system is still inferior. Intel
has reworked it multiple times. They even considered to cut down the vector
width from 16 to only 8 due to memory bandwidth and latency issues. But they
haven't, since a width of only 8 would rule out software rasterizing completely.
Well, I'm pretty sure that Larrabee's memory sub-subsystem becomes its weakest
point.

[Hanif-876]

Gran Turismo 5 is 1080p!

[likeaboss302]

No it's not. It doesn't run at a full 1920x1080 resolution. It's like 1280x1080 which is not 1080P. However, it is better then 1280x720 but not a ton.

[TABSF]

"No it's not. It doesn't run at a full 1920x1080 resolution. It's like 1280x1080 which is not 1080P. However, it is better then 1280x720 but not a ton"

This is true, in the menus it will run at 1920 by 1080p with no aa
but in game it will run at 1280 by 1080p with 2x aa

could look crushed.

[likeaboss302]

While everything I see of GT5 looks fantastic. I wish they would have gone with 1280x720 and squeezed out even more details then to go for a non-standard resolution. However, it seems the marketing people won out as most people won't even bother to find out it's not full 1080P.

[free2game365]

How is that 2.7x faster? I've got a GTX 285 and it hits about 1TFLOP

[likedamaster]

What I want to know is...

Can it run Crysis?

[ProblemSolver]

Given the amount of resources, Larrabee can run Crysis, for sure. The question
is, will it run efficiently? Well, that depends on how efficiently one can
implement all the GPU shaders / techniques on Larrabee. That's what Intel has
to prove within the next months.